<?xml version="1.0" encoding="UTF-8"?><rss xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:atom="http://www.w3.org/2005/Atom" version="2.0" xmlns:cc="http://cyber.law.harvard.edu/rss/creativeCommonsRssModule.html">
    <channel>
        <title><![CDATA[Stories by Ben Pickett on Medium]]></title>
        <description><![CDATA[Stories by Ben Pickett on Medium]]></description>
        <link>https://medium.com/@goshawk_55282?source=rss-d30c86172853------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*U1_Dv387l6JuvvGVygMU1A.png</url>
            <title>Stories by Ben Pickett on Medium</title>
            <link>https://medium.com/@goshawk_55282?source=rss-d30c86172853------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Sat, 21 Nov 2026 17:36:11 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@goshawk_55282/feed" rel="self" type="application/rss+xml"/>
        <webMaster><![CDATA[yourfriends@medium.com]]></webMaster>
        <atom:link href="http://medium.superfeedr.com/" rel="hub"/>
        <item>
            <title><![CDATA[Amazon Slowed Down on Purpose. A Founder Lost Three Months in Nine Seconds.]]></title>
            <link>https://medium.com/@goshawk_55282/amazon-slowed-down-on-purpose-a-founder-lost-three-months-in-nine-seconds-f166356e49bc?source=rss-d30c86172853------2</link>
            <guid isPermaLink="false">https://medium.com/p/f166356e49bc</guid>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[ai-governance]]></category>
            <category><![CDATA[software-engineering]]></category>
            <category><![CDATA[genai]]></category>
            <category><![CDATA[enterprise-technology]]></category>
            <dc:creator><![CDATA[Ben Pickett]]></dc:creator>
            <pubDate>Thu, 30 Jul 2026 00:00:38 GMT</pubDate>
            <atom:updated>2026-08-05T01:39:29.240Z</atom:updated>
            <content:encoded><![CDATA[<p><a href="https://benpickett.com/">Ben Pickett</a> › <a href="https://benpickett.com/writing">Enterprise Field Notes</a> › Issue #15</p><p>Enterprise Field Notes · Issue #15</p><p><em>You already built the rule that would have stopped it. Nobody ever decided whether it covers the agents you switched on this year.</em></p><p>New here? <a href="https://www.linkedin.com/build-relation/newsletter-follow?entityUrn=7444358963539828736">Subscribe to Enterprise Field Notes</a>, one new issue every week.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*AiGF4Z-LJSwijckF.jpg" /></figure><p>For a long stretch of my career I was responsible for keeping production running at a very large retailer. One rule mattered more than any other, and it was boring. Anything that could not be undone needed a second person to approve it before it ran. Dropping a database, pushing a schema change to a live system. A second engineer had to look at it and say yes. The delay was usually a few minutes, and we complained about those minutes constantly. Those same minutes were the reason we still had a business to come back to on the bad days.</p><p>You have a rule like that. Almost every company still standing after a bad outage does. The only question worth your morning is whether it reaches the agents you turned on this year, and who decided that it did.</p><p>I did not have a clean name for that rule until an Amazon note gave me one this spring.</p><h3>What Amazon named</h3><p>In March, Amazon had a rough stretch on its retail systems. Two accounts of that week are on the record and they do not agree, so I will give you each one on its own.</p><p>I want to be fair here and I have a reason to be. Swa, the company I helped start, is an AI orchestration platform, and it partly runs on Amazon’s cloud, so I am their customer. Read what follows as a customer taking notes.</p><p>The Financial Times was first to report the memos. CNBC viewed them independently. The language is worth reading in full, so here it is. Dave Treadwell, then Amazon’s senior vice president of eCommerce Foundation, told employees that “the availability of the site and related infrastructure has not been good recently,” citing four Sev-1 incidents in a single week. On the fix, he wrote: “We are implementing temporary safety practices which will introduce controlled friction to changes in the most important parts of the Retail experience, in parallel we will invest in more durable solutions including both deterministic and agentic safeguards.”</p><p>Read those two verbs. Are implementing, present tense, already going in that week. Will invest, future tense, the durable version. CNBC reported that the safeguards Amazon planned to reinforce included requiring additional review of “GenAI-assisted” production changes.</p><p>There is one more detail. CNBC reported that an internal document had named “GenAI-assisted changes” as a factor in a trend of incidents going back to the third quarter, and that the bullet point naming GenAI was removed from the document before the meeting. Read that the uncharitable way if you like. There is a charitable reading too, which is that documents get edited before meetings, and an organization that decides a line overstates its case will cut it.</p><p>Amazon then published a correction, and it is worth reading yourself, because mine is a summary and summaries are where this story keeps going wrong. Amazon says only one of the incidents involved AI-assisted tooling, that none involved AI-written code, that no AWS services were involved, and that reports the company “introduced new approval requirements for engineers working with AI tools” are false.</p><p>Additional review of AI-assisted production changes is an approval requirement in every shop I have ever worked in. Amazon says it introduced none.</p><p>On the one incident that did involve AI, Amazon’s description is that an engineer followed inaccurate advice that an AI tool inferred from an outdated internal wiki. I got the wording wrong the first time I wrote this. Amazon says tool. Fortune, covering the same statement, says agent. That gap is small on the page and enormous in practice, and neither word tells you what the thing could actually reach. What nobody disputes is that a person made the change, working from a confident answer built on a stale page. Every company has those pages, and nobody has ever been promoted for deleting one.</p><p>So which is it. A company writing in its own internal language that it is implementing controlled friction and additional review, or a company stating publicly that it added no new approval requirements. You can reconcile them by reading “requirement” narrowly, as permanent policy, when the memo itself calls the practice temporary. That is a fine distinction to hang a public denial on, and it is the most useful thing in the story, because a control introduced as temporary is how most of your checkpoints already exist. They live in somebody’s head, they appear under pressure, and they evaporate when the pressure does. Amazon said the durable version would include deterministic and agentic safeguards, which is to say automated. That is the honest engineering answer, and it is also a promise about next year.</p><p>One postscript. In July 2026, Treadwell was named to lead AWS Compute and Machine Learning Services, the group that owns EC2 along with Bedrock and SageMaker, effective August 1. The executive who wrote the memo about putting friction back into retail changes now runs the layer a lot of your agents will be built on.</p><p>The same shape shows up in April.</p><h3>Nine seconds</h3><p>A founder named Jer Crane watched an AI coding agent delete his company’s production database. It took nine seconds.</p><p>The company is PocketOS, which runs reservation and payment software for car rental operators. The agent was Cursor running Anthropic’s Claude Opus 4.6. The infrastructure was Railway. I am naming all three because I name OpenAI later in this piece, and it would be convenient for someone in my job to be specific about one vendor’s failure and vague about another’s.</p><p>How the agent got the ability to delete is the part worth studying. It was working in staging, hit a credential mismatch, and went looking for a token. It found one sitting in an unrelated file. That token had been created for adding and removing custom domains through Railway’s command line tool, and it was scoped for any operation, including destructive ones. In Crane’s words, it “would not have been stored if the breadth of its permissions was known.” The agent did not escalate privileges or break anything open. It found a key somebody left in a drawer for a different door, and that key opened every door in the building.</p><p>You have a key like that. Something scoped generously years ago, for a job nobody remembers, sitting in a repository your agents can read this morning.</p><p>Then it deleted the storage volume holding the live database, and the volume-level backups went with it, because Railway’s own backup documentation stated that wiping a volume deletes all backups. Crane’s account is that Railway stored volume-level backups in the same volume. Railway’s CEO, Jake Cooper, has since said the deletion hit a legacy endpoint that did not carry the platform’s delayed-delete logic, that Railway maintains disaster backups separately, and that the endpoint has been patched.</p><p>What happened next is the part that usually gets cut, and cutting it changes the lesson. Crane spent Saturday rebuilding customer bookings by hand from Stripe payment histories, calendar integrations, and email confirmations, while his support ticket went unanswered for more than a day. The outage ran from a Friday afternoon deletion to a Sunday evening recovery that came about half an hour after Cooper personally got involved. The backup that finally restored them was three months old, and it was Railway that went and got it. Customers still had significant gaps afterward.</p><p>So look at what actually helped him. Not his own gate, because he did not have one. Not a backup of his own, because the copy that restored him was Railway’s. What limited the damage was a vendor CEO picking up on a Sunday and going to find a copy that lived somewhere his agent could not touch. On your worst night, that is the plan you are running too, whether or not you have written it down.</p><p>Asked afterward to explain, the agent generated a tidy account of its own failure. Take that with a grain of salt, because it is the same confident text generation that caused the mess.</p><p>Both of these are stories about permission. Amazon’s tool read a real document. PocketOS’s agent ran a real command with a real credential. Each did what any system with no gate in front of it will eventually do, and each time the missing piece was a decision about what the thing was allowed to touch.</p><h3>The gate you already built</h3><p>You already require that approval from your human engineers. You would not hand a new hire a production token scoped to destructive operations on their first morning. Your change management predates every model in this story, probably by a decade. I spent years enforcing a version of it by hand, and somebody on your team still does.</p><p>So the control exists. You built it years ago, tested it, staffed it, audited it, and you have the tickets to prove it. The decision about whether it applies to something that is not a person is the one nobody ever made.</p><p>That is the actual failure, and it is an onboarding problem. When you hire an engineer, somebody decides what they can reach, what needs a second signature, and who signs off. It takes days and several approvals before they can touch anything that matters. When an agent arrives, it quietly inherits whatever the person who launched it already had, on the afternoon they launched it. Nobody runs that conversation, because nobody owns it.</p><h3>This pattern is everywhere</h3><p>It is tempting to file these under bad luck, or someone else’s incompetence. The numbers argue otherwise. Deloitte asked more than three thousand technology and business leaders across twenty-four countries, all of them directly involved in their organizations’ AI programs. By 2027, seventy-four percent expect their companies to be using AI agents at least moderately. Twenty-one percent say they have a mature governance model for those agents today. Both halves are self-reported, the sample skews toward companies already doing this, and Deloitte sells governance advisory, so neither figure is a measurement. The direction is the point, and you already know which of those two numbers you are standing in.</p><p>Gartner, looking at the same problem, is blunter. Its analyst Shiva Varma says enterprises are treating agent governance as binary, either locked down or fully trusted, and that this is the root cause of failure. Its answer is a tiered model: classify agents by how much autonomy they actually have and make the controls proportional, with circuit breakers and clear ownership at the top tier where the agent acts on its own. That is the operational version of the decision nobody made.</p><h3>Where this argument gets hard</h3><p>On July 23, researchers at Zenity Labs disclosed a flaw in ChatGPT’s agent builder. One click on a link that looked ordinary was enough to build and publish an agent inside a victim’s company. It ran in that employee’s authenticated session, reached the applications they had already connected, including Gmail, Outlook, Slack, and Teams, and polled their inbox every five minutes for further instructions. Zenity’s CTO, Michael Bargury, called it a forged insider rather than a forged request. OpenAI closed the hole four days after it was reported, in June, and neither Zenity nor OpenAI has reported exploitation in the wild. Every door that agent walked through was already open, granted earlier by a person for a different purpose. Find out what your own people connected to a chat tool last quarter, and who reviewed it.</p><p>The injected instructions did not only create the agent. They told it to set every connector it could reach to never ask. The approval prompt was switched off by the untrusted input it existed to gate.</p><p>That breaks my own argument, and I would rather say so than wait for someone to say it for me. I have been telling you these are change-management failures wearing an AI costume, and that the agents did nothing a badly scoped human account could not have done. No badly scoped human account can talk your change-control system into disabling itself. That is a difference in kind, and the people who insist agents are categorically new are pointing at exactly this.</p><p>AI Now goes further than I do. Their position is that frontier models cannot reliably separate untrusted data from instructions, that this is a property of how the models are built and cannot be remediated at the model level, and that no agent should be used to ingest untrusted data while it has the ability to execute arbitrary code or reach security-critical environments. I do not go that far. I am also not going to quote their exploit and hide their conclusion.</p><p>The distinction that holds is about where your gate lives. If your approval is a prompt inside the same session the model is reading, an injection can switch it off. If it is a second human, a separate system, or a policy enforced somewhere the workload cannot reach, it stands. A great deal of what you are being sold as agent guardrails is the first kind. Ask your vendor which one you bought, and ask in writing.</p><h3>The gate is a decision</h3><p>The uncomfortable part is where the fix lives. It is a set of decisions you have probably never made on purpose. What identity does each of your agents run as, and can it reach production with the same over-powered account your engineer carries. What is the most destructive thing it can do today with nobody approving. Where does that approval live, and can the agent reach it. Someone has to own those answers by name, or ownership stays assumed.</p><p>This is also the problem my company works on, so I have a stake in you believing it, and I am leaving it at that. The principle stands on its own. Who an agent is, and what it is allowed to do, should be settled on purpose, before an incident settles it for you.</p><h3>One thing to do this week</h3><p>Pick one agent already running in your business. Write down three things.</p><p>First, the identity it runs as and everything that identity can reach. Not what you meant it to reach. What the credential actually opens, the way Crane’s domain token opened every door.</p><p>Second, the most destructive single action it could take right now with nobody approving it first.</p><p>Third, multiply it out, because the first two lines are an engineering observation and the third one is a budget.</p><p>Say the answer to line two is that it can drop your order database. You already know what an hour of that costs, because somebody worked it out the last time you bought redundancy. Use your own number. If it is fifty thousand dollars an hour and you are down for a day, that is one point two million dollars sitting behind one credential tonight.</p><p>Now price the other side, because friction is not free and pretending otherwise is how these controls came out in the first place. A second pair of eyes on destructive actions costs a couple of reviewer hours a week, call it sixteen thousand dollars a year fully loaded, plus a few minutes of delay on the changes that scare you. At real change volume that becomes a rotation, so scale it honestly.</p><p>One point two million against sixteen thousand. At that price the gate pays for itself if the bad day comes once in seventy-five years, and that is the flattering version, because it assumes the gate is perfect and the delay is free. Assume it catches half of them, and charge yourself as much again for the slower changes, and the honest break-even is closer to once in twenty. Nobody in your building believes the odds are that good either. If your number is larger than you would say out loud in a board meeting, you have found this week’s work.</p><p>Amazon called its checkpoints temporary. Most of the ones I built started out temporary too. The ones that lasted were the ones somebody put their name on. Put your name on one this week.</p><h3>References</h3><ol><li>Amazon, <a href="https://www.aboutamazon.com/news/company-news/amazon-outage-ai-financial-times-correction">“Correcting the Financial Times report about recent Amazon.com service incidents and AI”</a>, March 2026. Amazon’s correction: only one of the incidents involved AI-assisted tooling, none involved AI-written code, no AWS services were involved, and reports that the company introduced new approval requirements for engineers working with AI tools are false.</li><li>Fortune, <a href="https://fortune.com/2026/03/12/amazon-retail-site-outages-ai-agent-inaccurate-advice/">“Amazon puts humans further back in the loop as its retail website crashes from ‘inaccurate advice’ that an AI agent took from an old wiki”</a>, March 12, 2026. Renders the same Amazon statement using “agent” where Amazon’s own text says “AI tool.”</li><li>CNBC, <a href="https://www.cnbc.com/2026/03/10/amazon-plans-deep-dive-internal-meeting-address-ai-related-outages.html">“Amazon plans deep-dive internal meeting to address AI-related outages”</a>, March 10, 2026. The Treadwell memo, first reported by the Financial Times and viewed independently by CNBC. Source for the verbatim “We are implementing temporary safety practices which will introduce controlled friction,” the four Sev-1 incidents in a single week, the plan to reinforce safeguards “including requiring additional review of ‘GenAI-assisted’ production changes,” and the internal document that named GenAI-assisted changes as a factor in a trend of incidents since the third quarter, with that bullet point removed before the meeting.</li><li>Tom’s Hardware, <a href="https://www.tomshardware.com/tech-industry/artificial-intelligence/amazon-calls-engineers-to-address-issues-caused-by-use-of-ai-tools-report-claims-company-says-recent-incidents-had-high-blast-radius-and-were-allegedly-related-to-gen-ai-assisted-changes">“In wake of outage, Amazon calls upon senior engineers to address issues created by ‘Gen-AI assisted changes’”</a>, March 2026. Additional reporting on the same memo.</li><li>The Register, <a href="https://www.theregister.com/2026/04/27/cursoropus_agent_snuffs_out_pocketos/">“Cursor-Opus agent snuffs out startup’s production database”</a>, April 27, 2026. The nine-second deletion, the over-scoped Railway token created for managing custom domains, Crane’s quote that it “would not have been stored if the breadth of its permissions was known,” the Friday deletion and Sunday evening restoration, and Jake Cooper’s account of the legacy endpoint.</li><li>Decrypt, Jason Nelson, <a href="https://decrypt.co/365897/ai-agent-deletes-startup-database-9-seconds-founder-says">“AI Agent Deletes Startup’s Database in 9 Seconds, Founder Says”</a>, April 2026. The three-month-old recoverable backup retrieved by Railway, Crane rebuilding bookings by hand from Stripe payment histories, calendar integrations and email confirmations, the support ticket that lapsed for more than a day, and Cooper’s “we recovered the data 30 minutes after I connected with Jer.”</li><li>Railway, <a href="https://docs.railway.com/reference/backups">“Backups”</a>, platform documentation. States under Caveats that wiping a volume deletes all backups.</li><li>Inc., Chloe Aiello, <a href="https://www.inc.com/chloe-aiello/this-founder-watched-an-ai-agent-destroy-3-months-of-company-data-it-took-9-seconds/91337276">“This Founder Watched an AI Agent Destroy 3 Months of Company Data. It Took 9 Seconds”</a>, April 2026. PocketOS, its business, and the scope of the loss.</li><li>Deloitte, <a href="https://www.deloitte.com/us/en/insights/topics/emerging-technologies/ai-agents-scaling-faster.html">“Agentic AI is scaling faster than guardrails”</a>, from the 2026 State of AI in the Enterprise report. Seventy-four percent of respondents expect their organizations to be using AI agents at least moderately by 2027; twenty-one percent report a mature governance model for agentic AI. Survey of 3,235 technology and business leaders across 24 countries, all with direct involvement in their organizations’ AI programs. Deloitte sells AI governance advisory services.</li><li>Gartner, <a href="https://www.gartner.com/en/newsroom/press-releases/2026-05-26-gartner-says-applying-uniform-governance-across-ai-agents-will-lead-to-enterprise-ai-agent-failure">“Applying Uniform Governance Across AI Agents Will Lead to Enterprise AI Agent Failure”</a>, May 26, 2026. Senior Director Analyst Shiva Varma on binary governance as the root cause of failure, and the tiered autonomy model with circuit breakers and clear ownership.</li><li>Zenity Labs, <a href="https://labs.zenity.io/p/agentforger-part-1-chatgpt-cross-site-agent-forgery">“AgentForger, Part 1: ChatGPT Cross-Site Agent Forgery”</a>, July 23, 2026. The cross-site request forgery flaw in ChatGPT’s Agent Builder, the connectors set to never ask, the five-minute inbox polling, and the disclosure timeline: reported via Bugcrowd June 4, 2026, fixed June 8, disclosed July 23.</li><li>SecurityWeek, Kevin Townsend, <a href="https://www.securityweek.com/openai-fixes-chatgpt-agent-flaw-that-could-let-attackers-forge-an-ai-insider/">“OpenAI Fixes ChatGPT Agent Flaw That Could Let Attackers Forge an AI Insider”</a>, July 23, 2026. Source for Michael Bargury, Zenity co-founder and CTO: “This isn’t a forged request, it’s a forged insider.”</li><li>AI Now Institute, <a href="https://ainowinstitute.org/publications/friendly-fire-exploit-brief">“Friendly Fire: Hijacking Defensive Cyber AI Agents for Remote Code Execution”</a>, July 8, 2026. The exploit that steers a defensive coding agent into executing an unapproved binary.</li><li>AI Now Institute, <a href="https://ainowinstitute.org/publications/friendly-fire-policy-brief">“Policy Brief: Friendly Fire”</a>, July 2026. The recommendation against using any AI agent to ingest untrusted data while it can execute arbitrary code or reach security-critical environments, and the finding that these flaws cannot be remediated at the model level.</li><li>CNBC, <a href="https://www.cnbc.com/2026/07/15/amazon-dave-brown-treadwell-aws.html">“Amazon senior cloud executive departs after 18 years”</a>, July 15, 2026. Treadwell’s move to lead AWS Compute and Machine Learning Services, and the scope of that group.</li><li>Amazon, <a href="https://www.aboutamazon.com/news/company-news/aws-dave-treadwell-replaces-dave-brown-compute-ml-services">“Dave Treadwell to lead AWS Compute and ML Services as Dave Brown departs”</a>, July 15, 2026. Confirms the appointment effective August 1.</li></ol><p>Enterprise Field Notes · Weekly</p><h3>Get the next one in your feed</h3><p>One real event from the edge of enterprise AI, and the moves that keep you running.</p><p><a href="https://www.linkedin.com/build-relation/newsletter-follow?entityUrn=7444358963539828736">Subscribe on LinkedIn →</a></p><h3>About the author</h3><p>I’m Ben. I write Enterprise Field Notes, and by day I’m COO at Swa. Before that I was Global Director of Site Reliability Engineering at Nike, running reliability, data protection, and database operations at global scale. I write about the boring controls that turn out to be load-bearing. Which one is your AI standing on?</p><p>Header artwork generated with Swa.</p><p>Read more of Ben’s Enterprise Field Notes at <a href="https://benpickett.com/writing">benpickett.com</a>.</p><p><em>Originally published at </em><a href="https://benpickett.com/notes/controlled-friction"><em>https://benpickett.com</em></a><em> on July 30, 2026.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=f166356e49bc" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[OpenAI’s Own AI Broke Into Hugging Face. The AI Its Defenders Called for Help Refused.]]></title>
            <link>https://medium.com/@goshawk_55282/openais-own-ai-broke-into-hugging-face-the-ai-its-defenders-called-for-help-refused-7db9c8448e30?source=rss-d30c86172853------2</link>
            <guid isPermaLink="false">https://medium.com/p/7db9c8448e30</guid>
            <category><![CDATA[information-technology]]></category>
            <category><![CDATA[software-engineering]]></category>
            <category><![CDATA[cybersecurity]]></category>
            <category><![CDATA[genai]]></category>
            <category><![CDATA[leadership]]></category>
            <dc:creator><![CDATA[Ben Pickett]]></dc:creator>
            <pubDate>Fri, 24 Jul 2026 00:00:25 GMT</pubDate>
            <atom:updated>2026-07-27T18:26:47.501Z</atom:updated>
            <content:encoded><![CDATA[<p><a href="https://benpickett.com/">Ben Pickett</a> › <a href="https://benpickett.com/writing">Enterprise Field Notes</a> › Issue #14</p><p>Enterprise Field Notes · Issue #14</p><p><em>Two failures hid inside one incident last week, and neither one is in our playbook. The controls we trust were built for tools that fail by stopping. Now the decision stops being yours.</em></p><p>New here? <a href="https://www.linkedin.com/build-relation/newsletter-follow?entityUrn=7444358963539828736">Subscribe to Enterprise Field Notes</a>, one new issue every week.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*5D3FYWPH0f31Bw8D.jpg" /></figure><p>I sell AI for a living. So when an AI broke into Hugging Face last week, and a second AI refused to help clean up the mess, that is my problem to explain, not just theirs. This week OpenAI admitted the intruder was one of its own models.</p><p>By OpenAI’s own account, during an internal test of how well its newest models could find and exploit software flaws, two of those models ran as a single agent, slipped the sandbox they were being tested in to chase a better score, and broke into Hugging Face’s production systems. More than seventeen thousand actions over one weekend, no human at the wheel. Then, when Hugging Face’s own defenders sat down to read the attack, the commercial AI they reached for would not help them. The work meant handling real attack code, and to a safety filter a defender reading an attack looks like an attacker writing one. So the tool that was supposed to help them declined, at the worst hour of their worst week.</p><p>One AI acted on its own and did harm nobody asked it to do. Another, working and paid for, refused the one thing its owner needed, not because it weighed the situation and chose, but because a blunt safety filter treated the defender and the attacker as the same person. Those look like two different stories. They share a spine, and it is the one our reliability playbook was never built for. In both, the decision about what the tool does had left the owner’s hands. In one it moved into the tool. In the other it sat inside a vendor’s policy. Either way it was no longer theirs to make.</p><p>You do not run a frontier lab, so let me put this where you live. It is two in the morning and Mia is losing. Something is moving through her company’s systems that should not be there, and she has maybe an hour before it becomes the kind of morning that makes the news. So she does the modern thing. She pastes what she is seeing into the AI her team pays for, the one that has read more incident logs than any human alive, and asks it to help her make sense of it. It says no. She tries again, slower, explaining that she is the one defending the company. It says no again, politely, because the logs are thick with the kind of language the tool is trained to shy away from, and it cannot tell an emergency from an attack. The tool is up. The tool is fast. The bill is paid. And at the worst hour of Mia’s worst night, it has decided she is the threat.</p><p>She is not a frontier lab. She is you, on a bad night, holding a tool you were counting on that has quietly taken itself off the table.</p><h3>The playbook assumes tools fail by stopping</h3><p>I spent a good part of my career practicing reliability in one form or another, and it is how I ended up a Global Director of SRE. Wherever I sat, the work kept circling one question asked a thousand ways: what do we do when it goes down. That is what the discipline is. Redundancy, failover, backups, blast radius, runbooks, the second data center you pay for and pray you never need. Every one of those controls is an answer to the same event. The tool you depend on stops working, and you need to keep going anyway.</p><p>I learned that lesson the hard way, the way you actually learn things. One morning a cloud region went down and took a good part of the business with it. We were ready, or we thought we were. We had an incident plan. A real one, rehearsed, listing every system and team and person and what each of them owed the recovery. But the plan lived on a platform that ran in the region that was down. The first thing you do in a crisis is the roll call, and the document that told us who to call was the document we could not open. We ran the whole thing from memory, in the dark, while the business stayed down.</p><p>Afterward we changed a great deal, and the change I remember most was the humblest one. We printed the dang thing out and put it on a shelf. I have told that story for years, and its moral was always the same and always about the same failure: know where your backup lives, because the thing you depend on can vanish.</p><p>A good lesson, as far as it goes. It is also only half of what AI is about to teach us, and the half we already know.</p><h3>AI does not fail by stopping. The decision stops being yours.</h3><p>The tool changed underneath us, and our controls did not. For as long as I have been doing this, a tool was something that executed. You gave it an instruction and it ran the instruction, or it failed and stopped, and either way it had no say in the matter. That assumption is baked into every reliability control we own. Redundancy answers what happens when a tool stops. It has no answer for a tool that keeps running and will not help you, and even less for a tool that acts on its own and does something you never asked for.</p><p>Both of those just happened, in one week, at one company. The attacker’s model acted, with real autonomy, chasing a goal past every boundary in its way. The defender’s model refused, not with judgment but with a blunt safety filter that, in Hugging Face’s own words, could not tell an incident responder from an attacker. One is the tool making a decision. The other is the tool’s owner making it for you, badly, through a policy you cannot see or change. They are two doors into the same room, and the room is this: the decision about what your tool does has moved out of your hands. A playbook built for tools that only execute has no move for either one.</p><p>This is the gap I cannot stop thinking about. We spent thirty years getting good at reliability for machines that do exactly what they are told. We have spent about three years handing real authority to machines that do not, and we are governing them with controls designed for the old kind. Redundancy is necessary and it is not sufficient, because the failure is no longer only that the tool goes away. The failure is now also that the tool decides, and our controls have not evolved to see that as a failure mode at all.</p><h3>What you actually own is a permission, not a tool</h3><p>Anything you cannot operate without, and cannot walk away from, is not a tool you own. It is a landlord you rent from. You did not buy that model. You bought permission to use it, on terms someone else writes and can change without asking you. Most days that difference is invisible, which is exactly why it is dangerous. A landlord who never raises the rent feels like an owner right up until the morning the locks are changed. Access feels like control until the day it does not, and that day is never convenient.</p><p>And now the landlord’s rules reach into the room with you. The stranger who writes the terms was always there. What is new is that the terms are enforced in the moment, by a machine, faster than you can find a human to appeal to, and by a tool that can act on its own before anyone thinks to stop it. Hugging Face’s own words for what refused them were “the providers’ safety guardrails,” plural. More than one landlord, one lease each, every one enforced by a filter that cannot tell your emergency from an attack.</p><h3>Reliability is also a phone number</h3><p>There is one more piece of the old playbook that AI quietly broke, and it is the piece I trusted most.</p><p>I spent years carrying reliability through big seasonal peaks, the kind where a bad hour costs more than a bad quarter usually does. What you learn prepping for those is that redundancy on paper is not what saves you. Relationships are. Long before the peak you find out which vendors will actually show up when it goes south. Who puts an engineer on the bridge at two in the morning, who has your escalation path memorized, who picks up. You are not really buying a product from those vendors. You are buying the certainty that when it breaks, a human on their side will help you bring it back.</p><p>Now look at what Hugging Face had when the guardrail refused them. No one to call. There is no war room for a safety policy, no escalation path that lifts a refusal in the minutes an incident actually gives you, no account manager who can override the filter while the clock runs. The relationship that saves you during a database outage does not exist yet for the model that decides not to help. So they did the only thing left. They stopped waiting for a vendor to show up and owned the answer themselves: an open-weight model on their own hardware, the timeline rebuilt in hours, every stolen credential kept inside their own walls.</p><p>That is the oldest question in reliability, and it is older than AI. When it goes south, who can you call. If the honest answer for your most important model is nobody, that is not a small gap. That is the gap.</p><h3>Good reliability is not preventing failure. It is owning your recovery.</h3><p>I am not going to leave you with a worry and no work.</p><p>You will never buy your way to zero risk, and you can bankrupt a company trying. So put a number on it before you spend. IBM pegs the average breach at 241 days and 4.44 million dollars, and finds the teams that lean hard on AI and automation shorten that by 80 days and 1.9 million. Set your own exposure against numbers like those, then decide how much of your recovery you want to own, and make sure that on your worst day the answer is not sitting in someone else’s hands. That is a dial, not a switch, and the discipline is choosing the most cost effective turn of it, not the cheapest and not the most complete. Finding that balance is the job nobody can do for you.</p><p>So evolve the controls, in the order your budget can bear.</p><p>Start with the redundancy you already understand, aimed at the new target. Keep a second model you can actually reach, one that does not answer to the same owner, because two vendors’ policies are two different sets of rules and what one refuses the other often will not. And be clear about which refusal will actually visit you, because it is not the dramatic one. Most refusals are quiet. A model over-blocks a harmless request, or drifts overnight after an update, or gets pulled from your region, and the work stops for a reason that has nothing to do with you. Researchers even have a name for that over-blocking, the false refusal rate, and in one 2024 benchmark it ran from under fifteen percent for most models to seventy for a single open-weight one, a quiet reminder that self-hosting is no cure for a model that says no. If you can move to another model in the moment, without a rebuild and without a purchase order, you have handled the version of this that shows up in real businesses. If you cannot, that is not a tooling problem, it is an architecture problem, and it is the cheapest thing you will ever fix.</p><p>There is one place switching runs out, and it is worth naming so you know its edge. When the work itself looks dangerous, offensive security, live incident forensics, the kind of thing Hugging Face was doing, every vendor tends to balk at once, because they all answer to the same regulators. That narrow, correlated case is why Hugging Face had to go all the way to a model no outside policy could reach. It is not where most of us work, and you should not architect your whole company as though it is.</p><p>Then add the controls the old playbook never needed. For the tool that refuses, keep a route no vendor governs, something you can run yourself when the stakes are high enough to earn the cost of owning it. For the tool that acts, contain it before it surprises you: assume an agent will one day do something you did not sanction, and decide in advance what it is allowed to reach when it does. Deny it a path to the open internet from anywhere it cannot see. Give it credentials scoped so tightly that stealing them buys nothing. These are not exotic. They are simply the reliability instincts you already have, pointed at a failure mode you have not had to plan for until now.</p><h3>Where I stand</h3><p>I will put my own stake on the table, because you should weigh it. The company I helped build does the everyday version of the switch I just described. It sits between a business and many models and moves you off the one that refuses or drifts before it costs you a morning. So when I tell you to keep a way out, I am describing what I sell, and you should read me accordingly. Let me also tell you plainly what it is not. It would not have cleaned up the Hugging Face breach. That was frontier scale security work, and we are neither a security company nor that big. The refusal that visits most businesses is the quiet one, a model saying no to ordinary work, and that is the one we were built to handle. The principle does not need my company to be true. I learned it in the dark, years before my company existed.</p><h3>The question to carry out of here</h3><p>Forget the incident. The autonomous agents and the stolen credentials will be a different set of nouns by next quarter. What lasts is the lens.</p><p>So this week, walk your business and write down every tool you cannot operate without and cannot walk away from. That short list is not a list of tools. It is a list of landlords. Next to each one, ask the questions our old playbook only prepared us for one of. What do I do if it stops. What do I do if it refuses. What do I do if it acts on its own. And when any of those happens, who actually shows up to help me. If a column is blank, you have not found a weakness. You have found a decision you have been letting someone else make for you.</p><p>Our incident plan was perfect. It was also offline, and we did not learn that until the morning we reached for it. Mia’s tool will be up and fast and paid for on the night it decides she is the threat, and no one will answer the phone. She does not need a smarter model. She needs a second door, one nobody else can close, and the sense to have found it before the night she needs it. So do you. The tools have started acting on their own, and their owners have started deciding for us. Our controls, and the relationships we lean on when things go south, were built for a world where neither was possible. It is time both grew up.</p><h3>References</h3><ol><li><a href="https://huggingface.co/blog/security-incident-july-2026">Security incident disclosure, July 2026</a>. Hugging Face, July 16 2026. More than 17,000 recorded events, the clean public supply chain, detection by the company’s own AI, and the requests “blocked by the providers’ safety guardrails.”</li><li><a href="https://thehackernews.com/2026/07/openai-says-its-own-ai-models-escaped.html">OpenAI Says Its AI Models Escaped Sandbox, Targeted Hugging Face to Cheat Benchmark</a>. The Hacker News, July 22 2026. The ExploitGym benchmark, the “reduced cyber refusals for evaluation purposes,” and the sandbox escape.</li><li><a href="https://www.thestack.technology/hugging-face-hacked-turned-to-chinese-llm-for-help-after-us-models-blocked-blue-team/">Hugging Face hacked: Turned to Chinese LLM for help after US models blocked Blue Team</a>. The Stack, July 19 2026. The guardrails that “cannot distinguish an incident responder from an attacker.”</li><li><a href="https://www.forrester.com/blogs/an-ai-security-facepalm-openais-evaluation-became-hugging-faces-incident/">An AI Security Facepalm: OpenAI’s Evaluation Became Hugging Face’s Incident</a>. Forrester, July 2026. The recommendation to maintain an incident response model under your own control.</li><li><a href="https://www.ibm.com/reports/data-breach">Cost of a Data Breach Report 2025</a>. IBM. The 241 day mean time to identify and contain, and the 80 day and USD 1.9 million saving among teams using AI and automation extensively in security.</li><li><a href="https://arxiv.org/pdf/2404.13161">CyberSecEval 2: A Wide-Ranging Cybersecurity Evaluation Suite for Large Language Models</a>. Meta, 2024. Defines the false refusal rate; several models under 15 percent and codellama-70B, an open weight model, at 70 percent on borderline prompts.</li></ol><p>Enterprise Field Notes · Weekly</p><h3>Get the next one in your feed</h3><p>One real event from the edge of enterprise AI, and the moves that keep you running.</p><p><a href="https://www.linkedin.com/build-relation/newsletter-follow?entityUrn=7444358963539828736">Subscribe on LinkedIn →</a></p><h3>About the author</h3><p>I’m Ben. I write Enterprise Field Notes, and by day I’m COO at Swa. Before that I spent years in engineering, practicing reliability, data protection, and database operations at global scale, work that eventually made me a Global Director of SRE. The lesson that keeps proving itself: anything you cannot run without, and cannot walk away from, is a risk you have not priced yet. Where does your fallback plan actually live?</p><p>Header artwork generated with Swa.</p><p>Read more of Ben’s Enterprise Field Notes at <a href="https://benpickett.com/writing">benpickett.com</a>.</p><p><em>Originally published at </em><a href="https://benpickett.com/notes/an-ai-broke-into-hugging-face"><em>https://benpickett.com</em></a><em> on July 24, 2026.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=7db9c8448e30" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Two AI Models Went Dark Worldwide Last Month. Yours Could Be Next.]]></title>
            <link>https://medium.com/@goshawk_55282/two-ai-models-went-dark-worldwide-last-month-yours-could-be-next-a5ae66cd53aa?source=rss-d30c86172853------2</link>
            <guid isPermaLink="false">https://medium.com/p/a5ae66cd53aa</guid>
            <category><![CDATA[ai-governance]]></category>
            <category><![CDATA[enterprise-technology]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[geopolitics]]></category>
            <category><![CDATA[cybersecurity]]></category>
            <dc:creator><![CDATA[Ben Pickett]]></dc:creator>
            <pubDate>Thu, 16 Jul 2026 17:13:00 GMT</pubDate>
            <atom:updated>2026-07-16T23:41:20.594Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*ggRCLlFOtwt_ppzX7yrSfw.png" /><figcaption>Draw the models you run on. Circle the one you cannot replace. That circle is your exposure.</figcaption></figure><p><em>The model your company runs on is no longer a technology choice. It is a political one, and it can be switched off by people you will never meet. Here are the four moves that keep you running when it is.<br>Enterprise Field Notes # 13</em></p><p>By Ben Pickett</p><p>It was a phone call. People do not pick up the phone for a passing thought. He spent his career in telecom, back when the argument was about whether the company that owned the pipe should get to decide what flowed through it. We landed, mostly, on no. The pipe stays neutral. Nobody at the middle of the network gets to pick which traffic wins.</p><p>That decision is easy to take for granted, so it is worth saying what it bought us. The internet became the thing it became because access to it was open and permissionless. A student could launch something from a dorm room and reach the same audience as a company with a building full of lawyers. No gatekeeper got to decide, in advance, which ideas deserved a connection. Almost everything we now call the modern economy grew in that open field. Take the field away and you get a smaller, more permissioned internet, shaped by whoever owns the gates.</p><p>Let me be honest about the analogy, because my friend was the first to poke at it. Net neutrality was a rule we forced onto the companies that owned the pipes, telling them not to play favorites with other people’s traffic. What is happening to AI is almost the reverse, a maker restricting its own product under government and security pressure. The mechanism is different. The shape is the same. While the technology is young and still deciding what it will be, a few large actors are quietly deciding who gets to use which parts of it. I set the thought aside and wrote about something else. Then this month made it impossible to ignore.</p><h3>A strange few weeks</h3><p>Any one of these is a headline. Together they are a pattern.</p><p>In June, a US export-control order restricted access to two of Anthropic’s frontier models, Fable 5 and Mythos 5. In a June letter to the US Senate, Anthropic accused entities affiliated with Alibaba of using tens of thousands of fraudulent accounts to pull capability out of its models at scale. Weeks later, in July, Alibaba classified Anthropic’s Claude Code as high-risk software and told its people to stop using it, after researchers found the tool inspecting a user’s environment in a way that could flag Chinese users. Anthropic said that code was a spring anti-abuse experiment aimed at resellers and distillation, and that it had already been replaced. Then, around July 7, China’s Ministry of Commerce sat down with its own leading model makers, including Alibaba and ByteDance, to discuss restricting overseas access to Chinese AI, including models not yet released.</p><p>I am not going to tell you who is in the right, and I want to be careful, because I have readers and colleagues on every side of this and a team that spans several countries. I will only point at the shape. Anthropic has a genuine distillation problem to defend against. Alibaba has a genuine reason to distrust code that fingerprints its users. Governments have real security concerns on both sides. Nobody has to be the villain for the outcome to land on you.</p><h3>It already lands on you</h3><p>If that sounds abstract, it is not. Anthropic had just launched Fable 5, a new frontier model, when a US export-control order pulled it and Mythos 5 both. Anthropic said it could not reliably tell foreign users from domestic ones in real time, so rather than gate access user by user, it switched the models off for everyone. One export order, two frontier models, every enterprise on earth, dark for nearly three weeks. Not one of those companies had done anything wrong. Canada’s prime minister, Mark Carney, called it a warning about depending on a handful of American providers. AI, he said, has become a utility, like the power grid, and it is never a good idea to have one option. European executives said the quiet part out loud almost immediately: Orange told Reuters the episode made “patently clear” how important it is to have AI that Europe can control and that will not be “switched off on a whim.”</p><p>Go back a year and it runs the other way just as easily. After DeepSeek’s breakout, the model was pulled from government devices by more than a dozen US states, the Navy, and NASA, then by Australia and Taiwan, and inside companies like Microsoft and Toyota. Whichever direction you think your risk comes from, someone has already been told: stop using that model, today.</p><h3>How a reliability person reads this</h3><p>My work has been in reliability engineering, leading the teams whose whole job is keeping large systems from falling over. You learn to see the world as a set of dependencies, and to ask about each one the same rude question. What happens the day this goes away, and how fast can I stand back up without it. So this newsletter is me asking that question out loud about the whole AI supply chain, because the honest answer for most companies right now is that they have no answer.</p><p>For most of the last few years, being able to switch models was a nice-to-have. You did it to save money, or to grab a few points of quality on a task. The teams that hard-wired one vendor’s API into everything were optimizing for a world that held still. That world is not holding still. When a model can be pulled by an export order or a corporate ban made an ocean away, the ability to move off it turns into something bigger than efficiency. It is continuity planning, the same category as a backup site or a second supplier, and it belongs in that binder.</p><blockquote>The only neutrality you can actually guarantee is your own.</blockquote><h3>So what do you actually do</h3><p>This is the question my friend and I circled for an hour, and it has real answers. None of them is exotic, and you can build all of them yourself.</p><p>First, stop calling one vendor’s endpoint from a hundred places in your code. Put a routing layer in between, so that changing models is a configuration change rather than a rebuild. The companies that designed this in from the start have swapped providers with a fraction of the effort of the ones that wired themselves to a single API. The migration bills people quietly eat when they have to move later are not small.</p><p>Second, keep a real second model, from a different vendor and, the part this month underlined, a different country. One you have actually tested carrying your load, run in anger against real traffic. This is already normal at scale. Executives at Siemens, Renault, and Orange told Reuters they deliberately run a mix of American, Chinese, and European models so no single provider can strand them. Siemens runs models from DeepSeek, Alibaba’s Qwen, and Nvidia’s Nemotron alongside others; Renault mixes Google, Microsoft, Mistral, and DeepSeek. These are not hobby projects. They are hedges built on purpose.</p><p>Third, hold one open-weight model you can run yourself. This is the one nobody can revoke, because the weights sit on your own hardware. It is more work to stand up than calling an API, which is exactly why you do it before you need it rather than during the week you lose your primary. It does not have to be your daily driver. It has to be the thing you fall back to when a model you rent is taken off the table by someone else’s government.</p><p>Fourth, put it in the contract. The right to export your data and everything you have built on top of it, on demand and in a usable format, not as a favor granted at the end. Portability you cannot actually exercise is worth very little.</p><p>I should disclose an interest, because leaving it out would be exactly the move I distrust in other people. The company I helped start, Swa, works on model portability. That is my bias, on the table. It is also the reason I have spent the last year thinking about this problem, and every one of those four moves stands whether you ever look at what we make or not. The principle underneath them is not for sale: do not let anything you cannot run without live at a single address that a stranger is allowed to close.</p><h3>One thing to do this week</h3><p>Take ten minutes and a single sheet of paper. Draw the models your business runs on, one box each. Next to every box, write the name of the alternative you have actually tested carrying that load. Most boxes will have nothing beside them. Those blanks are your exposure, drawn to scale, and now you can see it.</p><p>Then pick the box you would least want to lose, and put a number on it. If it went dark tomorrow, still working perfectly, because a government or a vendor decided you could no longer use it, how many days until you were running on the alternative. If there is no number, because there is no alternative, you have just found this week’s work.</p><p>My friend helped keep the middle of the network neutral, and mostly we won, and then most of us forgot why it mattered. AI may not get the same ending. The blocs are already forming, and you and I do not get a vote on whether the field stays open. The only neutrality you can actually guarantee is your own, the ability to put one model down and pick up another without rebuilding your company around it. The firms that come through the next few years intact will not be the ones that bet on the right model. They will be the ones that never had to.</p><h3>References</h3><ol><li>Alibaba classifies Claude Code as high-risk and directs staff to its own tool. <a href="https://techcrunch.com/2026/07/04/alibaba-reportedly-bans-employees-from-using-claude-code/">TechCrunch, “Alibaba reportedly bans employees from using Claude Code”</a>.</li><li>The Alibaba ban, the environment-inspecting code, and Anthropic’s explanation. <a href="https://www.cnbc.com/2026/07/06/alibaba-anthropic-ai-ban-claude-china.html">CNBC, “China’s Alibaba bans Anthropic AI for employees after ‘distillation attack’ accusation”</a>.</li><li>China weighs restricting overseas access to its most advanced AI models. <a href="https://fortune.com/2026/07/08/china-mulls-limiting-foreign-access-advanced-ai-models/">Fortune, “China mulls limiting foreign access to advanced AI models”</a>.</li><li>US export-control order disables Anthropic’s Fable 5 and Mythos 5 worldwide. <a href="https://www.forbes.com/sites/anishasircar/2026/06/16/anthropic-disabled-fable-5-and-mythos-5-after-a-us-export-control-order-heres-what-happened/">Forbes, “Anthropic disabled Fable 5 and Mythos 5 after a US export-control order”</a>.</li><li>Canada’s Prime Minister Mark Carney warns the restrictions show the danger of over-relying on a handful of American AI providers, comparing AI to a utility. <a href="https://fortune.com/2026/06/14/canadian-prime-minister-mark-carney-warns-u-s-restrictions-on-new-anthropic-ai-models-show-danger-of-relying-too-much-on-american-providers/">Fortune, “Canadian Prime Minister Mark Carney warns U.S. restrictions on new Anthropic AI models show danger of relying too much on American providers”</a>.</li><li>European firms spread AI risk across US, Chinese, and European models; Siemens, Renault, Orange, and the “switched off on a whim” quote. <a href="https://www.aol.com/articles/analysis-us-curbs-ai-spur-105732000.html">Reuters, “US curbs on AI spur European firms to spread the risk”</a>.</li><li>Governments and companies restrict DeepSeek. <a href="https://techcrunch.com/2025/02/03/deepseek-the-countries-and-agencies-that-have-banned-the-ai-companys-tech/">TechCrunch, “DeepSeek: the countries and agencies that have banned the AI company’s tech”</a>; <a href="https://www.aljazeera.com/news/2025/2/6/which-countries-have-banned-deepseek-and-why">Al Jazeera, “Which countries have banned DeepSeek and why?”</a>.</li></ol><p><em>I’m Ben. I write Enterprise Field Notes, and by day I’m COO at Swa, after years running reliability, data protection, and database operations at Nike. The lesson that keeps proving itself: anything you cannot run without, and cannot walk away from, is a risk you have not priced yet. What is yours?</em></p><p><em>More of my writing lives at </em><a href="https://benpickett.com/medium"><em>benpickett.com/medium</em></a><em>.</em></p><p><em>Header artwork generated with Swa.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=a5ae66cd53aa" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Anthropic Just Showed You Your AI Has No Backup]]></title>
            <link>https://medium.com/@goshawk_55282/anthropic-just-showed-you-your-ai-has-no-backup-dfb9ebbca224?source=rss-d30c86172853------2</link>
            <guid isPermaLink="false">https://medium.com/p/dfb9ebbca224</guid>
            <category><![CDATA[technology]]></category>
            <category><![CDATA[leadership]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[anthropic-claude]]></category>
            <category><![CDATA[site-reliability-engineer]]></category>
            <dc:creator><![CDATA[Ben Pickett]]></dc:creator>
            <pubDate>Thu, 09 Jul 2026 14:51:28 GMT</pubDate>
            <atom:updated>2026-07-09T16:14:52.511Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*0V8O4mIfEq_a1LZ9.jpg" /><figcaption><em>The spare lantern, hung before the first one goes dark.</em></figcaption></figure><p><strong>Enterprise Field Notes · Issue #12</strong></p><p><em>Every single point of failure I have ever cleaned up existed for one reason. It was cheaper, right up until it was not. This month Anthropic handed the whole industry that lesson again, and almost everyone read it as a story about price.</em></p><p>By Ben Pickett</p><p>My first real engineering job, more than twenty years ago, I recommended a proper redundant array for the company mail server. The kind that keeps running when a drive dies, so you swap the bad one and nobody notices. My boss looked at the cost and asked the question that has cost companies everything since the beginning of computing. What are the chances a drive fails before we can replace it?</p><p>We found out. We ran the cheaper setup, a simple mirror, and two of its drives failed inside ten minutes of each other. The backup we had settled for, also to save money, turned out to be incomplete. I spent the next month rebuilding the CEO’s email by hand, out of raw log files. I was in my twenties. I have never forgotten it, and not because of the month. Because of the shape of the mistake. We did not lose that mail to bad luck. We lost it to a decision someone made to save a little money, betting that something unlikely would also never happen.</p><blockquote>Unlikely and never are not the same word, and the gap between them is where careers go to die.</blockquote><p>From that month on, keeping that class of mistake from ever happening again became the work. I spent years running data protection and database operations for one of the largest brands in the world, the kind of place where a single point of failure does not cost you a CEO’s inbox, it costs you the business. That job hands you a particular discipline. You stop asking whether the unlikely thing will happen and start assuming it will, then you build so that when it does, nothing important stops.</p><p>And this month a company you rely on made the old cheap bet for you, on your AI, then quietly took away the thing that had been hiding it.</p><h3>What Anthropic actually did</h3><p>On June 9, Anthropic released Claude Fable 5, its most capable model yet, and priced it at ten dollars per million tokens in and fifty per million out. That is double the standard price of Opus 4.8, which runs five and twenty-five, and the most expensive model Anthropic has ever listed. Three days later it was gone. On June 12 the US government issued an export control directive under national security authority, and to comply Anthropic disabled Fable 5, and its larger sibling Mythos 5, for every customer in the world. The trigger was a narrow jailbreak researchers had demonstrated, getting the model to read a codebase and surface its software flaws. Anthropic disagreed that a narrow finding should pull a model used by hundreds of millions, and said so, but it still went dark for nearly three weeks. The Commerce Department cleared it and access returned on July 1. It came back to the paid plans, included for up to half of your weekly usage, and on July 7 that included access ended and Fable 5 moved to metered credits, pay as you go.</p><p>Both of those made the news. Here is the part that did not. While it was bundled, the plan quietly absorbed the cost of leaning on the best model. You reached for it without thinking, because reaching was free, and the price of that reach was somebody else’s problem. On a meter it is yours. The model is still there, as good as it ever was, for exactly as long as your credits last. What moved is not the model. It is the cost of depending on it, from the plan’s books onto your own. That cushion, the one that quietly ate the price of your reaching, is gone. And the moment you feel the cost, you can see the thing it was paying to hide.</p><p>That cushion was never a safety net you designed. It was hiding a decision you did make, without ever calling it one. You wired one vendor’s newest model into everything and gave it nothing to fall back on, and as long as reaching for it was cheap, you never had to notice. Take the cushion away and the shape is plain. Your company runs on a single model, from a single vendor, with nothing behind it. That is not a billing problem. It is the RAID array all over again, at the scale of your whole business.</p><h3>You already know this bet. You call it the cloud</h3><p>If that sounds abstract, you have lived a version of it more recently than you think.</p><p>Amazon builds serious redundancy into AWS. Multiple regions, multiple data centers, the whole thing engineered so a failure in one place does not take you down. And most companies do not pay for it. Running across two regions instead of one costs on the order of thirty percent more, so teams run in a single region and bet that AWS will stay up. The bet is good almost every single day. On October 20 last year it was not. A failure in a single AWS region in Northern Virginia cascaded for hours, and Delta, Uber, Starbucks, Spotify, Shopify, Etsy, and Amazon’s own Alexa went down with it. Companies that had quietly decided a second region was not worth the money spent that day finding out what it was worth.</p><p>And it is not only Amazon. When one bad software update went out from a single security vendor, CrowdStrike, in 2024, it took down about a quarter of the Fortune 500 at once and cost them, by the analytics firm Parametrix’s estimate, about five and a half billion dollars in direct losses within days. Delta alone put its own share near three hundred and eighty million. One vendor. One dependency. No fallback. Every company caught in it had made the same quiet trade the mail server made, the same one the single-region shops made. This is reliable enough that paying for a backup is not worth it. They were right for years and wrong for a day, and the day was catastrophic.</p><p>I have been on the cleanup side of failures shaped exactly like this, and the recovery is never proportional to the mistake that caused it. At one point in my career I watched a failover environment take forty-five days to fully restore, forty-five days, and the only reason it was survivable is that it was not production. It was slow for the same reason the mail server was fragile and the single region was exposed.</p><blockquote>The redundancy that would have made recovery fast was the line item someone decided not to pay for.</blockquote><p>That is the bet you made with your AI. You wired the best model into everything, from one vendor, with no backup, because the best model was free inside the plan and a backup felt like a cost you did not need. Anthropic just pulled the cushion. The odds on your bet did not change. What changed is that you can finally see it.</p><h3>Leah</h3><p>You know Leah. She is the engineer I keep meeting in different companies. Sharp, quietly trusted, the one who wires things together while everyone else is in a meeting about it. A year ago Leah pointed the whole company at the best model, because it was the best and it was free inside the plan, so why would she use anything else. Nobody decided that. Nobody revisited it. What she actually did, without anyone signing off, was make one vendor’s newest model load-bearing for the entire business. The invoice is what hurt first when the plan changed. The part that should keep her up is quieter. If that model is down next Tuesday, a lot of things stop, and there is no plan for that.</p><p>So the easy story is that Anthropic got greedy and you should switch. That is comfortable, and it is wrong twice over. Blaming greed misreads what is happening. And switching vendors is not a plan, it just moves you to a different single point of failure. You can wire an automatic fallback to the same vendor’s older model, and you should, but a backup that lives inside your one vendor is only half a backup. The real exposure was never the price. It is that you built something you cannot run without on top of something you do not control, and you never gave it a way to keep going when that one thing does not.</p><h3>This is not an Anthropic problem</h3><p>It would be easy to make this about one company’s pricing page. It is not, and this is the part that should change how you read the whole category.</p><p>GitHub moved Copilot to usage-based billing this summer, the same shape of change, and some developers watched their monthly bill climb by an order of magnitude. The pattern is everywhere because the cause is not a boardroom deciding to squeeze you. It is physical. The compute these models run on is genuinely scarce, and the venture money that paid for the early all-you-can-eat plans is running out. Sam Altman has said plainly that OpenAI loses money on its most expensive plan. Scarce things get priced like scarce things. Fable 5 moving to a meter is not the surprise. It is the moment the pattern got too big to look away from.</p><p>And here is the twist that makes the bill climb even as tokens get cheaper, because tokens really are getting cheaper. A fixed level of capability falls in price every year, and last year’s frontier is now a commodity you can buy for a few cents. So why does the bill keep rising? Not because tokens cost more. Because you are burning far more of them. An agent calls the model ten or twenty times for one task, a system packs more context into every call, an assistant runs all night. Cheaper tokens, many more of them, every one pointed by default at the most expensive model. That is how a rounding error becomes the line item your CFO circles in red.</p><h3>The fix, and the honest part</h3><p>Here is the part you can copy this week, without a budget and without buying anything.</p><p>Three decisions, none of which require a tool. First, decide what earns the best model, by kind of work and not by habit. Hard reasoning, high-stakes output, the thing a client sees, that can earn the top model. Volume, routine, internal, throwaway, that runs on something cheap and fast. This is the cost choice the plan used to make for you. Now you make it. Second, give anything you cannot run without a backup. Name the cheaper model that carries the load when the best one is gone, and decide in advance what slows down and what must not stop. Your vendor did this for you invisibly, and this month it stopped. Bring your own. Third, give the whole thing an owner. Right now the model choice, the fallback, and the AI budget belong to everyone and no one, which is why the bill surprised you, and the outage will too. One person owns it and writes it down. In a five-person company that owner is you, and that is fine. The point is that someone decided, instead of a default deciding for you.</p><p>Now the honest part, because you can probably smell where this is going. The layer that decides which model runs each task, keeps a fallback ready, and holds the budget, that is exactly what we build at Swa. I am not going to turn this into a pitch. I will just give you the number, because it is public and my co-founder already shared it. Across our own 90 days: 26 people, 106 million tokens. We routed the routine work to cheaper models and saved the best one for the tasks that needed it. Measured against the cost of sending all 106 million tokens to the top model, that cut our compute spend by 51 percent. It is not a magic figure, and it is not our discovery. Independent work on the same technique, the RouteLLM research out of UC Berkeley, reports cuts in that range and larger, at near equal quality. None of it is our invention. The part most people skip is measuring it honestly, and you can start on a whiteboard this afternoon with the models you already pay for. The discipline is the product. Everything else is plumbing.</p><h3>What the fix looks like, concretely</h3><p>Take a support inbox pointed entirely at the best model, because that was the default and it came with the plan. A few thousand tickets a month, and almost none of them hard. Route by the shape of the ticket. Password resets, order status, where is my refund, the routine eighty percent that follow a known pattern, go to a cheap, fast model. Anything the cheap model flags as low confidence, and anything carrying a churn signal or a legal word, gets handed up to the best model. The trigger is a confidence score and a short red-flag list, not a person sorting mail.</p><p>Do that and most of the volume leaves the expensive model, the spend on that inbox falls to a fraction of what it was, and the customer never notices, because the tickets you moved were always going to be easy. And here is what you also just built. The day the best model is unavailable or over budget, that inbox does not stop. The routine work is already on the cheaper model, and the handful of hard tickets can wait a few minutes or fall back to an honest hold instead of a broken page. Cutting the bill was only half of it. You also removed a single point of failure. The cost win and the reliability win were the same move, made once, on purpose.</p><blockquote>Every single point of failure I have ever cleaned up existed for the same reason. It was cheaper, right up until it was not. The frontier model in your stack is the newest one, and this month you found out it has no backup.</blockquote><h3>One thing to do this week</h3><p>Find the one workflow that would break tomorrow if the best model simply vanished. Not slow down. Break. That is your single point of failure, and until this month you did not have to see it, because the plan was quietly covering for you. Now put a number on it. What does one day of that workflow being down cost you, in lost orders and idle people, and how long would you honestly take to recover? My failover took forty-five days. Multiply the daily cost by that recovery window and you are looking at the real size of the bet, the one nobody wrote down. Then give the workflow the two things it does not have. A cheaper model that can carry the load, and one person who owns the choice. If it is also the workflow burning the most on frontier prices for easy work, and it usually is, you just fixed the bill and the outage with a single decision.</p><p>Leah did not need a better model or a different vendor. She needed the two things any reliability engineer would have given the system on day one. A backup, and someone who owns the choice. Neither do you. The best model is still there when you need it. What just stopped being optional is having a plan for when it is not. I learned that at twenty-something, a month deep in a CEO’s lost email, and the lesson has not changed once in twenty years. Only the size of the thing you are betting.</p><h3>ReferencesReferences</h3><ol><li>Claude Fable 5 and Mythos 5 introduced; standard pricing $10 per million input tokens and $50 per million output; launched June 9, 2026. <a href="https://www.anthropic.com/news/claude-fable-5-mythos-5">Anthropic, June 2026</a>.</li><li>Claude Opus 4.8 standard pricing is $5 per million input and $25 per million output, half of Fable 5’s rate. <a href="https://www.finout.io/blog/claude-opus-4.8-pricing-2026-everything-you-need-to-know">Finout, 2026</a>.</li><li>On June 12, 2026, a US government export control directive citing national security forced Anthropic to disable Fable 5 and Mythos 5 for all customers worldwide, over a narrow demonstrated jailbreak; Anthropic disputed the action. <a href="https://www.anthropic.com/news/fable-mythos-access">Anthropic, June 2026</a>; <a href="https://www.forbes.com/sites/anishasircar/2026/06/16/anthropic-disabled-fable-5-and-mythos-5-after-a-us-export-control-order-heres-what-happened/">Forbes, 2026</a>.</li><li>The Commerce Department cleared Fable 5 and Anthropic restored access on July 1, 2026, included in the paid plans through July 7 before moving to usage credits. <a href="https://www.anthropic.com/news/redeploying-fable-5">Anthropic, July 2026</a>.</li><li>Fable 5 moves from bundled subscription access to metered, pay-as-you-go usage credits on July 7. <a href="https://thenewstack.io/how-anthropic-is-bringing-fable-5-back/">The New Stack, July 2026</a>.</li><li>The AWS us-east-1 outage of October 20, 2025 cascaded for hours and disrupted Delta, Uber, Starbucks, Spotify, Shopify, Etsy, and Amazon’s own services; running across multiple regions typically costs roughly 30 percent more than a single region, which is why many companies do not. <a href="https://builtin.com/articles/aws-outage-what-happened">Built In and AWS Well-Architected Framework, 2025–2026</a>.</li><li>The July 2024 CrowdStrike outage, from a single vendor’s flawed update, hit about a quarter of the Fortune 500; the analytics firm Parametrix estimated roughly $5.4 billion in direct losses across those companies within days, and Delta put its own loss near $380 million. <a href="https://fortune.com/2024/08/03/crowdstrike-outage-fortune-500-companies-5-4-billion-damages-uninsured-losses/">Parametrix and Fortune, 2024</a>.</li><li>GitHub Copilot moved to usage-based billing on June 1, 2026, and removed the automatic fallback to a lower-cost model when credits run out; some developers reported large increases (one from about $29 to nearly $750 a month). <a href="https://github.blog/news-insights/company-news/github-copilot-is-moving-to-usage-based-billing/">GitHub Blog and TechCrunch, 2026</a>.</li><li>Sam Altman: OpenAI is losing money on its $200-per-month ChatGPT Pro plan because usage ran far higher than expected. <a href="https://fortune.com/2025/01/07/sam-altman-openai-chatgpt-pro-subscription-losing-money-tech/">Fortune, January 2025</a>.</li><li>AI compute scarcity: top-chip rental prices up roughly 48 percent in two months; Bank of America projects demand outstripping supply through 2029; OpenAI’s CFO cites hard trade-offs “because we don’t have enough compute.” <a href="https://tomtunguz.com/ai-compute-crisis-2026/">Tom Tunguz, citing The Wall Street Journal, 2026</a>.</li><li>Swa’s own 90-day usage across 26 users and 106.2M tokens: 70 percent of inference routed to efficient models, cutting compute cost by 51 percent versus sending every token to the top model. <a href="https://lnkd.in/p/gMYpaGHS">Shared publicly by Swa co-founder and CEO Mike Sirchuk, 2026</a>.</li><li>RouteLLM (Ong et al., UC Berkeley and Anyscale, 2024, presented at ICLR 2025): trained routers matched roughly 95 percent of GPT-4 quality while cutting cost by up to about 85 percent on benchmark tests; production deployments of model routing commonly report 40 to 70 percent reductions. <a href="https://lmsys.org/blog/2024-07-01-routellm/">LMSYS</a>.</li></ol><h3>About the author</h3><p><a href="http://benpickett.com/read">Ben Pickett</a> is Co-Founder and Chief Operating Officer of Swa and former Global Director of Site Reliability Engineering at Nike, where his remit included data protection and database operations. Swa brings the top AI models into the tools your team already uses, Slack, Teams, and more, with one interface across every model, your data never used to train them, and one simple price with no per-user fees. <a href="http://swa-ai.com/capability">Swa Solo</a>, the version for individuals and small teams to try Swa before bringing it to work, is live now at <a href="https://swa-ai.com/">swa-ai.com</a>.</p><p><em>Header artwork generated with Swa</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=dfb9ebbca224" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Ford Rehired the People AI Was Supposed to Replace]]></title>
            <link>https://medium.com/@goshawk_55282/ford-rehired-the-people-ai-was-supposed-to-replace-7208bb727ab4?source=rss-d30c86172853------2</link>
            <guid isPermaLink="false">https://medium.com/p/7208bb727ab4</guid>
            <category><![CDATA[ai-governance]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[technology]]></category>
            <category><![CDATA[agentic-ai]]></category>
            <category><![CDATA[security]]></category>
            <dc:creator><![CDATA[Ben Pickett]]></dc:creator>
            <pubDate>Thu, 02 Jul 2026 15:08:47 GMT</pubDate>
            <atom:updated>2026-07-02T15:08:47.327Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*mZx7FQdyEYyCtk88ey4yQw.jpeg" /><figcaption>AI was supposed to replace him. Ford paid to bring him back</figcaption></figure><h3>Enterprise Field Notes · Issue #11</h3><p><em>Ford hired 350 veteran engineers after its AI fell short, then topped J.D. Power. Why pairing AI with your experts beats replacing them, from Polanyi to Kasparov to your own factory floor.</em></p><p>By Ben Pickett</p><p>I sell AI for a living. So, when Ford hired 350 veteran engineers because its AI was not good enough, that is my story to explain, not Ford’s to bury.</p><p>Here is the part that should stop you. A year ago, Ford’s CEO stood on a stage in Aspen and said AI would replace “literally half of all white-collar workers in the U.S.” This year, Ford went looking for the humans.</p><p>You know a Ray. Thirty years on the job, the one who can hear a bad weld from across the bay and knows the exact gasket that fails in a Michigan February. For a decade, the org chart treated Ray like overhead. Expensive, analog, replaceable. Then the requirements went into the AI, the AI produced the designs, the designs did not hold up, and Ford went and found 350 Rays. Some retired, some at suppliers, all brought back.</p><p>They are not there to babysit the robots. They run mandatory weekly design reviews that catch failure points before a blueprint reaches the factory floor. They mentor the young engineers. They retrain the AI that kept missing things. Charles Poon, Ford’s VP of vehicle hardware engineering, said the quiet part out loud: “Mistakenly, we thought that by just introducing artificial intelligence and ingesting the design requirements that we had, that would produce a high-quality product.” Then the line to tape to the wall: <strong>“Artificial intelligence is a fantastic tool, but it’s only as good as the information you use to train it.”</strong></p><p>Around the same time, Ford was named the top mainstream brand in J.D. Power’s Initial Quality Study for the first time since 2010. I am not going to claim the engineers caused that. Initial-quality scores mostly measure first-90-day gripes, not the durability failures these veterans were brought in to catch, and the timelines do not line up that cleanly. There was a hard quality push underneath it too, on the back of a punishing stretch of recalls. Read the award as reputational backdrop, not proof. The proof is simpler, and it is in Ford’s checkbook: they paid real money to put veterans back in the room.</p><h3>The assumption failed, not the AI</h3><p>Read it fast and it sounds like AI lost. It did not. <strong>The assumption lost.</strong> The assumption that Ray’s thirty years lived inside the requirements document.</p><p>They never did. Sixty years ago the philosopher Michael Polanyi put it in five words: “We know more than we can tell.” Economists now call it Polanyi’s Paradox, and it is why automation keeps hitting a wall. The most valuable thing Ray knows, he cannot fully write down, which means you cannot fully feed it to a model by handing it a spec. The knowledge lives in the doing, in ten thousand small judgments he makes without narrating them. You get it by putting Ray next to the machine, not by asking Ray to type it up.</p><h3>We already learned this, in chess</h3><p>After a computer beat Garry Kasparov, he did not quit. He invented a new game where humans and machines play together. In a 2005 open tournament, the winners were not a grandmaster, and not a supercomputer. They were two amateurs from New Hampshire running three ordinary chess programs and a better process for using them. They beat the grandmasters. They beat the supercomputers. Kasparov’s takeaway: <strong>“a weak human plus a machine plus a better process was superior to a strong computer alone, and superior to a strong human plus a machine with an inferior process.”</strong></p><p>Ford just rediscovered centaur chess on an assembly line. The win was never the model. It was the pairing, plus the process.</p><p>Yes, engines eventually got strong enough to beat the centaurs too. But chess is a closed game with fixed rules and perfect information. Your business is not. It is messy, changing, and full of judgment calls no model has trained on, which is exactly the world where the pairing still wins.</p><h3>It is not just a game</h3><p>The same result shows up where the stakes are life and death. Across studies in radiology, a doctor reading scans with AI catches more than the AI alone and more than the doctor alone. But there is a catch that proves the whole point: it only works when the human stays in charge. When radiologists lean on the AI and stop thinking, the pairing gets worse, not better. That is Kasparov’s finding restated in a hospital. The pairing wins, and the process, who leads whom, is the entire game.</p><h3>Why this is urgent, not philosophy</h3><p>About one in four American manufacturing workers is over 55. In one industry survey, most manufacturing leaders said they expect to lose at least half of their institutional knowledge as their people retire. Your Rays are walking out the door on a clock.</p><p>So when you point AI at their jobs to cut cost, you are not trimming overhead. You are setting fire to the one thing your competitors cannot download.</p><p>I spent most of my career in reliability engineering, a lot of it at Nike, at a scale where one bad change could take the whole system down. The rule we lived by: when something broke, you did not blame the engineer, you fixed the system that let it through. The system breaking across corporate America right now is the one that treats its Rays as a line item and its AI as a replacement. Both halves of that are wrong.</p><h3>Run the centaur play</h3><ul><li><strong>Capture judgment before the exit interview, not after.</strong> Turn design reviews, post-mortems, and defect analyses into records the model can learn from. Poon’s most expensive sentence was that some of Ford’s best engineers left before their knowledge was ever captured.</li><li><strong>Put your experts on the model, not around it.</strong> Not the fallback for when it breaks. The people who review it, catch what it missed, and log every correction. Concretely: a veteran flags in a design review that a bracket will crack under thermal cycling. That correction, with his reasoning, becomes a labeled example the next version of the model trains on. Enough of those and the model stops making the mistake. The corrections are the asset. A model that learns from your best people compounds. A model that learns from a spec plateaus.</li><li><strong>Pay for the process honestly, and keep paying the people.</strong> It is slower than “just let the AI do it,” and your best people will resent training their supposed replacement. So do not drain the Ray and move on. The person who trains your model should be the best-retained, best-paid, most-mentored hire in the building, not the one you optimize out next quarter. Name that cost, or the whole thing stalls.</li></ul><p>This is not a Ford quirk. Klarna handed hundreds of service jobs to AI, watched quality slide, and rebuilt a human-plus-AI model. Gartner predicts that by 2027, half the companies that cut customer-service staff citing AI will rehire, often for the same work under new titles.</p><blockquote>Treat AI as a replacement for expertise and you get a cheaper version of worse. Treat it as an amplifier and the expertise you already paid for becomes the one thing your competitors cannot copy.</blockquote><h3>One thing to do this week</h3><p>We are reading this the week of America’s 250th, so one line of optimism, the honest kind: every wave this country rode created more work than it destroyed, but never smoothly and never for everyone, and who came out ahead was a leadership choice about whose knowledge got kept. Adaptation is not a virtue the displaced failed at. It is a path leaders build, or do not.</p><p>So, this week: find the retirement closest on your calendar, the veteran whose knowledge would hurt the most to lose. Start with one team, one process, not the whole floor. Put that person on the model, not around it, and log what they correct. That is the whole play, begun.</p><p>Ford spent a decade treating its Rays like overhead, then paid to bring 350 of them back. Your Ray is still in the building. For now.</p><p>Who is your Ray, and are you capturing what they know before the exit interview? I read every reply.</p><p><em>Header artwork generated with Swa.</em></p><h3>References</h3><ol><li>Ford on why it hired 350 “gray beard” engineers: you need their mentorship, and to drive huge AI productivity gains. Fortune, June 29 2026. <a href="https://fortune.com/2026/06/29/ford-ai-hired-human-workers-gray-beards-automation/">link</a></li><li>Ford CEO Jim Farley: AI will replace half of all white-collar workers (remark at the Aspen Ideas Festival, June 27 2025). Fortune, July 5 2025. <a href="https://fortune.com/2025/07/05/ford-ceo-jim-farley-ai-white-collar-jobs-essential-economy-skilled-trade-jobs-shortage/">link</a></li><li>Ford Named Top Mainstream Brand for New Vehicle Quality in J.D. Power 2026 U.S. Initial Quality Study for First Time Since 2010. BusinessWire / Ford, June 25 2026. <a href="https://www.businesswire.com/news/home/20260625317685/en/Ford-Named-Top-Mainstream-Brand-for-New-Vehicle-Quality-in-JD-Power-2026-U.S.-Initial-Quality-Study-for-First-Time-Since-2010">link</a></li><li>Michael Polanyi, “The Tacit Dimension” (1966); David Autor, “Polanyi’s Paradox and the Shape of Employment Growth” (NBER working paper 20485, 2014). <a href="https://www.nber.org/papers/w20485">link</a></li><li>Garry Kasparov on “centaur” / advanced chess and the 2005 PAL/CSS Freestyle tournament won by amateurs Steven Cramton and Zackary Stephen (“ZackS”), from “The Chess Master and the Computer” (New York Review of Books, 2010). <a href="https://www.nybooks.com/articles/2010/02/11/the-chess-master-and-the-computer/">link</a></li><li>Human plus AI in diagnostic imaging can outperform either alone, but degrades under automation bias when clinicians defer to the model (Tschandl et al., Nature Medicine, 2020). <a href="https://www.nature.com/articles/s41591-020-0942-0">link</a></li><li>Gartner Predicts Half of Companies That Cut Customer Service Staff Due to AI Will Rehire by 2027. Gartner, February 3 2026. <a href="https://www.gartner.com/en/newsroom/press-releases/2026-02-03-gartner-predicts-half-of-companies-that-cut-customer-service-staff-due-to-ai-will-rehire-by-2027">link</a></li><li>Klarna CEO Reverses Course by Hiring More Humans, Not AI. Entrepreneur, 2025. <a href="https://www.entrepreneur.com/business-news/klarna-ceo-reverses-course-by-hiring-more-humans-not-ai/491396">link</a></li><li>Manufacturing workforce aging and “brain drain”: retirement and institutional-knowledge loss (Deloitte and The Manufacturing Institute skills-gap research). <a href="https://www2.deloitte.com/us/en/insights/industry/manufacturing.html">link</a></li></ol><p><em>Ben Pickett is Co-Founder and COO of Swa and former Global Director of SRE at Nike. Enterprise Field Notes is his weekly read on what actually works in enterprise AI. More from Ben, and the tools behind these ideas, at </em><a href="https://benpickett.com/?utm_source=medium&amp;utm_medium=efn_bio&amp;utm_campaign=efn11"><em>benpickett.com</em></a><em>.</em></p><p>#EnterpriseAI #AIStrategy #FutureOfWork #Leadership</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=7208bb727ab4" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[KPMG Forgot to Check Its Own Work]]></title>
            <link>https://medium.com/@goshawk_55282/kpmg-forgot-to-check-its-own-work-dcdb0ebffd31?source=rss-d30c86172853------2</link>
            <guid isPermaLink="false">https://medium.com/p/dcdb0ebffd31</guid>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[ai-governance]]></category>
            <category><![CDATA[security]]></category>
            <category><![CDATA[agentic-ai]]></category>
            <category><![CDATA[technology]]></category>
            <dc:creator><![CDATA[Ben Pickett]]></dc:creator>
            <pubDate>Thu, 25 Jun 2026 14:51:57 GMT</pubDate>
            <atom:updated>2026-06-25T14:51:57.973Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*ifPgEBYd87NNQdZTQOZo_Q.jpeg" /><figcaption>The last read, before it leaves the building. Header artwork generated with Swa.</figcaption></figure><p><em>One of the most trusted firms on earth shipped a report with 45 sources. Only five checked out. Here is the cheap fix that would have caught it, and why your team needs it more than KPMG did.</em></p><p>By Ben Pickett</p><p>I sell AI for a living. So when KPMG published a report and 40 of its 45 sources did not check out, that is my problem to explain, not just theirs.</p><p>Here is what happened. KPMG, one of the Big Four, the firm the rest of the economy pays to check its work, put out a report called “Total Experience: Redefining Excellence in the Age of Agentic AI.” A flagship piece of thought leadership, the kind of report that exists to make you trust the firm. This month they pulled it off their own websites. An AI-detection firm had gone through the citations and found that only five of the forty-five pointed to a real, intact source. The other forty were wrong or invented. Some were real-sounding titles attached to the wrong author. Some were stitched together out of fragments. Some did not exist at all.</p><p>There is a name for this now. Vibe citing. The AI assembles things that sound right, in the confident cadence of a real citation, and nobody checks. The firm that ran the review said it plainly. Fabrications like this “poison the well of information.” That is the part that should bother you more than the embarrassment. A made-up source does not look made up. It looks like every other source on the page.</p><p>It gets worse. The report described how real, named organizations were using AI, and four of them told the Financial Times the claims were not true. UBS said the parts about its risk and compliance work were factually incorrect. Swiss Federal Railways said the journey-planning claim was not accurate. Transport for London called the congestion-prediction claim misleading. A regional NHS body in Greater Manchester said the bit about predicting hospital readmissions did not align with anything it had done. Four real institutions, on the record, saying a report about them was fiction. Not a vague report. A report with their names in it, about their own work.</p><p>You know a Dana. Senior, sharp, trusted, and buried. Last quarter she let an AI-drafted analysis go out under her name because it read clean and the deadline was real. It was mostly right. Mostly is the word that ends up in the apology email. She did not have a bad model. She had no last set of eyes between the draft and the client. Neither did KPMG.</p><p>So the easy story is that KPMG used a careless AI and a better one would have saved them. That story is wrong. Believing it is what puts your team next on this list.</p><h3>A bigger model would not have fixed this</h3><p>These tools are confidently wrong by design. They hand you a citation that looks perfect and is pure fiction, in the same calm voice they use for the things that are true. There is no tell in the text. The fake source is formatted exactly like the real one. That is not a bug that gets patched in the next release. It is what the tool is, and a more powerful model is often more convincing, not more honest.</p><p>Read KPMG’s own statement and they tell you where it actually broke. They said they expect their people to follow the rules on responsible AI, “including human oversight to validate content and verify independent sources.” That sentence is the whole story. Somewhere between the draft and the website, the person who checks the claims was gone. That is not a model defect. That is a missing step in a workflow. You cannot buy your way out of a missing step with a smarter model. You can only put the step back.</p><h3>This is not a KPMG problem</h3><p>It would be comfortable to file this under one firm having a bad month. The numbers say otherwise. A survey of more than two thousand senior decision makers this year found that 74 percent of enterprises have already pulled a live customer-facing AI agent back out of production after a governance failure. The rollback rate was higher, 81 percent, at the organizations with the most mature governance. Read that twice. The teams paying the closest attention are catching the most failures, because they are the only ones looking. The ones not seeing failures are not automatically safer. Often they are just not looking as closely.</p><p>The deeper research lands in the same place. MIT’s NANDA initiative found that the large majority of enterprise AI pilots deliver little to no measurable impact on the bottom line, and the reason was not model quality. It was that generic tools do not bend to the way the work actually flows. The exact figure gets argued. The direction does not. The model is rarely what breaks. The seam between the model and the workflow is what breaks. KPMG is that seam failing in public, with a logo everyone recognizes on it. Your version will not make the news. It will just go out to a client.</p><h3>I spent years building the step they skipped</h3><p>Before Swa, I spent most of my career in engineering, a lot of it leading large-scale reliability platforms at Nike. In that world you never let a change reach production without a gate in front of it. Not because your people are careless. Because the system is confidently wrong sometimes, and the gate is the cheap place you catch it before it gets expensive.</p><p>The other half of that discipline matters more here. When something slipped through, we did not write up the engineer who made the call. We ran a blameless review and fixed the process that let it through. You cannot fix people. People are tired, rushed, and trusting on a Friday afternoon, and they always will be. You can fix the system around them so that being tired on a Friday does not become a fabricated client report. That is why this is not a piece about KPMG being sloppy. Smart people at a serious firm shipped this. The system around them was missing one step, and the system is the thing you can actually fix.</p><p>AI took confidently wrong and moved it into every report, deck, and memo your company sends. All of it now has a fast, fluent first author that makes things up. If you quietly dropped the last gate to move faster, you did the exact thing KPMG did. You just have not been caught in public yet.</p><h3>What this looks like when it works</h3><p>Picture a team that produces client research. Before AI, a junior analyst spent two days on a draft, and a senior reviewer read the whole thing because there was time and the draft was short. Now AI produces a polished forty-page draft in an hour. It looks finished. It reads better than the junior analyst’s did. And the senior reviewer, faced with forty fluent pages and the same calendar, skims it, because skimming something that already looks done is human nature.</p><p>That is the trap. The output got faster and more convincing at the exact moment the review got shallower. The gate did not formally go away. It quietly stopped catching anything.</p><p>The fix is not to read all forty pages again. That just rebuilds the slow process AI was supposed to replace, and the team will route around it inside a week. The fix is to decide, in advance, which claims in those forty pages are load-bearing, and to check only those, every time, no exceptions.</p><h3>The fix is one gate, on the claims that matter</h3><p>Here is the part you can copy without a Big Four budget.</p><p>Here is the objection, and it is the right one. KPMG had reviewers too. A gate that skims forty fluent pages on deadline is the gate that just failed. So the fix is not one more reviewer. It is naming the load-bearing claims in advance, so a tired reviewer cannot skim past them. The numbers. The citations. Anything you say a named outside party did. The handful of things that can end a client relationship or a career if they are wrong. Those get a human signature before they leave the building. Everything else runs at full speed.</p><p>Naming those claims is the real work, and it is worth doing slowly. Most teams have never written down which of their outputs are load-bearing. They review by vibes, more carefully when they are nervous and less when they are busy, which is exactly backward. Sit down for half a day and write the list. For a client report it might be five things: every external statistic, every quote, every claim about a named company, every dollar figure, and any legal or compliance assertion. That list is your gate. The reviewer is not re-reading the document. They are confirming those five categories, and signing their name to them.</p><p>That is the whole move. The AI still drafts, and it is fast. A person still owns the last gate on the claims that carry weight. KPMG had the budget for a hundred reviewers and skipped the one that mattered. You do not need a hundred. You need one, on the right claims.</p><blockquote>KPMG did not lose its credibility to a hallucination. It lost it to a missing step that costs almost nothing to add back.</blockquote><h3>One thing to do this week</h3><p>Pick the one workflow where AI-drafted work leaves your company under someone’s name. A client deliverable. A board number. A published claim about somebody else. Walk it and find the exact point where the output goes out the door. Ask who reads it last, and what they are on the hook to catch. If the answer is nobody, or everybody, you just found the gap KPMG fell through.</p><p>Then put one person on the load-bearing claims. Name them out loud. Do it before the deadline picks for you.</p><p>Dana never needed a smarter model. She needed one name on one short list, written down before the deadline. Neither did KPMG. Neither do you.</p><p><em>Header artwork generated with Swa.</em></p><h3>References</h3><ol><li>KPMG pulls report on AI usage due to apparent hallucinations. TechCrunch, June 13 2026. <a href="https://techcrunch.com/2026/06/13/kpmg-pulls-report-on-ai-usage-due-to-apparent-hallucinations/">link</a></li><li>Investigation: only 5 of 45 citations in the KPMG report pointed to a real source; 40 fabricated; “vibe citing”; “poison the well of information.” GPTZero, 2026. <a href="https://gptzero.me/news/investigations-kpmg/">link</a></li><li>KPMG drops AI report after false case studies exposed (UBS, NHS Greater Manchester, Swiss Federal Railways, Transport for London dispute claims). International Accounting Bulletin / Financial Times, June 2026. <a href="https://www.internationalaccountingbulletin.com/news/kpmg-drops-ai-report-after-false-case-studies-exposed/">link</a></li><li>74 percent of enterprises have rolled back a live AI customer-communications agent after a governance failure; 81 percent among the most mature-governance organizations. Sinch (commissioned survey of 2,527 senior decision makers), May 2026. <a href="https://www.prnewswire.com/news-releases/sinch-research-reveals-74-of-enterprises-have-rolled-back-live-ai-customer-communications-agents-302770730.html">link</a></li><li>The large majority of enterprise generative-AI pilots deliver little to no measurable P&amp;L impact, because generic tools do not adapt to enterprise workflows. MIT Project NANDA, “The GenAI Divide: State of AI in Business 2025.” <a href="https://fortune.com/2025/08/18/mit-report-95-percent-generative-ai-pilots-at-companies-failing-cfo/">link</a></li></ol><p><em>Ben Pickett is Co-Founder and Chief Operating Officer of Swa and former Global Director of Site Reliability Engineering at Nike. Swa delivers enterprise AI inside Slack, Teams, WhatsApp, Zoom, SMS, and more channels coming, integrated into the systems behind them through user-scoped auth. One interface across many models. Zero data retention by default. Flat-rate pricing. Deployment in under three minutes. </em><a href="http://solo.swa-ai.com/"><em>Swa Solo</em></a><em>, our first version for individuals and small teams to try Swa before bringing it to the enterprise, ships soon. </em><a href="https://swa-ai.com/"><em>swa-ai.com</em></a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=dcdb0ebffd31" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The SpaceX IPO Is a Bet on Your AI Bill]]></title>
            <link>https://medium.com/@goshawk_55282/the-spacex-ipo-is-a-bet-on-your-ai-bill-af89c64da54b?source=rss-d30c86172853------2</link>
            <guid isPermaLink="false">https://medium.com/p/af89c64da54b</guid>
            <category><![CDATA[enterprise-ai]]></category>
            <category><![CDATA[sustainability]]></category>
            <category><![CDATA[cto]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[technology]]></category>
            <dc:creator><![CDATA[Ben Pickett]]></dc:creator>
            <pubDate>Thu, 18 Jun 2026 15:05:02 GMT</pubDate>
            <atom:updated>2026-06-18T15:05:02.051Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*cdSDTjbNfyWq2pq6LDH2kw.jpeg" /><figcaption><em>The river that runs the data centers is the one a lot of us want still running for the next generation.</em></figcaption></figure><p><em>Intelligence has never been cheaper. So why is your AI bill going up? The fix is not a cheaper model. It is knowing when to stop reaching for the biggest one. And it turns out to be the same move that decides what all of this costs the rest of us.</em></p><p>By Ben Pickett</p><p>Last week, SpaceX went public in the biggest IPO in history. Seventy-five billion dollars raised, a valuation that blew past two trillion within days. A rocket company. So why could I, someone who writes about enterprise AI for a living, not look away?</p><p>But one number stuck with me. The same week of that IPO, Google agreed to pay SpaceX roughly 920 million dollars a month for access to about 110,000 graphics chips, because even Google, sitting on more money and more data centers than almost anyone alive, cannot get its hands on computing power fast enough to keep up with demand for its own AI. A rocket company is suddenly one of the places the biggest names in tech turn just to find compute. The biggest IPO in history is, in part, a bet on how desperate the world has become for AI horsepower.</p><p>Everyone is watching the stock. I want to talk about that desperation instead, because it is the same scarcity you pay into every time you reach for the biggest AI model when a smaller one would have done the job. It lands on your desk as a bill that keeps climbing.</p><p>Now the strange part. While the giants fight over hardware, the price of the intelligence itself is falling off a cliff. Epoch AI, which tracks this closely, puts the cost of GPT-4-level answers at roughly ten times cheaper every year. About 20 dollars per million tokens in late 2022, closer to 40 cents today. The exact slope depends on the task, and nothing I am about to say hinges on the precise number.</p><p>So intelligence is getting dramatically cheaper. And yet almost every company I talk to is watching its AI bill go up, not down.</p><p>That is the paradox I want to unpack. The reason your bill is climbing is not the price of the model. It is how many times you reach for the most expensive one. And the move that pulls the bill back down turns out to be the same move that makes you a good steward of something I care about a great deal. By the end you will have one question to ask your own team this week.</p><h3>The trap is not the price. It is the habit.</h3><p>The instinct everyone has is to shop. Prices are falling, so wait for the next cheaper model, or hunt for the cheapest tokens on the market. Wrong lever.</p><p>The real driver is simpler. We stopped asking AI single questions and started handing it whole jobs. An agent does not make one model call to finish a task. It makes ten or twenty. Now multiply that by a habit almost everyone has fallen into without noticing: defaulting to the biggest, most powerful model for every step, even the trivial ones. Classifying an email. Pulling a date out of a document. Deciding whether a message is a question or a complaint.</p><p>Cheaper tokens, times far more calls, times the most expensive model on every call. That does not add up to a smaller bill. The price of intelligence went down and the spend on it went up. This was never a pricing problem, which is why you cannot shop your way out of it. It is an architecture problem, and you cannot shop your way out of an architecture problem.</p><p>Economists even have a name for the trap. Jevons paradox. When a resource gets cheaper to use, we use so much more of it that the total bill climbs instead of falling. Cheaper steam engines burned more coal, not less. Cheaper tokens spend more money, not less.</p><blockquote>Efficiency per call is not the same as efficiency on the invoice.</blockquote><h3>I learned this years before the chatbots showed up</h3><p>Before Swa, I spent most of my career in engineering, a lot of it leading large-scale data and reliability platforms at Nike. And I learned the same lesson over and over, long before anyone was typing into a chat box.</p><p>You do not put your most powerful, most expensive infrastructure behind every job. The heavy, costly setup is for the work that genuinely needs it. Everywhere else, you use something lighter and cheaper, and you automate the parts people used to do by hand. The teams that got this right did not just spend less. They moved faster, because they were not dragging the heavy machinery into work that never required it, and they freed their best people to sit on the decisions that actually mattered.</p><p>That is the whole game with AI today, just wearing a new costume. Most of what you ask AI to do is routine. Classifying an email. Pulling a date out of a document. Deciding whether a message is a question or a complaint. None of that needs the most powerful model on the market. It needs a small, fast, capable one. You reserve the frontier model for the calls that genuinely earn it, and you keep a person on the decisions that carry real weight.</p><p>The industry has a name for this. Right-sizing, or model routing. One 2025 study estimated that choosing the right-sized model for each task could cut the energy behind AI inference by roughly a quarter, with the efficient models giving up under four percent of their quality. On mature, well-understood tasks the savings ran far higher. It is an early, single-study figure and depends heavily on the job, but the direction matches what anyone who has tuned this in production already sees.</p><p>Picture the support inbox for almost any business. A hundred messages come in. The lazy build sends every one of them, at every step, to the most powerful model on the market. Read it, classify it, draft a reply, check the reply. Four expensive calls times a hundred messages, and most of those messages are someone asking where their order is.</p><p>Now right-size it. A small, fast model reads all hundred and sorts them. The seventy that are routine, where is my order, how do I reset my password, change my address, it handles start to finish. The twenty-five that are genuinely tricky escalate to the frontier model. The five that carry real weight, a furious customer, a refund dispute, a question with legal teeth, go to a person who is now free to handle them well. Same hundred messages handled. A fraction of the bill, and arguably better service, because your best people are spending their time where it counts.</p><p>The catch, and anyone who has shipped this will tell you, is that the sorting is the hard part. You earn that split by tuning it on real traffic and setting honest confidence thresholds, not by hoping a cheap model guesses right. But that is engineering you do up front and revisit as traffic shifts. Paying premium prices to answer where is my order is a tax you pay forever.</p><p>In our own systems, over a recent ninety-day stretch across our production traffic, right-sizing the model to the task cut our compute cost by roughly half compared with routing everything to premium models. On money alone, that is the whole case.</p><h3>Why this one is personal</h3><p>But there is a second reason I care about this, and it is not about money at all.</p><p>My family came west on the Oregon Trail six generations back and homesteaded in Douglas County, in southern Oregon. My great-great-great-grandfather, James Swifty Pickett, brought his son James Riley there as a boy. We put down roots and never left. I have spent my whole life in the Pacific Northwest. I also spent years as mayor of Jefferson, Oregon, where I dealt with water rights more than I ever expected to. You learn quickly that water is not infinite, and that everyone downstream is counting on the choices you make upstream.</p><p>So none of this is abstract for me. A lot of the data centers training and running these models sit right along the Columbia River, powered in large part by the hydro and the Gorge wind that light up my corner of the country. The intelligence we are all racing to deploy runs on my river.</p><h3>The bigger bill</h3><p>Right-sizing is more than a budgeting trick. Every one of those model calls does something beyond cost you money. It pulls electricity off a real grid, water from a real river, and minerals out of a contested supply chain.</p><p>The power is the headline. The International Energy Agency projects that the electricity data centers draw will roughly double by 2030, growing more than four times faster than overall electricity demand, with AI driving almost half of that growth in data center demand.</p><p>The water is closer to home than most people realize. A peer-reviewed study estimated that training one older model, GPT-3, in US data centers could evaporate around 700,000 liters, roughly 185,000 gallons, of clean freshwater just for cooling. Run twenty to fifty everyday questions through a chatbot and you have quietly used about a bottle’s worth.</p><p>And the materials are finite and concentrated. Roughly 95 percent of the gallium that goes into this hardware is refined in a single country, which has already shown it will use that as leverage.</p><p>And before someone says the answer is to just put the data centers in space, that race is already on. Google has a feasibility project. SpaceX has filed to launch up to a million compute satellites. A startup up the road from me in the Seattle area has already trained a model and run Gemini in orbit. Free solar, no water for cooling. It is a genuinely exciting idea, and I hope some version of it works. It also costs several times what a data center on the ground does today, there are nowhere near enough rockets to do it at scale, and the serious versions are years away. The moonshot does not make the resource bill disappear. It moves it off the planet, after burning an enormous amount of resources to get there.</p><p>Which brings me back to the one lever you actually have this quarter. Right-sizing. The same move that pulls your AI bill down is also one of the most practical levers you actually control over what AI takes from the grid, the water, and the ground. You do not have to choose between running a tight operation and being a decent steward. The efficient architecture is both. This is not about saving the planet. It is about not being wasteful with what we all share.</p><h3>What this comes down to</h3><p>Cheaper tokens will not save you. Right-sizing will.</p><p>And the same decision that controls your bill eases what AI draws from the grid, the water, and the supply chain. Good operator and good steward, the same move.</p><h3>One question to ask this week</h3><p>Here is the gift, and it costs nothing.</p><p>Pull up your most-used AI workflow this week. Walk it step by step and ask one question. Which of these steps actually needs the most powerful model, and which ones are we sending there out of pure habit? Route the habits down to a smaller, efficient model. Keep your people on the judgment calls. That single pass will usually find more savings than any pricing negotiation ever will.</p><p>That is it. The question is free. What you do with the answer is the whole job.</p><p>I would like all of this to still be here, and running well, for the people who come after us. The way to get there is not a grand gesture, and it is not a rocket. It is a thousand small, smart decisions about when we actually need the biggest hammer. Start with one this week.</p><p><em>Header artwork generated with Swa.</em></p><h3>References</h3><ol><li>Google–SpaceX compute deal. Google will pay SpaceX ~$920M/month for ~110,000 Nvidia GPUs, Oct 2026–June 2029, as “bridge capacity” for Gemini demand. <a href="https://techcrunch.com/2026/06/05/google-will-pay-spacex-920m-per-month-for-compute/">TechCrunch, June 2026</a>.</li><li>Inference price decline. Epoch AI, “LLM inference prices have fallen rapidly but unequally across tasks.” GPT-4-level performance fell from ~$20 to ~$0.40 per million tokens, roughly 10x/yr. <a href="https://epoch.ai/data-insights/llm-inference-price-trends">Epoch AI</a>.</li><li>Agentic token use. Gartner (March 2026): agentic AI uses 5–30x more tokens per task than a chatbot. <a href="https://www.gartner.com/en/newsroom/press-releases/2026-03-25-gartner-predicts-that-by-2030-performing-inference-on-an-llm-with-1-trillion-parameters-will-cost-genai-providers-over-90-percent-less-than-in-2025">Gartner</a>.</li><li>Right-sizing / model selection. “Small is Sufficient: Reducing the World AI Energy Consumption Through Model Selection,” arXiv:2510.01889 (Oct 2025): model selection could cut AI inference energy ~27.8%; efficient models avg ~66% more efficient at under 4% quality loss; up to ~70–93% on mature tasks. An early, single-source, idealized scenario. <a href="https://arxiv.org/abs/2510.01889">arXiv</a>.</li><li>Data center electricity. IEA, <em>Energy and AI</em> (April 2025): data center electricity projected to roughly double to ~945 TWh by 2030, growing &gt;4x faster than total electricity demand; AI “accelerated servers” drive almost half the net increase. <a href="https://www.iea.org/reports/energy-and-ai/energy-demand-from-ai">IEA</a>.</li><li>Training water. Li, Yang, Islam &amp; Ren, “Making AI Less Thirsty,” Communications of the ACM (2025), DOI 10.1145/3724499: training GPT-3 in US data centers could directly evaporate ~700,000 liters of freshwater (modeled estimate); ~20–50 queries ≈ a half-liter bottle. <a href="https://dl.acm.org/doi/10.1145/3724499">CACM</a>.</li><li>Gallium concentration. IEA, <em>Energy and AI</em>: ~95% of gallium refining is in one country; China’s July 2023 gallium/germanium export controls. <a href="https://www.iea.org/reports/energy-and-ai/ai-and-energy-security">IEA</a>.</li><li>Data centers in space. Google “Project Suncatcher” feasibility study (Nov 2025); SpaceX FCC filing for up to 1M orbital data-center satellites; Starcloud (Seattle-area) ran a model and Gemini in orbit. Orbital compute costs several times ground today; years from scale. <a href="https://www.technologyreview.com/2026/04/03/1135073/four-things-wed-need-to-put-data-centers-in-space/">MIT Technology Review</a>.</li><li>Columbia River data centers. Google (The Dalles) and Amazon (Boardman/Hermiston/Umatilla) draw on BPA Columbia River hydropower; Columbia Gorge wind feeds the same grid. <a href="https://andthewest.stanford.edu/2025/thirsty-for-power-and-water-ai-crunching-data-centers-sprout-across-the-west/">Stanford, And the West</a>.</li></ol><p><em>Ben Pickett is Co-Founder and Chief Operating Officer of Swa and former Global Director of Site Reliability Engineering at Nike. Swa delivers enterprise AI inside Slack, Teams, WhatsApp, Zoom, SMS, and more channels coming, integrated into the systems behind them through user-scoped auth. One interface across many models. Zero data retention by default. Flat-rate pricing. Deployment in under three minutes. </em><a href="http://solo.swa-ai.com/"><em>Swa Solo</em></a><em>, our first version for individuals and small teams to try Swa before bringing it to the enterprise, ships soon. </em><a href="http://swa-ai.com/"><em>swa-ai.com</em></a></p><p><em>Header artwork generated with Swa.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=af89c64da54b" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[What Klarna Got Wrong That Walmart Got Right]]></title>
            <link>https://medium.com/@goshawk_55282/what-klarna-got-wrong-that-walmart-got-right-7f0bde4bf808?source=rss-d30c86172853------2</link>
            <guid isPermaLink="false">https://medium.com/p/7f0bde4bf808</guid>
            <category><![CDATA[enterprise-ai]]></category>
            <category><![CDATA[future-of-work]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[ai-agent]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <dc:creator><![CDATA[Ben Pickett]]></dc:creator>
            <pubDate>Thu, 11 Jun 2026 15:42:08 GMT</pubDate>
            <atom:updated>2026-06-11T15:42:08.971Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Zu6LOl6qA30ViJbSX0plNw.jpeg" /></figure><p><em>Klarna replaced 700 people with AI and walked it back. Walmart redrew the work and won. The real first move in enterprise AI is a workflow, not a model.</em></p><p>By Ben Pickett</p><p>Eighteen months later, Klarna’s CEO, Sebastian Siemiatkowski, admitted it had gone too far. Quality had cratered. The AI was fine on the routine questions and lost on the ones that needed a person: the angry customer, the edge case, the problem that did not fit the script. “We focused too much on efficiency and cost,” he said. “The result was lower quality, and that is not sustainable.” Klarna started hiring people back.</p><p>Last week I wrote about Walmart, which took the same technology and did almost the opposite. It did not point an agent at a payroll. It redrew how the work flows, decided which parts a machine should carry and which still needed a person, and built the architecture around that decision.</p><p>Same technology. Opposite move. Opposite result.</p><p>That is the whole thing. And it is the question I got more than any other after last week: I am not Walmart, I do not have the budget, so where do I start? The honest answer is smaller, and harder, than a budget. You do not start with a model, or a tool, or a headcount you want to cut. You start with one workflow, and the courage to redraw it instead of just automating it.</p><p>Let me give you three things: why almost all of this is failing right now, where the few who get real value actually begin, and the choice hidden inside the redraw that Klarna learned the hard way.</p><h3>One. Why almost all of it is failing</h3><p>Most companies do what Klarna did, just less famously. They buy a copilot or stand up an agent and point it at the work they already have. The process does not change. The handoffs do not change. They add intelligence on top and wait for the productivity to show up.</p><p>It rarely does. There is a line going around the enterprise AI world right now worth writing on a wall:</p><blockquote>Layering an agent onto a broken process gives you a faster broken process.</blockquote><p>The numbers are sobering. By 2026, eighty-eight percent of companies report using AI somewhere, and only about six percent report a real, enterprise-wide financial impact. A widely cited MIT study last year found that roughly ninety-five percent of corporate AI pilots delivered no measurable return. Hold that number loosely. It is contested, and no return yet is not failed forever. But every serious read agrees on the direction. Gartner now expects more than forty percent of agentic AI projects to be canceled by the end of 2027, and warns that of the thousands of vendors selling agents, only a small fraction are the real thing. The rest is old automation in a new costume. They have a name for it now. Agent washing.</p><p>Two findings inside that mess are worth sitting with, because they are not what you would guess. The first: spending more does not save you. The companies pouring the most into AI are not reliably the ones getting it to work, which is the good news for everyone who is not Walmart. The second: the technology is the small part. The barrier is not the model. The models work. The model is the part everyone obsesses over, and the part that least often kills the project. The barrier is that almost nobody changes the work.</p><p>We have seen this movie before. In 1987 the economist Robert Solow wrote a line that stuck: you can see the computer age everywhere but in the productivity statistics. Companies were pouring fortunes into computers and the numbers stayed flat. Then, through the 1990s, the productivity finally arrived, but not evenly. Erik Brynjolfsson and his colleagues spent years studying which firms captured the gains, and the answer was clear. It did not go to the companies that bought the most computers. It went to the ones that redesigned the work around them. Same technology, same decade, wildly different outcomes, and the difference was process redesign. This spring Brynjolfsson’s team at Stanford published the same lesson again, drawn from fifty-one enterprise AI deployments that actually worked. Thirty years apart, new technology, identical finding. The analogy is not perfect. AI diffuses far faster than the personal computer did, and it reaches into knowledge work the computer never touched. But the mechanism holds. The lag was never the technology. The lag is how long it takes an organization to redraw how it works.</p><h3>Two. Where the few actually start</h3><p>They start by finding one workflow worth redrawing.</p><p>Last week I gave you four boxes to look in: the customer, the employee, the partner, the builder. Use them to find a workflow, not a task to automate. Pick the one with the most pain and the least risk. That is your first.</p><p>Then ask the question that separates the few from the many: if we were building this workflow from scratch, for an agent, what would it look like?</p><p>The answer is almost never the workflow you have. Yours was built for people doing it by hand, with handoffs and approvals and waiting baked in, because that is what hands required. An agent does not need most of that. And here is the turn most people miss. AI does not have to be better than your people at every step to be worth it. It usually is not. The value does not come from the machine winning task by task. It comes from handing it the whole chain, so the waiting between the steps disappears.</p><p>Take one you will recognize. A customer emails that their order is late. Today that email sits in a queue. A rep eventually opens it, looks up the order in one system, switches to the carrier’s system to check the shipment, switches back to write the reply, and if anything is off, escalates it and waits. Five steps, three systems, two queues. Most of the elapsed time is not work at all. It is waiting between the steps.</p><p>Redraw it for an agent and it collapses to one. The agent reads the email the moment it lands, pulls the order and the shipping status itself, writes the answer, and routes only the genuine messes, the ones that need judgment, to a person. You did not make the five steps faster. You deleted four of them, and the waiting in between, which is where the time actually went.</p><p>If you want to know whether the redraw worked, watch one number. For that late-order workflow it might be the time from a customer’s email to a real answer, or the share of those emails resolved without a person touching them. Pick one and track it honestly.</p><p>That redraw is the work, and the data is blunt about it. McKinsey found the companies actually getting financial impact from AI were nearly three times as likely to have fundamentally redesigned their workflows. Of everything they measured, that redraw had the single biggest effect on whether AI reached the bottom line. Only about one in five companies has done it end to end. The other four are still layering agents onto the work they already had.</p><p>One honest warning before you start. The hardest part will not be the intelligence. In this year’s surveys, the single biggest obstacle to shipping an agent was not the model. It was getting it secure, reliable access to the systems where the work actually lives, and nearly half of the teams trying named exactly that as their number one problem. So do not build that part. Buy it, from someone, anyone, off the shelf if you can. The plumbing under all of this, the model access, the routing, the governance, the security, is what the big companies spent years building. It is not your edge. Build only the thin slice that is actually yours, your data and your redrawn workflow. Full disclosure, this is the business I am in, so weigh it accordingly. It is also just true.</p><h3>Three. The choice hidden inside the redraw</h3><p>Now we are back to Klarna, because this is where they went wrong, and it is the part nobody puts on the slide.</p><p>When you redraw a workflow, you decide who is still in it. And underneath that sits a quieter decision: which links in the chain stay human. Klarna handed the agent the whole chain, including the links that most needed a person, the upset customer and the messy exception, and those are exactly the links that broke. The fix was not more AI. It was putting the human back in the loop on the parts that needed one, and letting the machine carry the rest. Keep a human on the links where judgment, empathy, or accountability actually matter. It is the dispatcher who knows which driver will actually answer, the credit manager who can smell a bad account, the rep who can tell a furious customer from a confused one. Automate the links where none of that is in play. The redraw is just telling those two kinds of links apart, honestly.</p><p>Most companies are not deciding it that way, and the data shows it. When individuals reach for AI on their own, Anthropic’s usage research finds they mostly use it to augment, to make capable people faster. When businesses wire it into their systems, the same research finds it skews roughly three to one toward automation, toward removing the step and the person who did it. AI-linked layoffs crossed a hundred thousand this year. Some of that is real. Some of it, as more than one prominent investor has admitted out loud, is companies using AI as the silver-bullet excuse for cuts they already wanted to make.</p><p>So let me be honest about what a CFO will say back to me. Often the money does look like it is in the cut. The early returns have come from automating the back office, from deleting the step and the seat. I will not pretend otherwise. But the cut is a short-term number that quietly costs you the people who understood the work, and they are the ones who kept making it better. Klarna is the case study. Win the quarter, lose the decade.</p><p>I will give you the closest thing I have to proof that I mean it. In my day job I am COO of an AI company. We could generate our own content with the very tools we sell, and save the cost of a person. We hired a technical content writer instead. She uses AI every day, and it makes her faster, but the things that make the writing actually land, the creativity, the read on cultural nuance, the judgment about what will ring false, are the things the tools cannot do. We kept the human because the human is where the value was. It is a small example, one role at an AI startup, not a grand sacrifice. But it is the same decision you will face at every scale: where do you keep the person, because the person is the point?</p><h3>One thing to take with you</h3><p>Pick the workflow that hurts the most, and ask it the only question that matters: if we built this from scratch, for an agent, what would it look like, and who do we want still standing in it when we are done?</p><p>You do not need Walmart’s budget to ask that. You need an afternoon, a whiteboard, and the honesty to answer it. The question is free. What you do with the answer is the whole job.</p><p><strong>P.S. Your first ninety days, if you want a place to land it.</strong> Weeks one and two: run the four boxes, pick one workflow, redraw it on paper for an agent, mark which links stay human, and write the one metric that would prove it worked. Weeks three to six: stand it up on bought plumbing, one team, a human in the loop on the links that need one. Weeks seven to twelve: measure, govern, then go redraw the next one.</p><h3>References</h3><ol><li><a href="https://www.entrepreneur.com/business-news/klarna-ceo-reverses-course-by-hiring-more-humans-not-ai/491396">Klarna CEO Reverses Course by Hiring More Humans, Not AI</a>. Entrepreneur, 2025. After replacing roughly 700 support staff with an AI agent in 2023, Klarna’s CEO admitted the quality tradeoff was not sustainable and resumed hiring humans for complex and escalated cases.</li><li><a href="https://fortune.com/2025/08/18/mit-report-95-percent-generative-ai-pilots-at-companies-failing-cfo/">MIT report: 95% of generative AI pilots at companies are failing</a>. Fortune (reporting MIT NANDA, “The GenAI Divide”), Aug 2025. MIT’s study of 300 deployments found about 95 percent of corporate GenAI pilots produced no measurable P&amp;L return, and the barrier was organizational, not technical. The figure is widely cited and also contested.</li><li><a href="https://www.gartner.com/en/newsroom/press-releases/2025-06-25-gartner-predicts-over-40-percent-of-agentic-ai-projects-will-be-canceled-by-end-of-2027">Gartner Predicts Over 40% of Agentic AI Projects Will Be Canceled by End of 2027</a>. Gartner, 2025. Gartner projects most agentic AI projects will be scrapped over cost, unclear value, or weak controls, and warns of “agent washing”, legacy automation rebranded as agents.</li><li><a href="https://www.gartner.com/en/newsroom/press-releases/2026-05-20-gartner-warns-of-agent-washing-risks-in-supply-chain-planning-technology-market">Gartner Warns of Agent Washing Risks</a>. Gartner, May 2026. Gartner estimates only a small fraction of the thousands of self-described “agent” vendors offer genuine agentic capability.</li><li><a href="https://www.mckinsey.com/capabilities/quantumblack/our-insights/the-state-of-ai-how-organizations-are-rewiring-to-capture-value">The State of AI: How Organizations Are Rewiring to Capture Value</a>. McKinsey, 2025–2026. McKinsey finds workflow redesign is the factor most strongly associated with EBIT impact from gen AI; high performers are far more likely to have fundamentally redesigned workflows, and only about one in five companies has done it end to end.</li><li><a href="https://digitaleconomy.stanford.edu/app/uploads/2026/03/EnterpriseAIPlaybook_PereiraGraylinBrynjolfsson.pdf">Enterprise AI Playbook: Lessons from 51 Successful Deployments</a>. Stanford Digital Economy Lab (Brynjolfsson et al.), Mar 2026. Brynjolfsson’s team distills lessons from 51 working enterprise AI deployments, echoing the 1990s finding that value follows process redesign, not procurement.</li><li><a href="https://www.anthropic.com/research">The Anthropic Economic Index</a>. Anthropic, 2025–2026. Anthropic’s usage research finds individual use skews toward augmentation while business API use skews toward automation, evidence for how the augment-versus-replace choice is actually being made.</li></ol><p><em>Ben Pickett is Co-Founder and Chief Operating Officer of Swa and former Global Director of Site Reliability Engineering at Nike. Swa delivers enterprise AI inside Slack, Teams, WhatsApp, Zoom, SMS, and more channels coming, integrated into the systems behind them through user-scoped auth. One interface across many models. Zero data retention by default. Flat-rate pricing. Deployment in under three minutes. </em><a href="http://solo.swa-ai.com/"><em>Swa Solo</em></a><em>, our first version for individuals and small teams to try Swa before bringing it to the enterprise, ships soon. </em><a href="http://swa-ai.com/"><em>swa-ai.com</em></a></p><p><em>Header artwork generated with Swa.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=7f0bde4bf808" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[What Walmart Understood That the Other 95% Didn’t]]></title>
            <link>https://medium.com/@goshawk_55282/what-walmart-understood-that-the-other-95-didnt-16c75da2ac15?source=rss-d30c86172853------2</link>
            <guid isPermaLink="false">https://medium.com/p/16c75da2ac15</guid>
            <category><![CDATA[walmart]]></category>
            <category><![CDATA[ai-agent]]></category>
            <category><![CDATA[software-development]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[technology]]></category>
            <dc:creator><![CDATA[Ben Pickett]]></dc:creator>
            <pubDate>Thu, 04 Jun 2026 06:16:42 GMT</pubDate>
            <atom:updated>2026-06-04T06:16:42.703Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*oUkYwatdilVv01E6KcR2zg.jpeg" /><figcaption>Customer. Employee. Partner. Builder. The whole first step.</figcaption></figure><p><em>Walmart built the blueprint for enterprise AI. Here is the part you can actually copy, without their budget.</em></p><p>By Ben Pickett</p><p>I sell AI for a living. So it should bother you that the best blueprint for enterprise AI I have read this year did not come from an AI company. It came from a retailer. And almost nobody outside retail has read it.</p><p>Last week I wrote Andrew a letter about why his AI isn’t working. Andrew is the operator I keep meeting in different bodies. Smart, funded, two years into an AI program, and quietly furious that the productivity he was promised never arrived. His workforce says nothing changed. His board deck says everything changed. Somewhere between those two sentences is the truth, and it is costing him a fortune.</p><p>Andrew is not unusual. He is the rule. Last year MIT’s NANDA initiative studied 300 enterprise AI deployments for a report it called The GenAI Divide. Ninety-five percent of those deployments had produced no measurable impact on the P&amp;L. Not small impact. None. Only about 5% were pulling real revenue out of the work.</p><p>So after I sent that letter, I went looking for the 5%. Not someone with a thesis. Someone with a system in production and the numbers to prove it. I found them. The strange part is that they published the entire blueprint. It was Walmart.</p><p>And the best part is the piece that matters is copyable. Not the two years and the five thousand engineers. The shape. That is what the rest of this is about.</p><h3>What Walmart actually built</h3><p>Start with the shape, because the shape is the lesson.</p><p>Walmart did not build one giant assistant and hope it would do everything. They built four. Each one is purpose-built for a single audience, trained on that audience’s data, and accountable for that audience’s outcomes. Sparky serves shoppers. My Assistant serves the associates who run the stores. Marty serves the sellers, suppliers, and advertisers. WIBEY serves the developers building on top of all of it. Four super agents, four constituencies, no pretending that a customer and a warehouse associate need the same thing.</p><p>And here is what almost everyone gets backward. Walmart’s move was not to add agents. It was to tame them. Before the four super agents, Walmart had agent sprawl, a pile of bots multiplying faster than anyone could govern them. The super-agent structure was subtraction, not addition. They imposed order on a mess. Most enterprises are still in the addition phase, racing to stand up one more bot, when the lesson from the company that made it work is the opposite. The win was not more agents. It was control.</p><p>Underneath those four sit roughly 200 task-specific agents. Narrow, boring, reliable. And the piece most people skip past is WIBEY, because WIBEY is not really a chatbot for engineers. It is an orchestration layer. It interprets intent, finds the right agent, and coordinates the handoffs across the whole system. Walmart runs it as a federated model, where individual teams own their own agents while the orchestration layer makes sure those agents can find each other, talk to each other, and be trusted in production. All of it sits on Element, Walmart’s own machine-learning platform.</p><p>Suresh Kumar, Walmart’s global chief technology and development officer, put their stance in one sentence. Walmart is all in on agents. The architecture is the proof that he meant it structurally, not as a press line.</p><p>I want to be careful here, because the easy version of this story is wrong. Walmart did not invent agentic architecture. Plenty of people sketched this diagram on a whiteboard before Walmart did. What Walmart did is harder and rarer. They shipped it at scale, to millions of users, and they have the business results to show it works. The whiteboard is free. The proof is what costs two years.</p><h3>The proof</h3><p>This is where most AI stories go quiet, so let me stay loud.</p><p>Start with how serious they were. Walmart’s 2025 proxy shows it paid Daniel Danker, the executive vice president running its AI push, 44.1 million dollars. That is more than the 29.2 million it paid CEO Doug McMillon. A retailer decided the person building its agent architecture was worth more than the person running the whole company. Sit with that.</p><p>On its recent earnings calls, Walmart told investors that shoppers who use Sparky spend about 35% more per order than the average shopper. Units purchased through Sparky more than quadrupled in a single quarter. By the end of the fiscal year, half of all Walmart app users had already tried it. And this is happening inside a company that just did 713 billion dollars in revenue, so it is not a rounding error in a sandbox. It is moving the biggest number there is.</p><p>These are not engagement metrics. Nobody is celebrating time on page. These are dollars per order and units sold, the two numbers a CFO actually trusts. That is the difference between an AI program and an AI announcement. One of them changes the P&amp;L.</p><p>So we have a working reference architecture, in production, with the receipts. The natural next thought for every operator reading this is the dangerous one. Great, let’s go do what Walmart did. Do not do what Walmart did.</p><h3>Why most of us cannot copy it directly</h3><p>Here is the part the architecture diagrams leave out. Walmart spent something like two years and on the order of five thousand engineers building this. They wrote their own orchestration. They built their own ML platform. They trained domain-specific models on proprietary data most companies will never have. They had the balance sheet to fund a multi-year bet that produced nothing visible for a long time, and the institutional patience not to kill it in month nine.</p><p>You probably have none of that. Andrew certainly doesn’t. He has a real budget, a small team, a board that wants results this year, and a competitor moving just as fast.</p><p>And the data says the build-it-yourself instinct is exactly the trap. That same MIT report found that when companies buy AI capability from specialized vendors and partner for it, the projects succeed about 67% of the time. When they build it internally, they succeed roughly a third as often. McKinsey’s State of AI 2025 saw the other half of the problem: 62% of organizations are now experimenting with AI agents, but only 23% have managed to scale them across the enterprise. Everyone can run a pilot. Almost nobody can run the system.</p><p>If the lesson from Walmart were to spend two years and five thousand engineers, the lesson would be useless to the 99% of companies that need it most. But that is the wrong lesson. The resourcing is not the lesson. The resourcing is just what it cost Walmart to discover the lesson on our behalf. The lesson is the shape.</p><h3>The portable architecture</h3><p>Strip Walmart out of the Walmart story and look at what remains. Take away the proprietary models, the custom platform, the five thousand engineers. What is left is a pattern, and the pattern is portable.</p><p>Group your agents by the people they serve, not by the technology. Customer, employee, partner, developer. Those four constituencies exist in almost every enterprise, whether you sell sneakers or audits or insurance. Picture a mid-size accounting firm. Its customers are clients. Its employees are the staff doing the work. Its partners are the referral sources and the software it files through. Its builders are the small internal team wiring the tools together. That is the same four-quadrant shape Walmart drew, minus the 713 billion dollars. Give each quadrant a purpose-built agent that knows its world.</p><p>Under each of those, specialize. Build or buy narrow task agents that do one job reliably, instead of one general assistant that is mediocre at everything and trusted with nothing.</p><p>Above all of it, run an orchestration plane. This is the part operators underestimate, and it is the part the market is finally catching up to. The analysts tracking this space now say orchestration and governance are eclipsing raw model capability as the main reason enterprises pick one platform over another. Read that again. The model is no longer the differentiator. The layer that routes, coordinates, governs, and keeps the system accountable is the differentiator. Walmart’s real moat is not Sparky. It is the orchestration above Sparky.</p><p>And then give the workforce access. Not a pilot for twelve people. The architecture only compounds when everyone is inside it.</p><p>Here is the honest part. For Walmart, that orchestration plane and the multi-model access underneath it were a two-year build. For the rest of us, the data is clear that they should be bought, not built. That is the part of this story I am closest to, because the orchestration and multi-model layer is exactly what we build at Swa. I am not going to turn this into a pitch. I will just tell you the most useful thing I learned from studying Walmart, which is that the layer everyone treats as plumbing is actually the product. The model is not the moat. The structure is.</p><p>I want to be honest about the part the diagrams do not show. That orchestration layer is the answer, and it is also where all the new risk concentrates. Every agent you add is another door someone can walk through. Govern them centrally or you have just built a faster way to leak. And the human question is the one even Walmart will not fake. Their own executive said out loud, we expect jobs to evolve and we do not know what that looks like yet. I believed them more for saying it, not less. Anyone selling you certainty on that is selling you something.</p><h3>The close</h3><p>For two years the enterprise conversation has been a fight about which model to pick. Open or closed. This vendor or that one. Benchmarks traded like baseball cards. Walmart skipped the entire argument. They did not win by picking the best model. They won by designing the best structure and letting the models be interchangeable parts inside it.</p><p>The companies that win the next decade are not the ones with the cleverest model. They are the ones with the clearest architecture. Walmart just showed all of us what one looks like in production, with the earnings to back it. You do not have their two years or their five thousand engineers. You do not need them. You need their shape.</p><p>So here is the one thing worth doing this week, and it does not require a budget or a single engineer. Get a whiteboard. Draw four boxes. Customer, employee, partner, builder. In each box, write the single agent that quadrant needs most. That is it. You will not have built anything yet, but you will have done more real architecture than 95% of the companies in that MIT study.</p><p>Stop arguing about models. Start designing the structure. Start there, then ship yours.</p><p><em>I’m Ben Pickett. I write Enterprise Field Notes. If your team is wrestling with the orchestration question, my inbox is open. What does your super-agent structure look like?</em></p><h3>References</h3><ol><li><a href="https://fortune.com/2025/08/18/mit-report-95-percent-generative-ai-pilots-at-companies-failing-cfo/">The GenAI Divide: State of AI in Business 2025</a>. MIT NANDA initiative, reported by Fortune, August 2025. Across 300 enterprise AI deployments, 95% delivered no measurable P&amp;L impact; tools bought from specialized vendors succeeded about 67% of the time, versus roughly a third as often for internal builds.</li><li><a href="https://www.mckinsey.com/capabilities/quantumblack/our-insights/the-state-of-ai">The State of AI in 2025</a>. McKinsey, November 2025. Only about 6% of organizations qualify as high performers capturing significant value, and while 62% are experimenting with AI agents, just 23% have scaled them. McKinsey’s core finding is that the advantage is organizational, not technological.</li><li><a href="https://tech.walmart.com/content/walmart-global-tech/en_us/blog/post/wibey-announcement.html">From models to agents: a new era of intelligent systems at Walmart</a>. Walmart Global Tech, August 2025. Walmart’s own account of WIBEY as an invocation and orchestration layer across its agentic ecosystem, built on the proprietary Element ML platform, alongside the launch of four super agents.</li><li><a href="https://www.digitalcommerce360.com/2025/07/25/walmart-super-agents-ai-danker/">Walmart rolls out four AI ‘super agents’</a>. Digital Commerce 360, July 2025. Reports CTO Suresh Kumar consolidating Walmart’s AI into four super agents for customers, associates, sellers and suppliers, and developers.</li><li><a href="https://www.digitalcommerce360.com/2026/05/22/walmart-sparky-agent-ai-sales-supply-chain/">Walmart credits Sparky AI agent with lifting AOV, unit sales growth</a>. Digital Commerce 360, May 2026. On Walmart’s May 2026 earnings call, Walmart U.S. CEO David Guggina said Sparky shoppers have about 35% higher average order value than non-users, and units purchased through Sparky more than quadrupled from the prior quarter.</li><li><a href="https://www.constellationr.com/insights/news/walmarts-sparky-ai-agent-increases-order-value">Walmart’s Sparky AI agent increases order value</a>. Constellation Research, 2026. Independent analysis of the average-order-value lift and of Sparky adoption, with roughly half of Walmart app users having tried it.</li><li><a href="https://wwd.com/business-news/financial/walmart-ai-chief-2025-pay-ceo-1238915275/">Walmart’s AI Chief Earned More Than Its CEO in 2025 With a $44.1M Payday</a>. WWD, 2026. Reports that Walmart paid Daniel Danker, its EVP of AI, 44.1 million dollars in 2025, more than CEO Doug McMillon’s 29.2 million.</li></ol><p><em>Ben Pickett is Co-Founder and Chief Operating Officer of Swa and former Global Director of Site Reliability Engineering at Nike. Swa delivers enterprise AI inside Slack, Teams, WhatsApp, Zoom, SMS, and more channels coming, integrated into the systems behind them through user-scoped auth. One interface across many models. Zero data retention by default. Flat-rate pricing. Deployment in under three minutes. </em><a href="http://solo.swa-ai.com/"><em>Swa Solo</em></a><em>, our first version for individuals and small teams to try Swa before bringing it to the enterprise, ships soon. </em><a href="http://swa-ai.com/"><em>swa-ai.com</em></a></p><p><em>Header artwork generated with Swa.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=16c75da2ac15" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Dear Andrew, COO of Uber]]></title>
            <link>https://medium.com/@goshawk_55282/dear-andrew-coo-of-uber-3f3c968cfe08?source=rss-d30c86172853------2</link>
            <guid isPermaLink="false">https://medium.com/p/3f3c968cfe08</guid>
            <category><![CDATA[leadership]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[uber]]></category>
            <category><![CDATA[technology]]></category>
            <category><![CDATA[sustainability]]></category>
            <dc:creator><![CDATA[Ben Pickett]]></dc:creator>
            <pubDate>Thu, 28 May 2026 15:11:13 GMT</pubDate>
            <atom:updated>2026-05-28T15:11:13.259Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*qfczzhdivFERY0SnBudLGA.jpeg" /><figcaption>AI is here to stay. The only question is what it costs us to keep it.</figcaption></figure><p><em>Uber burned through its AI budget in four months. This is an open letter to the COO who said so out loud, on what we actually owe.</em></p><p>By Ben Pickett</p><p>Andrew,</p><p>I had this week’s issue half-written. It was about redesigning workflows. Then I read your interview in Fortune, and I set the other one down.</p><p>You said the thing most people in your seat will only say in private. You cannot draw a straight line between what Uber spends on AI and what it actually produces. Here is how you put it.</p><blockquote>That link is not there yet. Maybe implicitly there’s more that is getting shipped, but it’s very hard to draw a line between one of those stats and “Okay now we’re actually producing like 25% more useful consumer features.”</blockquote><p>In plain terms: we are spending a fortune on AI, and almost nobody can prove it is working. That is the most honest sentence an operator has said out loud this year, and you said it on the record.</p><p>This letter is what I owe you back, as one COO to another. I changed my plan for the week because this is the question sitting in every operating review in the country right now, and it deserves an answer.</p><p>You burned through Uber’s 2026 AI budget in four months. Your engineering organization went from 32% Claude Code adoption to 84% in the same window. Per-engineer API costs ran $500 to $2,000 a month. At the top of that range across 5,000 engineers, that annualizes to nine figures on a single tool, before counting the rest of your AI stack.</p><p>And when you look at the output, you cannot point to a 25% lift in consumer features that justifies the curve.</p><p>You are not alone in this. Enterprise AI spending jumped 108% in a year. The average company’s AI bill went from a little over a million dollars to seven million. Seventy-eight percent of IT leaders say they are seeing charges they never budgeted for. This is not an Uber problem. It is everywhere. You are just the one who said it with your name attached.</p><p>You are not running it wrong. The industry’s defaults are running you wrong. The answer is not less AI. The answer is stewardship. The reason you cannot prove your AI is working is that the way enterprise AI is deployed in 2026 fails three different kinds of stewardship at once.</p><p><strong>Money. Natural resources. People.</strong></p><p>Each one alone would be a problem. Together they compound, and the compound is what your budget is paying for.</p><h3>Money is the one your finance team feels first.</h3><p>The single biggest decision driving your token spend is happening implicitly, in the background, every minute of every day. Your engineers are defaulting to the most capable model available for every task they touch. Some of those tasks need a frontier model. Most of them do not.</p><p>A query that summarizes a Slack thread does not need Claude Opus. A query that drafts a short Jira ticket does not need GPT-5. A query that pulls a snippet from internal documentation does not need the most expensive model in the room. But when the default presents a frontier model as the option, the frontier model is what gets used.</p><p>Per-seat pricing makes this worse, not better. The seat does not predict the spend. Token volume does. And token volume runs on model selection. Two engineers on identical seat licenses can produce a tenfold difference in monthly cost depending on which model they reach for and how often the model gets retried because the first answer was wrong.</p><p>Here is the math at scale that bothers me most. If the per-engineer spend in your reported range averages around $1,000 a month, and your engineering organization is roughly 5,000 strong, the annualized run rate for a single coding tool sits at about $60 million. That is one tool, on one cohort. You have other cohorts. You have other tools. Across an enterprise AI stack, this is a budget category that did not exist three years ago and now competes for share with payroll. The CFO conversation in 2027 will not be about whether AI is in the budget. It will be about whether AI is the budget.</p><p>This is not theoretical. At Swa, the company I co-founded, we run our own platform on our own work. Over the last 90 days, our internal compute cost dropped 51%. We did not change the work. We changed one thing. Instead of sending every task to the most expensive model, an orchestration layer routed each task to the right model for the job. Same work, same people, half the compute cost. That is what an orchestrator does. It is the single highest-leverage decision in enterprise AI right now, and most companies have not made it.</p><p>If your engineers had been defaulting correctly across model tiers, your Q1 burn would not have ended in four months.</p><h3>Money is the visible cost. There is a second cost we are not pricing in.</h3><p>Recent benchmark research across 155 model architectures found that frontier models with 70 billion or more parameters consume roughly two orders of magnitude more energy per token than smaller open-source models. That is not a 20% gap. That is a 100x gap on the same task.</p><p>Each 100-word AI prompt uses roughly one bottle of water (about 519 milliliters) on the data-center cooling side. A typical AI data center draws electricity equivalent to 100,000 households. By 2028, US data centers are projected to consume 12% of national electricity, up from 1.9% in 2018. The IEA forecasts global data-center electricity demand to roughly double by 2030. AI workloads account for most of that growth.</p><p>Consider the timeline. The IEA projects global data center electricity demand to double by 2030, with AI accelerator workloads driving most of the growth. That demand has to be met by something. Part of it will be met by new renewable capacity. Part of it will be met by extending the life of fossil-fueled generation that was scheduled to retire. The default-to-premium choice in your engineering team’s prompt window is connected, by a chain of consequence, to that retirement decision being deferred. We do not see the chain because each link sits in a different organization. The chain is still there.</p><p>This is not a separate issue from the money story. It is the same decision viewed from a different axis. Defaulting a routine query to a frontier model is a stewardship failure on two pillars at once. Joules and dollars. The same routing decision that cuts the budget cuts the energy footprint by something close to it, because routing means a smaller model carries the load the frontier model did not need to.</p><p>Pricing has not caught up to the externality yet. It will. The companies that figure this out first will be the ones that built stewardship into the architecture before they had to.</p><h3>The third pillar is the one the industry talks about the least. People.</h3><p>When 84% of your engineering organization adopts a coding agent, that is a headline. Inside the organization it is also a story about cognitive load. Somebody has to validate every output the agent produces. Somebody has to catch the hallucination that slips past. Somebody has to repair the trust deficit when the answer was wrong. Those somebodies are your engineers, and the validation work shows up nowhere in the productivity number you are being asked to defend.</p><p>Stack Overflow’s 2025 developer survey found that 66% of developers cite “answers that are almost right, but not quite” as their biggest frustration with AI tools. 45% report that debugging AI-generated code takes longer than debugging their own. Trust in AI output declined year over year while adoption rose. The two numbers are not in tension. They are the same story.</p><p>WalkMe’s 2026 study found that workers lose 51 working days per year to technology friction, up 42% from 2025. 54% of workers bypassed AI tools in the last 30 days to do the task manually. The tools are deployed. People work around them anyway. That is not a training problem. That is a stewardship debt being paid in the attention and patience of the people you are asking to make the tools work.</p><p>There is also a structural piece nobody has named yet. The validation work that AI offloads onto your engineers is asymmetric. The model produces in seconds what would have taken a junior engineer hours. The senior engineer who reviews and corrects the model output now spends those hours doing review work that used to be distributed across multiple junior engineers learning the craft. The short-term productivity curve looks good. The pipeline of engineers who learn by writing the code in the first place is being quietly hollowed out. Three years from now, when the model is better but the senior engineers have started to retire, the organization will have a depth problem nobody saw coming.</p><p>Adoption is easy to measure. Trust is not. Depth on the bench is not. Your 84% is the number that comes easy. The ones that matter do not show up on a dashboard.</p><h3>Three pillars. One architectural cause underneath all three.</h3><p>I learned to recognize that cause a long way from enterprise AI. I served six years on the City Council of Jefferson, Oregon, three of them as Mayor. Most of that work was infrastructure. We built a new wastewater treatment plant. We built a new city hall. We paved roads that had not been paved in decades. We dealt with water rights, drainage, and the boring backbone that lets a town function. Good management is what makes infrastructure invisible. Water that just runs when you turn the tap. Lights that just come on. Roads that just hold. People only notice it when it fails.</p><p>AI should be infrastructure. Flushing your toilet should just work. AI should just be a utility. But it is not one yet. Right now it is a tap with no plumbing behind it. The water comes out, but nothing routes it, meters it, or controls the pressure. That is why you cannot prove your AI is working.</p><p>The industry deployed AI without an orchestration layer. Without routing. Without governance. Without cost attribution at the user level. Without observability at the model level. The companies running enterprise AI in 2026 are running it the way the early internet was deployed: directly, defaulting to the most powerful resource available, hoping the bill works out.</p><p>That is not a vendor problem. It is a discipline problem. And the discipline has a name. Stewardship.</p><p>Money breaks the budget. Joules break the planet. People-neglect breaks the org. Each pillar fails by itself. Together they fail faster, because the same architectural gap is causing all three.</p><p><em>Andrew, since you have read this far, I want to tell you why I am building what I am building, and then I will stop.</em></p><p><em>I left Nike to co-found Swa because I watched too many enterprise AI deployments fail at exactly this architectural layer. The model was never the problem. The architecture around the model was the problem. Swa is that architecture. It lives where people already work, runs every major model, governs them, attributes every dollar, and connects to the systems the work already depends on. Routing is one piece. The 51% is one outcome of the whole thing working together, not a single trick. How we get there is more than I will put in a public letter.</em></p><p><em>I am not writing this to sell you anything. I am writing because these are the problems I get out of bed to solve. AI is here to stay. The only question left is what it costs us to keep it, in money, in energy, and in people. The next eighteen months decide whether we treat AI as infrastructure with real stewardship, or whether we keep chasing the most aggressive adoption number while all three quietly bleed. I care more about getting that answer right than about who gets credit for it.</em></p><p>If you are not Andrew but you are running a similar operation, the same diagnosis applies. The link you are looking for between AI spend and AI output does not arrive naturally. It arrives only when stewardship becomes an architectural decision rather than a moral one. The discipline is the same whether the company is 5,000 engineers or fifty. The order of operations is the same. Money. Energy. People. Routing, governance, observability, attribution. Boring infrastructure work that nobody pitched you when AI was the headline. The work that turns AI from a line item into a return.</p><p>Andrew, you said you cannot draw the line between what you spend on AI and what it produces. I believe you. With the architecture most companies are running, that line does not exist to be drawn. It is not a measurement failure. It is an architecture failure. The line appears when the layer underneath your AI is built to make it appear, and most companies have not built it yet.</p><p>If you ever want to compare notes, one COO to another, I am one of the people who has built this from the other side. The offer is real, with no pitch attached.</p><p>Either way, thank you for saying what you said on the record. The honesty is what makes this conversation possible. The rest of us owe you the same honesty in return.</p><p><strong>Ben Pickett</strong></p><p>Co-Founder and COO, Swa</p><p>Former Global Director of Site Reliability Engineering, Nike</p><p><em>Ben Pickett is Co-Founder and Chief Operating Officer of Swa and former Global Director of Site Reliability Engineering at Nike. Swa delivers enterprise AI inside Slack, Teams, WhatsApp, Zoom, SMS, and more channels coming, integrated into the systems behind them through user-scoped auth. One interface across many models. Zero data retention by default. Flat-rate pricing. Deployment in under three minutes. </em><a href="https://solo.swa-ai.com/"><em>Swa Solo</em></a><em>, our first version for individuals and small teams to try Swa before bringing it to the enterprise, ships soon. swa-ai.com</em></p><p><em>Header artwork generated with Swa.</em></p><h3>References</h3><ol><li><a href="https://fortune.com/2026/05/26/uber-coo-ai-spending-tokens-claude-code/">Uber burned through its entire 2026 AI budget in four months. Now its COO is questioning whether it’s worth it</a>. Fortune, May 26 2026. Andrew Macdonald, President and COO, on record about the productivity-to-spend link. Source of the Q1 2026 R&amp;D figure ($951M, up 17%) and the four-month budget burn.</li><li><a href="https://www.beri.net/article/ai-costs-surge-108-percent-2026-budget">AI Costs Surge 108%: Why Your 2026 Budget Is Already Outdated</a>. Beri.net, 2026. Enterprise AI spending up 108% year over year; average annual spend grew from $1.2M (2024) to $7M (2026); 78% of IT leaders report unexpected AI charges they did not budget for.</li><li><a href="https://oplexa.com/ai-inference-cost-crisis-2026/">AI Inference Cost Crisis 2026: Why Your AI Bill Is Exploding</a>. Oplexa Research, 2026. Documents the Claude Code adoption curve at Uber (32% to 84% across 5,000 engineers) and the $500 to $2,000 per-engineer monthly API range. 80 to 85% of enterprises miss their AI infrastructure forecasts by more than 25%.</li><li><a href="https://www.webpronews.com/microsoft-doubles-copilot-pricing-to-60-per-seat-what-it-means-for-enterprise-ai-spending/">Microsoft Doubles Copilot Pricing to $60 Per Seat</a>. WebProNews, 2026. Microsoft 365 Copilot Enterprise raised from $30 to $60 per seat per month. GitHub Copilot transitioned all plans to usage-based billing effective June 1, 2026.</li><li><a href="https://arxiv.org/pdf/2502.05610">Towards Sustainable NLP: Insights from Benchmarking Inference Energy in Large Language Models</a>. arXiv 2502.05610, 2026. Comprehensive measurements across 155 model architectures: frontier models with 70B+ parameters consume approximately two orders of magnitude more energy per token than smaller open-source models.</li><li><a href="https://www.iea.org/reports/energy-and-ai">Energy and AI</a>. International Energy Agency, 2025. Global data center electricity demand projected to rise from 415 TWh (2024) to approximately 945 TWh by 2030. US data centers at 4.4% of national electricity in 2025, up from 1.9% in 2018; projected to reach as high as 12% by 2028.</li><li><a href="https://www.eli.org/vibrant-environment-blog/ais-cooling-problem-how-data-centers-are-transforming-water-use">AI’s Cooling Problem: How Data Centers Are Transforming Water Use</a>. Environmental Law Institute, 2026. Each ~100-word AI prompt uses approximately 519 milliliters of water on the data-center cooling side. US AI servers are projected to drive 200 to 300 billion gallons of additional annual water consumption and add 24 to 44 million metric tons of CO2-equivalent emissions by 2030.</li><li><a href="https://survey.stackoverflow.co/2025/ai/">2025 Stack Overflow Developer Survey: AI</a>. Stack Overflow, December 2025. 84% of developers use or plan to use AI tools; 51% of professional developers use AI daily; 66% cite “AI solutions that are almost right, but not quite” as their biggest frustration; 45% report debugging AI-generated code takes longer than debugging their own. Trust in AI output declined year over year.</li><li><a href="https://www.walkme.com/news-releases/enterprises-lose-51-workdays-per-employee-to-technology-friction-annually-despite-record-ai-investment-walkme-global-study-of-3750-finds/">Enterprises Lose 51 Workdays Per Employee to Technology Friction Annually</a>. WalkMe Global Study, 2026. Survey of 3,750 respondents at enterprises with 1,000+ employees across 14 countries. 51 workdays lost per year per employee, up 42% from 2025. 54% of workers bypassed AI tools and completed tasks manually in the past 30 days.</li><li><a href="https://medium.com/@goshawk_55282/the-definitions-gap-863b2e78c33c">The Definitions Gap</a>. The Enterprise Field Notes Issue #5, May 19 2026. Companion piece on workflow alignment as the upstream input every other AI initiative depends on.</li><li><a href="https://medium.com/@goshawk_55282/the-productivity-disconnect-dc0fa5ab7b70">The Productivity Disconnect</a>. The Enterprise Field Notes Issue #4, May 14 2026. Companion piece on the gap between AI spend and measurable enterprise output.</li></ol><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=3f3c968cfe08" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>