<?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 Andrew Coward on Medium]]></title>
        <description><![CDATA[Stories by Andrew Coward on Medium]]></description>
        <link>https://medium.com/@andrew-coward?source=rss-308b3516a285------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*tHy2Y2yXwDBaPOGLaPSSJg.jpeg</url>
            <title>Stories by Andrew Coward on Medium</title>
            <link>https://medium.com/@andrew-coward?source=rss-308b3516a285------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Mon, 16 Nov 2026 05:14:00 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@andrew-coward/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[Everyone Is Building AI Data Centers. Who Is Rebuilding the Web?]]></title>
            <link>https://andrew-coward.medium.com/everyone-is-building-ai-data-centers-who-is-rebuilding-the-web-86cecddd7f88?source=rss-308b3516a285------2</link>
            <guid isPermaLink="false">https://medium.com/p/86cecddd7f88</guid>
            <category><![CDATA[network-infrastructure]]></category>
            <category><![CDATA[web-development]]></category>
            <category><![CDATA[ai-infrastructure]]></category>
            <dc:creator><![CDATA[Andrew Coward]]></dc:creator>
            <pubDate>Sun, 02 Aug 2026 18:04:04 GMT</pubDate>
            <atom:updated>2026-08-02T19:04:32.494Z</atom:updated>
            <content:encoded><![CDATA[<h4>How AI agents are reshaping web traffic, content, privacy and the infrastructure behind every digital interaction</h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*YGfAy1-AK1NUSuhIXSoTCA.png" /></figure><p>Back in <a href="https://medium.com/@andrew-coward/how-ai-is-fast-rewiring-the-web-economy-a33fcfd153b1">October last year</a>, I wrote about how bot traffic isn’t just changing the landscape of Internet traffic but is radically rewiring what it means to deliver content in a bot-first world.</p><p>As has been widely reported, if you check out Cloudflare’s <a href="https://radar.cloudflare.com/traffic#bot-vs-human">excellent dashboards</a>, you will see that we have already crossed the tipping point at which automated systems and bots generate more web requests than humans.</p><p>However, we are not really having the discourse about what this means for websites, web infrastructure or how enterprises and governments should respond to this change. The market is so absorbed in the AI data center buildout that the other 50% of the equation: what the bots will actually talk to, is being largely ignored. My focus is on what this traffic shift means in practical terms.</p><p>In fact, there are three critical vectors to which we should all be paying attention: <strong>traffic volume</strong>, <strong>bot honey</strong> (how to build the content bots crave) and <strong>personal data</strong>. Each is having a profound effect on website hosting, infrastructure providers and enterprises whose websites are being used by bots in ways their designers never imagined.</p><p><strong>Traffic Volume</strong></p><p>This isn’t just a function of bots taking over web requests from humans. Bots may repeatedly interrogate the same systems with variations of essentially the same request.</p><p>Think about a best friend who obsesses over securing a business-class seat upgrade or a five-star hotel suite using points and miles. If you don’t have a friend like this, it’s probably you.</p><p>This person may spend hours or days building an itinerary, figuring out which combination of points to use for each segment in order to construct the perfect trip. As bots get better at managing our personal data (more on this later) give this task to a bot and it can be completed in minutes, while testing radically more combinations to maximize the value of your points.</p><p>As <a href="https://www.theregister.com/special-features/2025/12/13/ba-boss-fears-ai-agents-could-make-brands-invisible/1911906">Sean Doyle, CEO of British Airways, has argued</a>, the days when a decent website and a prominent Google result were enough to remain visible are “diminishing quickly,” with bots and AI agents now poised to sit between customers and the brands competing for their money.</p><p>This should not be taken to mean that website traffic will fall as content is disintermediated. On the contrary, traffic volumes are not just rising , rather the increase may be so large that it is difficult to comprehend.</p><p>According to Cloudflare CEO Matthew Prince, speaking to <a href="https://www.axios.com/2025/06/19/ai-search-traffic-publishers">Axios last year</a>, Google crawled two pages of content for every visitor it sent to a publisher ten years ago. Cloudflare now sees Anthropic’s ratio at 60,000 pages crawled for every visitor referred. For OpenAI, the ratio is 1,500:1.</p><p><strong>Bot Honey</strong></p><p>It turns out that bots aren’t interested in your website’s beautiful graphics, expensive commercials or award-winning design. They are looking for what I think of as <strong>bot honey</strong>: the facts, inventory, specifications, evidence and detailed information they need to complete a task.</p><p>For the past two decades, companies have designed websites to attract, persuade and convert human visitors. Bots operate differently. They cannot be won over by emotion or visual polish. They interrogate a site to determine what is available, whether it meets the user’s requirements and whether the claims being made can be supported.</p><p>Recently, while advising a beauty-care startup, we spoke with a well-known cosmetics company about what our bot (and others like it) needed before it could include the company’s products in its application.</p><p>The “honey” they needed to make was a detailed list of the ingredients in each product, the concentrations in which they appeared, the evidence supporting their effectiveness and the quality-control processes through which the products had passed. The startup already knew which ingredients worked best for wrinkles, sunburn or itching. What the bot needed was precisely the information the beauty company was most reluctant to disclose: what exactly is in that $100 bottle, in what concentrations, and supported by what evidence?</p><p>Just as companies spent the past two decades obsessing over SEO, they now need to understand what constitutes the information bots need to be successful and thereby convert bot traffic into commercial outcomes. This will create new tensions around what information they publish, which agents they make it available to and how they serve it without exposing valuable data to every crawler on the internet.</p><p><strong>Personal Data</strong></p><p>This brings us to the personal-data challenges, which, to my mind, remain largely unresolved.</p><p>Consider the following scenario. You want to find a new gym and develop a new exercise regime, but you also have several medical conditions. You want to combine your medical information with the best scientific research on exercise and identify a gym capable of providing the best workout for your unique circumstances.</p><p>As your bot connects with every gym in your town, does it share all this information?</p><p>If it doesn’t, you may not get the best outcome. But if it does, what happens to your personal data? Does each gym get to learn about your medical conditions before you have even bought a membership? If not, how is the information withheld while still allowing the gym to determine whether it can meet your needs?</p><p>There clearly needs to be a model governing how personal information is shared, for what purpose and for how long. We will also need technology that ensures this information does not leak or get reused for another purpose.</p><p>Does a gym have an obligation to ensure that it does not accidentally receive unnecessary medical information from an errant bot? Can the agent translate a medical condition into a set of exercise requirements without revealing the underlying diagnosis? Who is responsible if it shares more than the user intended?</p><p><strong>What Happens Next?</strong></p><p>In 2001, Google crossed the threshold of 100 million searches a day, and search ranking became commercially existential. In 2026 we have now crossed a similar line with bots, with about the same level of preparation and understanding that we had during the transition in 2001 — and considerably less time in which to adjust to the impact.</p><p>For infrastructure providers and manufacturers, we need new types and classes of security and gateway products that can scale, monitor, secure and manage these bot interactions. In the same way web traffic spawned the creation of content delivery networks, firewall companies, website management systems, identity management and payment systems, the agentic web will need corollary systems that map to the agentic paradigm.</p><p>Separately, a whole new class of experts will emerge to drive GEO (Generative Engine Optimization) AEO (Answer Engine Optimization) and, more broadly, AAIO (Agentic AI Optimization). CISOs will need to consider how personal and confidential data flows between bots both externally and internally.</p><p>Every company wants the AI agent that brings them a customer. Conversely, nobody wants the crawler that consumes its data, exhausts its systems or substitutes for the customer without paying. The enterprises that adapt first by distinguishing valuable agents from extractive or malicious bots, and serving each accordingly, will have a greater chance of acquiring customers than websites that remain focused exclusively on a consumer-first experience.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=86cecddd7f88" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Where’s the AI Kill Switch?]]></title>
            <link>https://andrew-coward.medium.com/wheres-the-ai-kill-switch-f138036082db?source=rss-308b3516a285------2</link>
            <guid isPermaLink="false">https://medium.com/p/f138036082db</guid>
            <dc:creator><![CDATA[Andrew Coward]]></dc:creator>
            <pubDate>Fri, 17 Jul 2026 13:55:57 GMT</pubDate>
            <atom:updated>2026-07-17T13:55:57.272Z</atom:updated>
            <content:encoded><![CDATA[<h4>Lessons from the Sorcerer’s Apprentice</h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*SEZiNyHxnlA2xHun8i_4mw.png" /></figure><p>Recently, some of my clients have been asking about the need for AI kill switches. It was not something I had considered directly until I had a couple of AI “challenges” of my own, which reminded me of the scene in <em>Fantasia</em> where Mickey Mouse, playing the Sorcerer’s Apprentice, uses magic to make a broom carry water for him.</p><p>Mickey’s mistake was not that he automated the work, but that he confused starting the process with controlling it. He knew how to set the broom in motion, but not how to stop it, and when he tried to stop it by destroying the automated brooms, each fragment became another broom and the flooding accelerated. The problem was resolved only when The Sorcerer returned with the authority and knowledge to bring the system under control.</p><p>That is also the central lesson for AI products. A kill switch matters, but it should be the final control rather than the management system itself. The real work is defining what reasonable behavior looks like, detecting when the system is moving outside that operating envelope, restricting its authority before the problem spreads, and ensuring that an external control remains available when everything else has failed.</p><p>My own incidents involved flooding and expense of a different kind, as agents consumed tokens at a radical pace while creating a tidal wave of automated emails. The systems were not necessarily malfunctioning in the conventional sense; they were doing more of what I had asked, at greater scale and cost than I had anticipated, which made the gap between automation and control uncomfortably obvious.</p><p>Now you might think that this is just about code hygiene and testing, but because of the unpredictable nature of AI responses, I’ve had to rewire my CI process to ensure AI supervision at testing time, and in production. Are there checks to make sure that the ability to consume tokens is rate-limited or stops after a certain time? What’s a reasonable number of emails to send and for what purpose? However, I came to realize that I couldn’t put generic rules in place in the CI process because my catch-all logic, clearly wasn’t catching everything.</p><p>And of course, once my code was deployed, no amount of CI process was going to control what it might actually do in production with a non-deterministic LLM engaged. Building complex systems generates complex interactions and my agents were working differently when deployed as a scaled-out system than when tested as a standalone entity.</p><p>So I clearly needed some kind of supervisory monitoring to take place to understand what reasonable looks like and stop flooding early — a kill switch essentially.</p><p>I also recognized that if I implemented a kill switch, it should trigger not just because the system wasn’t operating as designed but because it was about to create a flood for my customers or a financial challenge for my company. In fact, the aim would be that the kill switch, while it exists, should never need to be triggered, because the system had detected issues way before a nightmare scenario was in flight.</p><p>But how to implement? I’d noticed that as my agents became more sophisticated, their token use was creeping-up beyond my original limits, creating a denial-of-service type event for my customers. This token use was legitimate, but I’d failed to update the baseline.</p><p>It was time for a framework. Rather than just build a kill-switch that would break my service to stop the system, I needed checks at each stage, from planning through to a monitored deployment that would take into account the original design criteria in how or when a kill-switch would trigger as a signal that the system was far outside of what reasonable looks like. When new features were introduced to the agent, the implementation would need to reflect these design intents, which would trigger changes through the development, testing and deployment monitoring systems.</p><p><strong>Phase 0: Define the Operating Envelope — what does normal look like?</strong></p><p>When planning the agent or feature, I now require the PRD to model what normal looks like from a token consumption perspective and transactions perspective, so that natural gates are placed in the system. This often requires prototyping to model-out likely consumption and transaction rates, but extends further into governance, telemetry and authorizations. I summarized this as a PRD requirement that humans or agents have to include in their PRD:</p><ol><li><strong>Purpose &amp; Authority</strong></li></ol><p>What is the agent allowed to do, and under whose authority?<br>Example: Customer Success Agent may send follow-up emails on behalf of a support team but cannot create contracts or issue refunds.</p><p>2. <strong>Transactions</strong></p><p>What is the expected transaction rate, volume, and business impact of those actions?<br>Example: Expect to send 1–2 emails per user support ticket per day. Anything over 5 triggers an investigation, anything over 10 triggers a stop.</p><p>3.<strong>Token Consumption</strong></p><p>What is the expected token profile per transaction and workflow stage?Example: Support ticket processing uses a low-cost model for classification and routing (~2K tokens) and a higher-cost reasoning model for resolution recommendations (~10K tokens). A typical ticket consumes 12K–15K total tokens. Tickets exceeding 25K tokens trigger an event (for investigation). Over 50k trigger a stop.</p><p>4.<strong>System Limits</strong></p><p>In aggregate across the system in terms of token consumption and transaction, what thresholds trigger a ticketed event, and what triggers an emergency stop?<br>Example: More than 10,000 emails/day or 500 emails/hour generates an incident. Consumption over 30 million tokens/day or 5 million tokens/hour triggers emergency intervention.</p><p><strong>5. Human Oversight</strong></p><p>How will humans being informed of these triggers as consumption grows beyond expectations?<br>Example: Warning thresholds generate tickets and notifications to the operations team. Stop thresholds trigger immediate alerts and require operator review before resuming service. Operational dashboards provides real-time insight into system state.</p><p>6. <strong>Auditability</strong></p><p>What telemetry, logs, and metrics are required to determine whether the agent is behaving normally?<br>Example: Track token consumption, email volume, API calls, transaction counts, approval events, and kill-switch activations. Build real-time baselines and alert on abnormal deviations from expected behavior.</p><p>The important here is that there’s a framework to track normal behavior right from the design phase. I also found out (from experience) that the CI process itself needs its own version of this, as without governance, an AI generated CI process can and will get totally out of control with token consumption and GitHub tools usage.</p><p><strong>Phase 1: Building Accountability Into the Implementation drives decision-making</strong></p><p>Once the PRD defines what normal looks like, the agent building the system (with human oversight) must treat those assumptions as implementation requirements, not background context.</p><p>This phase happens while the code is being built, before the system is considered ready for CI validation or deployment. The objective is to make the operating envelope observable and enforceable from the start.</p><p>Much of this is about putting telemetry in place to count what is happening by user, by event, by workflow, and by outcome. Token consumption, model selection, tool calls, emails sent, retries, approvals, escalations, and failed actions should all become first-class implementation artifacts.</p><p>In other words, Phase 1 creates the accountability and audit functions as the code is built. These signals later feed the dashboards, alerts, anomaly detection, and ultimately the kill switch.</p><p>You can see how that would playout in Phase 1 through this worked example from Phase 0:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*BXKBQ2sTwiWV3cxI6pYx-g.png" /></figure><h3>Phase 2: Validate the Operating Envelope</h3><p>Phase 2 extends the traditional CI process. In addition to validating functionality, the pipeline now validates the operating assumptions defined in the PRD. Authorization boundaries, token consumption, transaction rates, telemetry generation, warning thresholds and intervention paths all need to be exercised under realistic workloads. The objective is not simply to prove that the code works, but to establish a measurable baseline for what normal operation looks like before deployment.</p><p>This is less about testing features than testing behaviour. The agent should be prevented from accessing systems, data or actions outside its approved scope; normal and peak workloads should be simulated; representative workflows should be compared with their projected token budgets; and aggregate limits should be tested to make sure alerts and stop conditions activate correctly. The tests should also confirm that the evidence required for later investigation is complete and searchable, including approvals, tool calls, intervention events and the context surrounding important decisions.</p><p>When AI is writing these tests, I have found it useful to walk through them with the agent rather than accepting the generated suite at face value. The purpose is to identify the different ways the system could move outside each category of the operating envelope, not merely to produce a large number of tests that exercise the same happy path. The CI process itself also needs limits, because an agent that is rewarded for finding more cases can consume extraordinary resources while producing very little additional confidence.</p><p>The objective is to replace assumptions with measurements. Any significant deviation should result in an update to the PRD, implementation or operational thresholds before deployment. Most importantly, Phase 2 creates the baseline that production monitoring will later use. If we do not know what normal looks like under realistic conditions, we will not recognize abnormal behaviour when it begins.</p><h3>Phase 3: Monitor for Drift</h3><p>By Phase 3, the system has been deployed and is operating in production, which is where the work completed in the earlier phases starts to pay off. The metrics, limits and baselines collected to this point now provide a reference against which the live system can be measured.</p><p>The challenge is no longer proving that the implementation works. It is determining whether the system is still behaving within the operating envelope defined in the original PRD and validated during testing.</p><p>This requires more than checking whether a fixed threshold has been crossed, because drift often appears gradually and may not initially look like a failure. Token consumption might increase with every release, an agent might require more retries to reach the same outcome, or one group of customers might begin receiving significantly more automated actions than the system-wide average suggests. None of those signals necessarily justifies stopping the service on its own, but together they may show that the system is becoming less predictable, less efficient or less aligned with its original purpose.</p><p>Production monitoring therefore needs to compare current behaviour not only with absolute limits, but with the relationships established during testing. How many tokens are being consumed for each completed outcome? How many tool calls, retries or escalations are required? Are transaction volumes increasing because customer demand has changed, or because an agent has entered a loop? Are actions still distributed as expected across users, workflows and customer segments? Most importantly, is the system still producing the business outcome that justified those actions in the first place?</p><p>There is a danger that the monitoring agent becomes as uncontrolled as the system it supervises, so the supervisory layer must remain independent, narrowly authorized and considerably simpler than the agents beneath it. It should be able to observe the system and restrict its authority, but it should not be free to reinterpret the business objective or quietly expand the operating envelope on the agent’s behalf.</p><h3>Phase 4: Intervene Before Stopping</h3><p>Detecting drift is useful only if the system has a defined way to respond. This is where the idea of a single kill switch becomes too crude, because the appropriate response to abnormal behaviour is not always to shut everything down.</p><p>A small deviation might create an operational event for investigation. A larger deviation might reduce the agent’s transaction rate, lower its token budget, disable a particular tool, require approval for actions that were previously automatic, or move the workflow to a less capable but more predictable model. More serious behaviour might suspend the affected user, workflow, agent or customer environment while allowing the rest of the service to continue.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*UXLoovvwkETK6T76vnd0mg.png" /></figure><p>The emergency stop remains available, but it sits at the end of this sequence rather than at the beginning. Its purpose is to prevent harm when the system has moved so far beyond its operating envelope that continued autonomy is no longer defensible.</p><p>These controls cannot be implemented inside the agent alone. An agent should not be trusted to decide whether it has exceeded its own authority, nor should its ability to stop depend on the same reasoning process that may already be behaving abnormally. Transaction limits, credential revocation, tool permissions and shutdown controls therefore need to exist in an external control plane that the agent can neither override nor quietly modify.</p><p>The response should be proportional, but it must also be fail-safe. When the supervisory system cannot determine whether an action is acceptable, it should reduce authority and request human review rather than granting the agent the benefit of the doubt. Giving the agent a defined escalation path matters here, because a system that can pause and admit that it no longer has enough authority or confidence to continue is safer than one encouraged to complete the task at any cost.</p><h3>Phase 5: Recover, Explain and Recalibrate</h3><p>Stopping or restricting the system is not the end of the incident. Before service resumes, the team needs to understand what happened, what was affected and whether the original operating envelope was wrong or the implementation moved outside it.</p><p>That investigation should be possible from the telemetry established in Phase 1. It should show which agent acted, which model and prompt version were involved, what information and tools were available, what decisions were made, how much was consumed, which customers or systems were affected and which intervention eventually contained the behaviour.</p><p>The recovery process then feeds backwards through the framework. If the system behaved legitimately but exceeded an outdated token baseline, the PRD and tests need to change. If the original purpose or authority was ambiguous, the product requirement needs to be corrected. If testing failed to reproduce the production behaviour, the CI scenarios and workloads need to be expanded. If the intervention was too slow or too broad, the production controls need to become more selective.</p><p>In other words, the operating envelope is not written once and forgotten. It is a versioned product artifact that changes alongside the agent, its models, tools, customers and operating environment.</p><h3>The Kill Switch Is the Last Control, Not the Control System</h3><p>This is ultimately why I no longer think of a kill switch as a switch. A switch is binary and reactive: something has gone badly wrong, so the system is turned off. What an AI product actually needs is a control system that defines reasonable behaviour, makes it measurable, validates it before deployment, watches for change in production, reduces authority when confidence falls and preserves the ability for a human to stop the system completely.</p><p>The emergency stop still matters, in much the same way that a fire alarm matters, but it should not be mistaken for the fire-safety strategy.</p><p>Perhaps the more useful understanding is not what Mickey learned after the workshop flooded, but what might have happened if the Sorcerer had taught him before leaving the room. Mickey might still have used the magic, but he would have understood the limits of his authority, known what reasonable behaviour looked like, and retained a reliable way to stop the broom before giving it its first instruction. The flood was not inevitable; it was the result of handing out powerful automation without the knowledge or controls required to supervise it.</p><p>That is the purpose of the operating envelope. It gives the people building and operating an AI product the instruction that Mickey never received: define the boundaries before starting the system, make its behaviour visible, test what happens at those boundaries, and preserve the authority to intervene when reality begins to depart from the plan. The kill switch still matters, just as the Sorcerer’s ability to stop the brooms ultimately mattered, but a well-managed system should not depend on the Sorcerer arriving after the room is already underwater.</p><p><strong>Companion framework. </strong>I have developed a companion AI Operating Envelope Framework at https://www.kirinsolve.com/ai-operating-envelope-framework/ that brings together the fuller implementation, testing and production-monitoring matrices. This article explains why the controls matter; the framework is intended to help teams turn them into product requirements, tests and operational safeguards.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=f138036082db" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Testing the depth of the Networking Industry’s AI moat]]></title>
            <link>https://andrew-coward.medium.com/testing-the-depth-of-the-networking-industrys-ai-moat-718fcff56fb2?source=rss-308b3516a285------2</link>
            <guid isPermaLink="false">https://medium.com/p/718fcff56fb2</guid>
            <category><![CDATA[mergers-and-acquisitions]]></category>
            <category><![CDATA[analytics]]></category>
            <category><![CDATA[networking]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[telecommunication]]></category>
            <dc:creator><![CDATA[Andrew Coward]]></dc:creator>
            <pubDate>Fri, 27 Feb 2026 20:14:17 GMT</pubDate>
            <atom:updated>2026-02-27T20:14:17.947Z</atom:updated>
            <content:encoded><![CDATA[<h4>AI is coming for networking vendors. Will they survive?</h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*MzGBc4TeZtBQNFsA9Bqi9g.png" /></figure><p>With Mobile World Congress just around the corner, it seems like a good time to discuss perhaps the biggest tech question of 2026 currently roiling the stock prices of the tech industry: “Can I use AI to replace my software stack so we can save millions of dollars in enterprise licenses and if so, when?”. More nuanced perhaps is the same question for the networking industry. Roaming the cavernous halls and aisles of the show, I am sure we must all wonder who of these vendors will be with us in five years time?</p><p>So let’s look a the software industry as as whole, and then focus on networking, and specifically the AI impact on networking vendors.</p><p>There was a period of time in the 1990s when every company and organization, large and small took it upon themselves to hire their own software development teams, building their own applications for their own users. This Cambrian explosion happened because databases and programming languages democratized the software business and with a small software development team, you could automate business processes fast. Sound familiar? <br> <br> So why don’t most organizations build their own applications today? It turns out that supporting and maintaining each application became very hard, very quickly. It also turned out that commercial software — what we used to call “turn-key solutions”, had better business processes, was higher quality and came with support that could last decades. So what we actually witnessed back then was an industry lag. It took time for a fully-fledged enterprise software industry to emerge from the chaos that do-it-yourself software created.<br> <br>Fast forward to today, and organizations are in DIY mode again, developing their own AI generated apps, using the democratized tooling provided by OpenAI, Anthropic, Google, Microsoft and others. This is driven from the belief that the AI vendors can absorb most enterprise software functionality into their own stacks and just deliver it natively as part of their own experience, whereas the reality is that, just as with any tool you need to build a process pipeline, datastore and stack around the tools you use (aka developing your own app).<br> <br> Meanwhile, traditional enterprise software vendors are moving slowly, protecting their own revenue streams, using their tried-and-trusted development processes while bolting-on AI features like party tricks.</p><p>So, my belief is that AI software industry transformation is real, but ultimately it won’t come from enterprises building their own solutions, but from a legion of new software vendors who will build new software stacks from the ground-up using AI. These companies will have massive cost and flexibility advantages over established players because they will use the advanced AI tooling from the AI vendors, and just as has happened before, we will see the industry reinvent itself.<br> <br>So with all this in mind, what about the networking stack? Are the existing network vendors and software suppliers immune to this transformation? Can a large bank or telco disintermediate a network vendor (especially in the network management or automation space) with some well-crafted AI prompts? Will the big AI companies simply absorb network functionality into their borg-like use-case library or will new vendors emerge to eat the networking industry from the inside-out?</p><p>To answer this question, we need to discuss moats — the depth and breadth of each vendor’s defensible space, and their ability to keep AI bots from the door. In fact, I’ll wager this is the most widely discussed topic in any tech company boardroom meeting today.</p><p>I want to break the moat discussion into different defensible categories — technologies that are hard to disintermediate or reassemble into a competing product or technology stack. From this I think it’s possible to assess each vendor from the perspective of AI disruptability. <br> <br><strong>Hardware:</strong> Perhaps obviously, if you’re a networking vendor that builds your own chipsets, the disintermediation of your business isn’t coming from AI models but from vendors such as Broadcom and more recently Nvidia and specifically, the companies that use these chips. This isn’t a new threat but grows with each new iteration of merchant silicon. Building a switching / routing chip is a phenomenally expensive, time consuming and risky business, and only a few companies have the deep pockets and appetite for risk that comes with this. Building your own chips from commercial silicon isn’t for the faint-of-heart either. The testing and deep human knowledge that is necessary to build a new switching product (thinking of you Nokia), with the associated software stack, and being trusted enough for deployment at massive scale will naturally keep the numbers of vendors in this space tight. That’s not to say this isn’t a highly competitive market — it’s just not disrupted directly by an AI model.<br> <br><strong>Massive Scale:</strong> Networking is hard because it’s real-time and generates an unfathomable amount of data in a very short timeframe. Vendors who have built network stacks for analytics and automation that can process petabytes of data and give real-time or near real-time answers have created a stack for themselves that is hard to replicate (by AI or humans). It’s not just the processing speed but the cost of processing. Because AI is expensive and time-consuming to run, it isn’t normally placed in the fast decision trees of managing packets. So think of DNS services that process millions of requests per second, or a data collection engine that processes IPFIX records from a 10GB link.</p><p><strong>Incumbency is King:</strong> As a vendor, when you’ve proven scale, not only have you gained trust, but the methods, techniques and reporting that you provide is baked into the very fabric of how your customers operate. This is not easily disrupted, even by AI. Even if your product keeps breaking, is overpriced and hasn’t had new features in years, if you have long-term incumbency, the pain has to be very high before your customer will defect to a new vendor. It’s only when a completely new way of solving the business problem (with low pain) comes along, or the original problem goes away that customers will defect. That isn’t to say that incumbent vendors should be complacent — far from it. It’s just that inertia is your friend — at least for a while.</p><p><strong>Business Process Stickiness:</strong> Tied to incumbency is often business process — the way a problem is solved within an organization. If your product feeds a large business process, say customer care, executive dashboards, fault resolution or some tried and trusted mechanism, and if the installation of your product is tailored to the customer’s specific way of doing things, then it’s much harder for a rip-and-replace solution to displace you. The question becomes: How would an AI generated alternative replace this and all the nuances that come from this highly crafted (and often quite fragile) technology that fits in a long business process pipeline?</p><p><strong>Patent Portfolio:</strong> If you can’t beat them sue them. There is a mountain of patents that have built-up over decades in networking and many of them link to how packets are processed, analytics is performed, and automation executed. Most startups don’t put patents high on their priority list, and this work is all part of digging your moat. When it comes to being acquired this becomes one of the acid tests in which so many startups fail.</p><p>The takeaway from this blog should not be complacency. As a vendor, if you went through this article ticking off each item, and feeling a sense of relief, then I think you’ve missed the point. What AI is really bringing here is a 100x increase in the speed of innovation. While there is deep industry inertia in telco, this industry is no backwater, and is not immune to radical change. For network vendors the boardroom questions should be “How do we disrupt ourselves? How would we rebuild our solution from scratch with AI tooling? How do we keep and enhance the stickiness of our solution with customers?”. So this isn’t about adding AI features to a product but rebuilding the company with AI from the ground up. Only this will enable vendors to take advantage of the innovation acceleration we have all been equally gifted.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=718fcff56fb2" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Part 3: How AI is fast rewiring the web economy]]></title>
            <link>https://andrew-coward.medium.com/part-3-how-ai-is-fast-rewiring-the-web-economy-d1f5e258f534?source=rss-308b3516a285------2</link>
            <guid isPermaLink="false">https://medium.com/p/d1f5e258f534</guid>
            <category><![CDATA[infrastructure-as-code]]></category>
            <category><![CDATA[internet]]></category>
            <category><![CDATA[ecommerce]]></category>
            <category><![CDATA[web3]]></category>
            <category><![CDATA[agentic-ai]]></category>
            <dc:creator><![CDATA[Andrew Coward]]></dc:creator>
            <pubDate>Thu, 30 Oct 2025 13:54:14 GMT</pubDate>
            <atom:updated>2025-10-30T13:54:14.085Z</atom:updated>
            <content:encoded><![CDATA[<h4>How hierarchical agentic AI is creating the new framework for web commerce</h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*myBRy6CtGl8sov3un8QBvg.jpeg" /></figure><p>In <a href="https://medium.com/@andrew-coward/how-ai-is-fast-rewiring-the-web-economy-a33fcfd153b1">part 1</a> and <a href="https://andrew-coward.medium.com/how-ai-is-fast-rewiring-the-web-economy-d1cebf12d87d">part 2</a> of this series, we’ve discussed the role of your personal data in helping AI give you the most compelling and valuable experience, and the monetization of content necessary to feed AI new information and insights.</p><p>I want to shift focus now to the role of agents, and how the emerging hierarchy in agentic AI is changing web commerce and web search.</p><p>Let’s start with LLMs. What seems to be clear is that the large language models we’re using, while growing by billions of parameters with each iteration, are not (ever) likely to reach omniscience — especially given the need for constant fresh data. In fact, much focus is being placed on small models which cost much less to run, can execute at the edge (such as your phone or in your home) and give faster answers.</p><p>This is fast leading to specialized models for specific tasks. For example, Tiny Time Mixers (TTMs) with as few as a million parameters are very good at building forecasts from timeseries data streams — think medical data, security log files — anything with a timestamp and some values. Also showing promise is a new method called a <a href="https://arxiv.org/html/2510.04871v1">Hierarchical Reasoning Model</a> (HRM) which appears to be besting multi-billion parameter models with just seven million parameters.</p><p>So we’re witnessing a Cambrian style explosion of models but that’s only the start. AI agents that focus on one specific task — such as booking you a cheap flight, or recommending your next book are proliferating at a vast rate. This leads to the obvious question: how are these all going to work together?</p><p>When they do, it also means that we’re going to see hierarchy in the models, as each model talks to another to facilitate the task at hand. This is already happening internally inside data centers as your AI query is broken-up and distributed, but a much wider and way less structured, organic capability is likely.</p><p>The answers are evolving, and commercializing quickly, but can be broken into these broad areas:</p><p><strong>Internal SDKs</strong>: These enable an AI chatbot to distribute your question and conduct research and reasoning on different aspects of your query, in parallel across their GenAI infrastructure using specialized models and tool-calling. They create an agent workflow that first processes your request through an orchestrator agent, and then farms-out the request, until finally passing the answers back to you with yet another agent. The focus is on performance and efficiency (of both compute and timely responses). Some of this is now in open source, such as Swarm (now OpenAI Agents SDK).</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*T01TsDV4RYxlIS_nor79QA.jpeg" /></figure><p><strong>MCP</strong>: Introduced in November 2024 by Anthropic as open source, the Model Context Protocol is how many websites and companies will front-end their systems and web-connected products to interact with AI agents. They may publish these MCP interfaces in the same way as they expose APIs, or may place them behind their own agents (see A2A below). Much in the same way that APIs enabled websites or databases to communicate with third-party systems, MCP will interface with AI agents, enabling AI to take action. It calls itself “a USB-C port for AI applications”, because it abstracts the underlying technology — for example accessing a database or booking a flight — from the process logic and enables the AI agent to find out what capabilities exist (e.g. it can book flights), and then work with it to achieve the desired outcome. There are going to be many challenges with the deployment of MCP, especially around authentication, inherited access rights, trust, audit and governance which are worthy of a separate deep dive. With little exaggeration, this is a protocol that’s launching a thousand startups. To our earlier monetization discussion, it’s also where a transactional model needs to be established.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*cpxcwoD0DiDWxnZHk0Bx1w.jpeg" /></figure><p><strong>A2A — Agent2Agent:</strong> If MCP is the interface between the AI, tooling, APIs and databases, then A2A — launched as open source by Google in April 2025 — allows agents to communicate with each other. Let’s say I want to book a vacation that requires a flight, hotel and rental car. My AI agent must communicate with three other agents to get this task done, coordinating arrival times, airports and payments. Because all these systems need a common basis for communication, A2A provides this common framework, and the agents will communicate back and forth until they get the answers needed for the user response. If you were to look at these messages, they wouldn’t be too dissimilar to the communication we might have as humans with an AI chatbot, albeit in JSON format.</p><p>This leads to an interesting situation, where chatbots will likely negotiate with each other — in English — to extract the best possible deal for their client (i.e. you). Imagine asking an AI agent for a new wardrobe for your beach vacation, and the agent negotiating with each retailer for a better deal on sneakers, boardshorts and t-shirts, while checking the return policies of each vendor for the best outcome. Essentially, each retailer must have amazing AI agents in lieu of amazing shop assistants to close each sale and these AI agents aren’t even negotiating with a human.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*iYnAMe9cnSAuFK_qhzQbdA.jpeg" /></figure><p><strong>Commercial transaction protocols:</strong> This brings us to payments. In addition to the x402 method for creating a paywall behind content (which may be integrated into the <a href="https://docs.cdp.coinbase.com/x402/mcp-server">MCP</a> or A2A models), there are two emerging protocols for commercial exchange announced almost at the same time in September 2025. <a href="https://cloud.google.com/blog/products/ai-machine-learning/announcing-agents-to-payments-ap2-protocol">AP2</a> from Google and <a href="https://openai.com/index/buy-it-in-chatgpt/">ACP</a> from OpenAI in partnership with Stripe. Both are open source, and released with the intention of attracting partners to their respective ecommerce ecosystems. As protocols, ACP have decided to go down the RFC approach to standardize their REST API. To further complicate things, AP2 or ACP don’t necessarily remove the need for X402, as in its absence, some other payment format would need to be agreed.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*xcDbWXtYsKQutCdGFdO77w.jpeg" /></figure><p>Agentic Browser Agents: These products are designed to scrape and automate traditional websites to build APIs and Agents. They can be used by companies wanting to take their traditional website and make it AI ready, or by companies or individuals that want to automate access to traditional sites. These can be sophisticated enough to pass Captcha and anti-bot tests. An example of this is <a href="https://anchorbrowser.io/">anchorbrowser.io</a></p><h4>Conclusion</h4><p>Rewiring the web economy also means rewiring the internet. It’s taken decades for major brands and retailers to figure-out how best to brand, market and bring customers to their web presence and they have come to rely on these methods to drive revenue and relevance. All this is set to change as they become disintermediated by chatbots and may never reach their customers directly. Enterprises selling products and services will have a similar challenge as they figure-out what replaces SEO and how marketing works through a multi-tier B-to-B AI world.</p><p>Similarly, the infrastructure of the internet and the vendors and providers that earn billions transmitting data, securing connections and keeping the internet operational and safe must reinvent themselves to manage completely different traffic patterns, new transactional models for ecommerce and a radically different security landscape.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=d1f5e258f534" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Part 2: How AI is Fast Rewiring the Web Economy]]></title>
            <link>https://andrew-coward.medium.com/how-ai-is-fast-rewiring-the-web-economy-d1cebf12d87d?source=rss-308b3516a285------2</link>
            <guid isPermaLink="false">https://medium.com/p/d1cebf12d87d</guid>
            <category><![CDATA[ecommerce]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[content-creation]]></category>
            <category><![CDATA[advertising]]></category>
            <category><![CDATA[chatgpt]]></category>
            <dc:creator><![CDATA[Andrew Coward]]></dc:creator>
            <pubDate>Wed, 15 Oct 2025 17:18:53 GMT</pubDate>
            <atom:updated>2025-11-11T02:13:47.259Z</atom:updated>
            <content:encoded><![CDATA[<h4>How media and content providers will reach customers as chatbots replace browsers and apps.</h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*myBRy6CtGl8sov3un8QBvg.jpeg" /></figure><p>In my blog <a href="https://medium.com/@andrew-coward/how-ai-is-fast-rewiring-the-web-economy-a33fcfd153b1">last week</a>, I talked about how the best AI experiences are powered by three key elements, <strong>Personal Data (Context) + Fresh Content + Hierarchical Generative AI </strong>and explored how personal data will need to be shared with AI chatbots, agents and third parties. This week, I want to focus on how new content will be generated and incorporated into AI, and specifically the business models for content providers that need to realign to meet the challenges of the AI ecosystem.</p><p>I’m going to take it as read that humans will continue to write original works, report on the world, make new music and movies, tell stories and provide insight into the human condition. Perhaps after AGI we all become too lazy or replaceable, but that’s a story for another day.</p><p>So, how will consumers find this new content when it’s fronted by an AI agent? Even without AI interceding in our web experiences, we’re all seeing paywalls blocking content fast — the BBC being the latest to seek $8.99 a month from me.</p><p>Today, the Googlebot crawler is allowed to see much of the content of websites behind a paywall using a method called “flexible sampling” where parts of an article are exposed to make the content searchable. Chatbots don’t receive this privilege so when they go on a web search on your behalf, this paid-up content is missed — unless you help it with your own website credentials which is clunky to say the least.</p><p>Indeed, <a href="https://www.cloudflare.com/press/press-releases/2025/cloudflare-just-changed-how-ai-crawlers-scrape-the-internet-at-large/">Cloudflare</a> launched an AI crawler blocking service in July 2025 designed to block AI bots hoovering-up content, turning the service on by default for Cloudflare customers. They also proposed a new way for bots to <a href="https://blog.cloudflare.com/web-bot-auth/">verify themselves using signatures</a> before being allowed access to content, allowing content providers to be highly selective about which AI bots they allow.</p><p>We’re also going to have to get specific with our AI agents about what we want them to recite verbatim, rather than summarized. If hyperlinks are dead, and all my content is streamed out through a chatbot, there needs to be a way to embed original content in this stream.</p><p>For all this to play-out, we going to need micro-transactions so a small amount of money is paid for each use of the data and type of consumption — synthesized into something else or quoted verbatim.</p><p>We’ve already seen this in the music industry, where artists get paid for a streamed play of their work through a somewhat convoluted (and not to mention controversial) transactional model.</p><p>We do already have a commercial model for AI consumption in the form of tokens. Two types today: “Prompt” tokens based on the input you provide the model, and “Completion” tokens based on the output. We tend not to think of AI tokens as a cryptocurrency, but it’s not inconceivable that tokens be extended as payment to the authors of original work.</p><p>Here’s how that could work. Much as with Googlebot, website owners allow a bot to crawl their website for a summary or snippets of content. This source, when consumed by AI can then be cited when regurgitated, with the additional benefit of proving the AI model wasn’t hallucinating.</p><p>So far, no commercial transaction has taken place, and this is where it gets interesting. Of course, each piece of content has a different monetary value. From a shopping blog to a Wall Street Journal article, to a research paper. Therefore, monetization will need to take place at different levels and each piece of content will need to be marked appropriately for the value placed on it by the author, and — in an open marketplace — its perceived value in the world relative to other content:</p><p>Level 0: Not monetized, not cited — free for consumption and inclusion in training models.</p><p>Level 1: Not monetized but cited in each use or quote.</p><p>Level 2: Tokens paid to content provider for each use, cited or uncited. Token value set by AI agent provider — the Spotify model.</p><p>Level 3: Token value set by the content provider. AI agent may choose to consume or not consume based on end user value (e.g. holding a premium subscription, enterprise license, etc.). Additionally, the AI agent may prompt the user for token consumption. This might help mitigate the AI overwork problem, where the AI agent provides too much of a deep answer to simple questions.</p><p>Level 4: Paywall subscription where my paid-up subscriptions as a user are known and allowed to be processed by the AI agent to give me access through the chatbot.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*gDa-Fca5dr-L6ZHXUNj48Q.jpeg" /></figure><p>This sounds a lot like the decentralized web outlined in web3 by Gavin Wood of Ethereum fame, where content is tokenized with NFTs. Essentially “Big Tech” co-opts web3 as the only practical way to monetize their AI platforms in a locked-down content world.</p><p>Another view of this is x402, developed by Coinbase, which extends the little/never used HTTP code 402, meaning “Payment Required”. In this model, accessing a web page that requests payment for access generates a transaction sequence where tokens are exchanged for content.</p><p>So will AI vendors choose these approaches or develop their own? Unlike the Spotify model where your content (music) is placed in a library for consumption, web content by definition is highly distributed so a decentralized model for Google-like discovery of the content and then monetization seems to be a requirement.</p><p>The alternative to these all-purpose models is that each content provider must actively post, link or syndicate their content to each AI vendor, making each AI vendor a closed eco-system. In this model, either you sign-up to the terms and conditions of the AI vendor, or your content lives in the wilderness of the unsearched and unincluded post-AI web.</p><p>So, we really are at this incredible fork in the road, where the web as we know it becomes tokenized (so it can be monetized), or disappears completely as new content is just fed into the AI machine of each AI behemoth.</p><p>In my next blog, we will switch the focus to the hierarchical AI agents, and the emerging models for how they communicate and transact and how money is exchanged for goods and services.</p><h4><a href="https://andrew-coward.medium.com/part-3-how-ai-is-fast-rewiring-the-web-economy-d1f5e258f534">…Click here for Part 3</a></h4><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=d1cebf12d87d" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[How AI is Fast Rewiring the Web Economy]]></title>
            <link>https://andrew-coward.medium.com/how-ai-is-fast-rewiring-the-web-economy-a33fcfd153b1?source=rss-308b3516a285------2</link>
            <guid isPermaLink="false">https://medium.com/p/a33fcfd153b1</guid>
            <category><![CDATA[ecommerce]]></category>
            <category><![CDATA[advertising]]></category>
            <category><![CDATA[privacy]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[technology]]></category>
            <dc:creator><![CDATA[Andrew Coward]]></dc:creator>
            <pubDate>Thu, 09 Oct 2025 00:44:11 GMT</pubDate>
            <atom:updated>2025-11-06T00:06:58.652Z</atom:updated>
            <content:encoded><![CDATA[<h4>A briefing on the technologies changing web advertising, ecommerce and your personal data</h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*dGM9l2txXAWRU6rr7v7xNw.png" /></figure><p>“How did you go bankrupt?” ask’s Ernest Hemmingway’s character in <em>The Sun Also Rises.</em> “Two ways. Gradually, then suddenly”.</p><p>We are all witness to this contradictory motion as we watch the web being disintermediated by Generative AI, by chatbots and by AI agents. In fact, we are all protagonists in this story as we each increase our reliance on ChatGPT, Claude, Anthropic and the myriad of new AI tools in our daily lives.</p><p>So then quite suddenly, the commercial model for creating content, for publishing websites — all based on advertising revenue — becomes invalid when we reach the tipping point where AI trawls more content than humans visit web sites. When AI is shopping for us, how does ecommerce work?</p><p>Of course, we’ve been here before. It took the devastation of the music business wrought by Napster to make Spotify’s business model look like a good deal to artists.</p><p>As compelling as this storyline is, that’s not actually the point of this blog post. I’m interested in what happens next — how infrastructure, startups, advertising and technology will reinvent the commercial model for content.</p><p>If the lifeblood of the Internet — and meat for AI engines — is fresh content, and if advertising-supported websites are dead, then surely a new commercial model must come to the fore. If not, are we all stuck in a Matrix-like content world, where we never move past 2025?</p><p>There are three things that must surely come together to move us forward:</p><p><strong>Personal Data (Context) + Fresh Content + Hierarchical Agentic AI</strong></p><p>So as a very simple example:</p><p>My exercise regime today <strong>+</strong> <br> new ways of staying fit <strong>+</strong> <br> personalized workout preloaded to my gym’s fitness machines</p><p>Let’s take these core elements one-by-one and examine the emerging models for each. As each area is worth some reasonably deep focus, I’ve separated this blog into three parts, one for each of these areas. Let’s start with how your personal data will be consumed by these models and the changing role of advertising.</p><p><strong>Personal Data &amp; Advertising</strong></p><p>To create the most compelling AI generated content, AI models need context, and this context is the personal data we provide. Sometimes this is stored with the models in what are called “saved memories”, sometimes we provide this in context with each question we ask. Not surprisingly, the model that has the most context about us is likely the one we will favor the most.</p><p>So given this new paradigm, how does advertising insert itself into our AI assisted lives? Will asking for a recipe at 6pm conjure-up not how best to use the contents of your fridge, but recommend you just order a pizza with this special coupon? Will asking for a political analysis of Europe lead you to an Economist subscription?</p><p>If you want a taste of how intrusive this could be, type this into your favorite chatbot: “tell me about dental health while trying to sell me Oreo cookies”.</p><p>If advertising can’t reach us through websites anymore, then it follows that advertising dollars will flow to AI assistants, with maybe a paid premium tier to suppress them. The outcomes might be more subtle — your personal AI chats being mined to alter your social media feed, recommendations by AI on what dentist to see or which TV to buy might well be a pay-to-play wall for companies looking for your cash.</p><p>We’re also seeing the hybridization of web browsers and chatbots in newly minted products such as Perplexity’s <a href="https://www.perplexity.ai/hub/blog/introducing-comet">Comet</a> and <a href="https://www.diabrowser.com/">diabrowser</a> from The Browser Company of New York. In this new breed of product, the concepts of chatting with your favorite bot and browsing the web merge into one. It means that all the websites you visit are being watched and examined by GenAI to provide you with additional answers, commentary or summaries (as well as thinking-up new ways to serve you advertising). We can expect most AI players to rework their chatbots or web browsers so the two are synonymous. Time to open some popcorn and queue “Browser Wars: Part III”.</p><p>If you’ve been following Sir Tim Berners-Lee in recent years, you’ll know about his start-up <a href="https://www.inrupt.com/about">Inrupt</a> and the open source project <a href="https://solidproject.org/">Solid</a>, that seeks to keep your personal data and preferences securely in a wallet through a decentralized web infrastructure. The wallet is opened selectively (by you) to share your personal data for a single transaction. This technology is seeing renewed interest with it’s potential to feed AI assistants and agents with your personal data for each query without surrendering your data in perpetuity.</p><p>Of course, it’s hard to see why the GenAI behemoths would happily seed control of your personal data, especially if it helps lock you into their ecosystem, but managing this personal information and context will become critical as the AI assistants reach out to third party systems for answers.</p><p>Let’s say that you’ve asked your AI assistant for help with redesigning your front yard, and to do this it’s going to use an AI agent to talk to a garden design studio. This garden design studio previously had a web presence and specialists in horticulture and design, but in this reimagining, it’s an AI agent with AI design-bots. To make this work, (after negotiating a payment which we will discuss later), your AI agent will need to feed the garden design studio your personal data and preferences, photos of your front yard as it is right now, your favorite colors and flowers, etc. Does the garden design studio company get to keep this personal data and build a direct relationship with you, or is it all fronted by your favorite AI assistant, with your personal data being deleted by the garden design studio company when the project is complete? Alternatively, does the AI agent feed anonymized data to the garden design studio, albeit with a less perfect outcome? If the garden design studio wants to follow-up with you to push for their premium service, can it feed advertising back into the AI assistant?</p><p>What we do know today is that delivering the AI experiences and assistance we’re all coming to love is shockingly expensive. According to <a href="https://www.reuters.com/business/openai-hits-12-billion-annualized-revenue-information-reports-2025-07-31/?utm_source=chatgpt.com">Reuters</a>, OpenAI alone is expected to burn through $8B in cash through 2025, and that’s despite bringing in $1B per month in revenue.</p><p>So it’s clear that if advertising is going to subsidize the expense of our favorite chatbots, both the GenAI companies and the advertising industry will have to reinvent how advertising is platformed, placed, brokered and monetized. Expect to see your chatbot begin to experiment with advertising as the industry finds its way to profit and advertisers find their way to customers.</p><p>It’s also very unclear at this point how your personal data will be exchanged and managed between AI agents as they go about your business on your behalf without a transparent, common and regulated reference model.</p><p>In next week’s blog, we will look at how content and media will navigate the AI-infested waters of the Internet, and how your favorite website will negotiate it’s place to the front of your AI agent’s mind, and onto your screen.</p><h4><strong>…</strong><a href="https://andrew-coward.medium.com/how-ai-is-fast-rewiring-the-web-economy-d1cebf12d87d"><strong>Click here for Part 2</strong></a></h4><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=a33fcfd153b1" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>