<?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 Carla Stone on Medium]]></title>
        <description><![CDATA[Stories by Carla Stone on Medium]]></description>
        <link>https://medium.com/@carla_1448?source=rss-cf24ffb761af------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*E_5YhQGtnzxavcFSemgznA@2x.jpeg</url>
            <title>Stories by Carla Stone on Medium</title>
            <link>https://medium.com/@carla_1448?source=rss-cf24ffb761af------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Wed, 18 Nov 2026 17:22:59 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@carla_1448/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[Humans in the Loop Need a Loop Worth Being In]]></title>
            <link>https://medium.com/@carla_1448/humans-in-the-loop-need-a-loop-worth-being-in-fccd5aca6b01?source=rss-cf24ffb761af------2</link>
            <guid isPermaLink="false">https://medium.com/p/fccd5aca6b01</guid>
            <category><![CDATA[decision-making]]></category>
            <category><![CDATA[ai-governance]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[enterprise-technology]]></category>
            <category><![CDATA[future-of-work]]></category>
            <dc:creator><![CDATA[Carla Stone]]></dc:creator>
            <pubDate>Wed, 18 Nov 2026 13:03:00 GMT</pubDate>
            <atom:updated>2026-11-18T13:03:00.310Z</atom:updated>
            <content:encoded><![CDATA[<h4>Human oversight of AI only works when the person approving can see why. Otherwise it’s a signature, not a safeguard.</h4><p><em>By Carla Stone, Founder of ILYVEN</em></p><p>Earlier this month, Thoughtworks added a new term to its Technology Radar and filed it under “Caution”: human-in-the-loop theater. Their definition is blunt. It is “keeping a human in an agent’s loop as an empty approval step rather than a real check,” where the human approves “actions, prompts or diffs they can’t realistically evaluate.” They go further and describe that approval step as “an accountability sink: no one truly owns the decision” (<a href="https://www.thoughtworks.com/en-us/radar/techniques/human-in-the-loop-theater">Thoughtworks</a>).</p><p>That phrase has stayed with me all week, because it names something every enterprise rolling out AI agents is about to discover.</p><p>For the last few years, “human in the loop” has been the default answer to AI risk. Put a person at the end of the workflow, give them an Approve button, and the system is supervised. Regulators lean on the idea too. Article 14 of the EU AI Act requires high-risk AI systems to be designed so they “can be effectively overseen by natural persons” (<a href="https://artificialintelligenceact.eu/article/14/">EU AI Act, Article 14</a>).</p><p>The important word there is <em>effectively</em>. A human in the loop is only as good as the loop they are standing in.</p><h3>What the research is finding</h3><p>A study posted to arXiv in October by researchers at the National University of Singapore and the MIT Media Lab looked closely at what people actually do when they supervise an AI agent. Participants worked through programming tasks with a coding agent that sometimes produced correct code, sometimes failed visibly and sometimes introduced silent defects. The researchers reported that participants “were often highly confident while missing silent defects.” They performed generic checks, failed to generate diagnostic evidence, or misinterpreted the evidence in front of them. The paper’s title captures it: “Looks Fine to Me” (<a href="https://arxiv.org/abs/2610.27801">arXiv</a>).</p><p>That doesn’t mean the participants were careless. It means the output looked plausible, and plausible is exactly what modern AI is best at producing.</p><p>The EU AI Act anticipates this. The same Article 14 says the people assigned to oversight must be enabled “to remain aware of the possible tendency of automatically relying or over-relying on the output” of a high-risk system, and it names that tendency directly: automation bias.</p><p>So the research and the regulation point to the same uncomfortable fact. Putting a person next to an AI system does not automatically produce oversight. It can produce the <em>appearance</em> of oversight, which is worse, because it lets everyone stop worrying.</p><h3>Why approvals go hollow</h3><p>I don’t think the problem is the human. I think the problem is what the human is given.</p><p>Consider what an approver typically sees when an agent asks for sign-off: a proposed action, maybe a summary, maybe a diff. What they almost never see is the reasoning behind it. Which inputs did the system rely on? Which assumptions did it make? Which earlier decisions is it building on, and what has changed since those decisions were made?</p><p>That is the gap I keep coming back to. AI’s real problem isn’t intelligence. It’s reliability, and a big part of reliability is that these systems forget why decisions were made. When the “why” is missing, a reviewer isn’t checking a decision. They’re checking whether the output <em>looks like</em> a decision. Those are very different jobs.</p><p>I saw a version of this long before AI agents existed. In construction, the costly problems I watched were rarely caused by missing information. They were caused by information that was scattered and disconnected, so the person signing off couldn’t see how one change connected to everything else. An approval made without context is still an approval. It just isn’t a judgment.</p><h3>The pressure is rising</h3><p>Last week, CNBC reported that United Airlines will restrict third-party AI bots from its digital platforms, warning that letting AI assistants access account credentials could lead to unintended purchases or other unauthorized account activity. The same piece quoted Shomit Ghose of the University of San Francisco, who cautioned against letting agents make decisions “anytime [your] health, wealth, or privacy are involved” (<a href="https://www.cnbc.com/make-it/2026/11/12/ai-agents-human-in-loop.html">CNBC</a>).</p><p>That is the consumer version of a question every enterprise now faces. As agents take on more steps, the number of things a human could approve goes up, and the time available to understand each one goes down. If the review model doesn’t change, rubber-stamping isn’t a failure mode. It’s the default.</p><h3>Designing a loop worth being in</h3><p>Thoughtworks recommends triage: reserve human review for changes a human can genuinely judge, using risk-tiered human oversight, and lean on other forms of verification for the rest. I agree, and I’d add a few principles for any organization designing human oversight of AI.</p><p><strong>1. Show the why, not just the what.</strong> An approval request should carry its reasoning: the sources used, the assumptions made and the prior decisions it depends on. If a reviewer can’t trace it, they can’t truly judge it.</p><p><strong>2. Surface what changed.</strong> Many bad approvals aren’t wrong when they’re made. They become wrong when an underlying fact shifts. Reviewers should be able to see when an input has changed since something similar was last approved.</p><p><strong>3. Match the review to the stakes.</strong> Not every action deserves a person’s attention. The expensive, irreversible or hard-to-explain ones do. Spreading human attention evenly means spending it badly.</p><p><strong>4. Make disagreement easy.</strong> If questioning an AI recommendation takes five times the effort of approving it, you’ve designed for approval, not oversight.</p><p><strong>5. Record the human’s reasoning, too.</strong> Oversight should leave a trail. Months later, a name next to “approved” tells you almost nothing. “Approved because the revised lead time still fits the schedule” tells you exactly what to recheck when the lead time moves.</p><p><strong>6. Check whether the oversight works.</strong> If reviewers approve nearly everything they see, ask whether that reflects a great system or a meaningless step. You won’t know unless you look.</p><h3>Where this leaves us</h3><p>“Human in the loop” was supposed to be a promise that a person is responsible. Too often it has become a checkbox that moves responsibility onto someone who was never given the information to carry it.</p><p>The future of enterprise AI won’t be decided by whether a human is in the loop. It will be decided by whether that human can see why.</p><p>That’s the problem I’m focused on at ILYVEN, starting in construction: preserving the reasoning behind decisions so the people accountable for them have something real to review. Today that is active development and a long-term vision, not a finished answer. But the principle is clear to me. Oversight without context is ceremony. Oversight with context is judgment.</p><h3>Thoughts from the Woman in the Red Dress</h3><blockquote>I’ve put my name on a lot of decisions over the years, and the ones I trusted most were the ones where I understood the reasoning. I don’t want a future where people are asked to approve what a machine did without being able to see why it did it. If we’re going to keep humans in the loop, we owe them a loop that actually respects their judgment.</blockquote><p><strong>Carla Stone</strong><br>Founder, ILYVEN<br><a href="https://ilyven.co/">ilyven.co</a></p><p><strong>Sources (verified loading Nov 18, 2026):</strong></p><ol><li>Thoughtworks Technology Radar, “Human-in-the-loop theater” (Nov 4, 2026): <a href="https://www.thoughtworks.com/en-us/radar/techniques/human-in-the-loop-theater">https://www.thoughtworks.com/en-us/radar/techniques/human-in-the-loop-theater</a></li><li>Dissanayake et al., “Looks Fine to Me: Understanding Oversight in Agentic Interactions,” arXiv (Oct 2026): <a href="https://arxiv.org/abs/2610.27801">https://arxiv.org/abs/2610.27801</a></li><li>EU Artificial Intelligence Act, “Article 14: Human Oversight”: <a href="https://artificialintelligenceact.eu/article/14/">https://artificialintelligenceact.eu/article/14/</a></li><li>CNBC, “United Airlines won’t let AI agents book your flight. Expert says you shouldn’t let a bot do that for you, anyway” (Nov 12, 2026): <a href="https://www.cnbc.com/make-it/2026/11/12/ai-agents-human-in-loop.html">https://www.cnbc.com/make-it/2026/11/12/ai-agents-human-in-loop.html</a></li><li>ILYVEN: <a href="https://ilyven.co/">https://ilyven.co</a></li></ol><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=fccd5aca6b01" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Procurement Decisions That Age Badly]]></title>
            <link>https://medium.com/@carla_1448/procurement-decisions-that-age-badly-5e061967069d?source=rss-cf24ffb761af------2</link>
            <guid isPermaLink="false">https://medium.com/p/5e061967069d</guid>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[construction]]></category>
            <category><![CDATA[procurement]]></category>
            <category><![CDATA[construction-management]]></category>
            <category><![CDATA[risk-management]]></category>
            <dc:creator><![CDATA[Carla Stone]]></dc:creator>
            <pubDate>Tue, 17 Nov 2026 13:19:54 GMT</pubDate>
            <atom:updated>2026-11-17T13:19:54.488Z</atom:updated>
            <content:encoded><![CDATA[<p>A purchase order is a decision made with yesterday’s prices, yesterday’s lead times and yesterday’s assumptions. Most projects never check whether it still holds.</p><p><em>By Carla Stone, Founder of ILYVEN</em></p><p>On Friday, the Associated General Contractors of America reported that the producer price index for inputs to new nonresidential construction rose 13.3% from October 2025 to October 2026, the most in more than four years. Steel mill products were up 32.9% year over year, aluminum mill shapes 23.3%, and copper and brass mill shapes 23.9%. AGC’s chief economist, Ken Simonson, added that “lead times for steel and insulation have lengthened, making it harder for contractors to keep projects on schedule” (<a href="https://www.agc.org/news/2026/11/13/construction-input-costs-rise-133-october-2025-october-2026-price-increases-spread-fuel-metals-other">AGC</a>).</p><p>The same day, Associated Builders and Contractors put the year-over-year increase in nonresidential construction input prices at 11.5%, and its chief economist, Anirban Basu, said tariff-affected materials like iron, steel and copper are “still experiencing rapid price escalation” (<a href="https://www.abc.org/News-Media/News-Releases/abc-october-construction-materials-prices-unchanged-energy-and-lumber-prices-fall">ABC</a>).</p><p>Most coverage treats those numbers as a cost story. I think they are a memory story.</p><p>Every one of those price moves lands on a project that already made decisions based on the old number.</p><h3>A procurement decision is a bet on conditions</h3><p>When a team awards a package, approves a product or sequences a long-lead order, it isn’t just buying material. It is making a decision that rests on assumptions: this price holds through buyout, this product arrives before the install window, this manufacturer can still make what was specified.</p><p>Those assumptions are rarely written down next to the decision. They live in an estimator’s spreadsheet, a supplier quote with an expiration date, an email thread or someone’s memory.</p><p>When conditions are stable, that gap doesn’t matter much. When prices move by double digits in a year and lead times stretch, the gap becomes the risk.</p><h3>Three ways procurement decisions age</h3><p><strong>Lead times move.</strong> Procurement data from Terrapin Construction Group, reported by IndustrialSage in September, put 15 kV medium-voltage switchgear at 52 to 80 weeks, against about 24 weeks before the pandemic. Pad-mount transformers in the 0 to 5 MVA range were at 40 to 65 weeks, against roughly 8 (<a href="https://www.industrialsage.com/switchgear-lead-times-2026/">IndustrialSage</a>). A schedule built on the old number isn’t wrong because someone made a mistake. It’s wrong because the world moved and the schedule never heard about it.</p><p><strong>Substitutions ripple.</strong> When the specified product can’t arrive in time, the team looks for an alternate. That’s reasonable. But a substitute is a new decision layered on an old one. The original selection may have been tied to a warranty, a fire rating, a dimension another trade coordinated around, or an owner preference captured in a meeting. If the reasons behind the original choice aren’t attached to it, nobody can check whether the substitute still satisfies them.</p><p><strong>Prices escalate after commitment.</strong> ConsensusDocs notes that in a firm-fixed or lump sum agreement, constructors, subcontractors and suppliers typically bear the risk of cost increases, even unforeseen ones, and that paying higher costs is usually considered a business risk. It also puts it bluntly: “Hope is not a risk management strategy” (<a href="https://www.consensusdocs.org/resources/price-escalation-clause/">ConsensusDocs</a>). A bid priced in one market and bought out in another carries an assumption that has already expired.</p><h3>Why these failures are silent</h3><p>None of this shows up as a single dramatic event. A lead time slips a few weeks at a time. A substitute gets approved in one submittal while a dependent trade keeps working from the original. A price increase gets absorbed into contingency without anyone asking which other commitments were priced on the same assumption.</p><p>Each change is visible to someone. What’s missing is the connection between the changed fact and every decision that depended on it.</p><p>When I worked in operations, estimating and subcontractor coordination, the costly problems rarely came from missing information. They came from scattered information. The updated lead time existed. The new price existed. It just wasn’t connected to the commitment built on the old one.</p><p>I see the same pattern in enterprise AI. AI’s real problem isn’t intelligence. It’s reliability, because it forgets why decisions were made. A procurement record that captures what was bought, but not why or under what assumptions, has the same flaw. It can tell you the answer. It can’t tell you when the answer stopped being true.</p><h3>Treat assumptions as expiring data</h3><p>A few practices any team can adopt without new software:</p><ol><li><strong>Write the assumptions down with the decision.</strong> For every major package or long-lead item, record the price basis, the quoted lead time, the date of the quote and the install date it supports.</li><li><strong>Give each assumption a review date.</strong> Quotes expire. Lead times need to be re-confirmed. Put a date on every critical assumption, not just on the contract.</li><li><strong>Record why the original product was selected.</strong> Performance, rating, warranty, coordination, owner preference. That list is what any substitution has to be checked against.</li><li><strong>Map the dependents.</strong> Which trades, submittals and schedule activities rely on this item? When it changes, those are the people who need to know.</li><li><strong>Use the contract tools that already exist.</strong> Escalation clauses tied to an objective index, allowances, alternates and early procurement are established ways to share price risk instead of hoping it never arrives.</li><li><strong>Review commitments, not just costs.</strong> A cost report tells you what you spent. A commitment review asks which earlier decisions no longer rest on true facts.</li></ol><h3>The decision nobody revisits</h3><p>Our industry is very good at tracking what was ordered. It is far less consistent at tracking why, and under what conditions that order made sense. In a year when input prices are rising at double-digit rates and critical electrical equipment is quoted in years rather than weeks, that gap is no longer an administrative detail. It is where margin and schedule quietly leak.</p><p>The most dangerous procurement decision isn’t the bad one. It’s the good one that nobody noticed had expired.</p><h3>Where ILYVEN fits</h3><p>ILYVEN is early. Our vision is a more reliable intelligence layer, starting in construction, that connects information, preserves the reasoning behind decisions, identifies emerging risks and helps organizations understand the consequences of changes before they become costly mistakes. Procurement is exactly the kind of place where that matters. That is the direction we are building toward, not a released capability.</p><h3>Thoughts from the Woman in the Red Dress</h3><blockquote>I’ve never thought of a purchase order as paperwork. It’s a promise built on conditions, and conditions change faster than most project records do. When prices and lead times move the way they are moving right now, the question isn’t only what we bought. It’s whether the reasons we bought it still hold.</blockquote><p><strong>Carla Stone</strong><br>Founder, ILYVEN<br><a href="https://ilyven.co/">ilyven.co</a></p><p><strong>Sources (verified loading Nov 17, 2026):</strong></p><ol><li>Associated General Contractors of America, “Construction Input Costs Rise 13.3% From October 2025 to October 2026 as Price Increases Spread From Fuel &amp; Metals to Other Goods &amp; Deliveries” (Nov 13, 2026): <a href="https://www.agc.org/news/2026/11/13/construction-input-costs-rise-133-october-2025-october-2026-price-increases-spread-fuel-metals-other">https://www.agc.org/news/2026/11/13/construction-input-costs-rise-133-october-2025-october-2026-price-increases-spread-fuel-metals-other</a></li><li>Associated Builders and Contractors, “ABC: October Construction Materials Prices Unchanged, Energy and Lumber Prices Fall” (Nov 13, 2026): <a href="https://www.abc.org/News-Media/News-Releases/abc-october-construction-materials-prices-unchanged-energy-and-lumber-prices-fall">https://www.abc.org/News-Media/News-Releases/abc-october-construction-materials-prices-unchanged-energy-and-lumber-prices-fall</a></li><li>IndustrialSage, “Switchgear Lead Times Are the Data Center’s Real Constraint” (Sept 1, 2026): <a href="https://www.industrialsage.com/switchgear-lead-times-2026/">https://www.industrialsage.com/switchgear-lead-times-2026/</a></li><li>ConsensusDocs, “Tariffs and Price Escalation Resource Center”: <a href="https://www.consensusdocs.org/resources/price-escalation-clause/">https://www.consensusdocs.org/resources/price-escalation-clause/</a></li><li>ILYVEN: <a href="https://ilyven.co/">https://ilyven.co</a></li></ol><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=5e061967069d" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The Enterprise Hallucination Budget]]></title>
            <link>https://medium.com/@carla_1448/the-enterprise-hallucination-budget-10a55fa62fa0?source=rss-cf24ffb761af------2</link>
            <guid isPermaLink="false">https://medium.com/p/10a55fa62fa0</guid>
            <category><![CDATA[enterprise-technology]]></category>
            <category><![CDATA[ai-reliability]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[generative-ai-tools]]></category>
            <category><![CDATA[risk-management]]></category>
            <dc:creator><![CDATA[Carla Stone]]></dc:creator>
            <pubDate>Mon, 16 Nov 2026 13:20:00 GMT</pubDate>
            <atom:updated>2026-11-16T13:20:32.042Z</atom:updated>
            <content:encoded><![CDATA[<p>Every company using AI is already spending an error budget. The real question is whether anyone decided where it should be zero.</p><p><em>By Carla Stone, Founder of ILYVEN</em></p><p>Last week, Paligo published a survey of 411 business leaders across the UK, US, Germany and Benelux. Seven in ten said an enterprise AI application in their organization had produced a materially wrong answer or action at least monthly over the past year. Only 15.6% said their organization measures the cost of reaching a reliable, verified answer (<a href="https://www.globenewswire.com/news-release/2026/11/12/3401317/0/en/ai-errors-hitting-business-finances-and-delaying-projects-globally.html">Paligo via GlobeNewswire</a>).</p><p>Read those two numbers together. Wrong answers are routine. Accounting for them is rare.</p><p>That is a budget nobody approved.</p><h3>Every system has an error budget. Most are unwritten.</h3><p>Teams that run large software systems have long used the idea of an error budget: decide in advance how much unreliability a service can tolerate, then spend that allowance deliberately instead of pretending it is zero (<a href="https://sre.google/sre-book/embracing-risk/">Google SRE</a>).</p><p>Generative AI needs the same discipline, because “how often is it wrong?” has no single answer. NIST’s Generative AI Profile calls the problem confabulation: “the production of confidently stated but erroneous or false content” by which users may be misled or deceived (<a href="https://nvlpubs.nist.gov/nistpubs/ai/NIST.AI.600-1.pdf">NIST AI 600–1</a>).</p><p>The word that matters there is <em>confidently</em>. A wrong answer that sounds unsure gets checked. A wrong answer that sounds certain gets used.</p><h3>Tolerance belongs to the decision, not the model</h3><p>EY’s paper on managing hallucination risk inside its own organization makes this concrete. It suggests that for audit, a “good” control regime might allow fewer than one unsupported claim per 1,000 generated outputs, with at least 98% of responses carrying clear source information. Tax could accept up to five per 1,000, provided flagged outputs are escalated to human review within 12 hours. Consulting, where outputs are advisory and interpreted by experts, might tolerate up to ten (<a href="https://www.ey.com/content/dam/ey-unified-site/ey-com/en-gl/technical/documents/ey-gl-managing-hallucination-risk-in-llm-deployments-01-26.pdf">EY</a>).</p><p>You don’t have to adopt those numbers to adopt the logic. The same model can be acceptable in one workflow and unacceptable in another. Tolerance is a property of the decision, not the tool.</p><p>A practical way to draw the lines:</p><ul><li><strong>Generous budget:</strong> drafting, brainstorming, and summaries for a reader who will check the original. Errors are cheap and visible.</li><li><strong>Tight budget:</strong> answers that inform a person’s decision, such as a policy lookup, a specification question or a schedule summary. Every answer should show its source.</li><li><strong>Zero budget:</strong> outputs that become commitments, such as a price, a contract term, a safety requirement, an approval, or a record someone else will rely on without re-checking.</li></ul><p>Zero doesn’t mean the model never errs. It means an unsupported output cannot pass through to action. The system has to cite, abstain, or hand the question to a person.</p><h3>Where the budget actually leaks</h3><p>Here is what most hallucination conversations miss: many costly AI errors are not invented from nothing. They are assembled from real information that has stopped being true.</p><p>In the Paligo survey, 27% of respondents said investigations had identified outdated retrieved information, and 25% found information from multiple sources combined incorrectly. For applications in or near production, only 22% said their organization can always identify which version of a source was used. Among respondents who had encountered errors, 51% said a typical issue takes at least five employee hours to investigate, verify and correct.</p><p>EY’s own taxonomy names the same failure: “out-of-date references,” where a system uses superseded versions of rules or standards, leading to non-compliance and rework.</p><p>Agents raise the stakes because they act instead of just answering. In a Nylas survey of 1,030 U.S. professionals published October 20, 54% said an AI agent had made a mistake in the past 90 days, most often by triggering a workflow the team had to reverse. Only 53% said they could produce a full audit log of what their agents did the previous week (<a href="https://www.nylas.com/blog/pr/ai-agent-mistakes-survey-state-of-agentic-ai-h2-2026/">Nylas</a>).</p><p>This is why I keep coming back to the same point. AI’s real problem isn’t intelligence. It’s reliability, because it forgets why decisions were made. An answer that was right when it was given can become wrong without anyone touching it, because the facts underneath it moved and nothing connected them back.</p><h3>A construction stress test</h3><p>Construction is a useful stress test because so many answers expire.</p><p>Consider a typical scenario. A superintendent asks an assistant which finish was approved for the third floor. The answer cites a real submittal. But a later change superseded it, and the assistant has no way of knowing the reason behind the original approval no longer holds. Nothing was hallucinated in the classic sense. The answer is still wrong, and it arrives in a confident voice.</p><p>When I worked in operations and estimating, the costly problems rarely came from missing information. They came from scattered information: the right fact existed somewhere, disconnected from the decision that depended on it.</p><h3>How to write your hallucination budget</h3><p>Any team can start this week:</p><ol><li><strong>List the decisions AI touches,</strong> not just the tools you’ve bought.</li><li><strong>Assign each decision a budget:</strong> generous, tight or zero.</li><li><strong>For tight and zero budgets, require sources on every answer,</strong> including the version and date of the source.</li><li><strong>Give the system permission to say “I don’t know.”</strong> Abstention is a feature, not a failure.</li><li><strong>Measure the cost of verification,</strong> not just the cost per request. If you don’t count the five hours of checking, you will underprice every error.</li><li><strong>Keep the reason with the decision,</strong> so when a fact changes, you can see which answers and approvals depended on it.</li></ol><h3>The budget you can’t see is the one you overspend</h3><p>Most enterprises already have a hallucination budget. It was set by default, by whichever team adopted a tool first, and it is paid in rework and quiet re-checking that nobody tracks.</p><p>The organizations that get real value from AI won’t be the ones whose models never err. They will be the ones that decided in advance where error is affordable and where it is not, and built systems that remember why.</p><h3>Where ILYVEN fits</h3><p>ILYVEN is early. Our vision is a more reliable intelligence layer, starting in construction, that connects information, preserves the reasoning behind decisions, identifies emerging risks and helps organizations understand the consequences of changes before they become costly mistakes. That is the direction we are building toward, not a released capability.</p><h3>Thoughts from the Woman in the Red Dress</h3><blockquote>I don’t need AI to be perfect. I need it to know when it isn’t sure, and to show me what it relied on. The answers that worry me aren’t the strange ones; they’re the confident ones built on facts that quietly changed. Decide where your tolerance is zero before a tool decides it for you.</blockquote><p><strong>Carla Stone</strong><br> Founder, ILYVEN<br> <a href="https://ilyven.co/">ilyven.co</a></p><p><strong>Sources (verified loading Nov 16, 2026):</strong></p><ol><li>Paligo via GlobeNewswire, “AI errors hitting business finances and delaying projects globally” (Nov 12, 2026): <a href="https://www.globenewswire.com/news-release/2026/11/12/3401317/0/en/ai-errors-hitting-business-finances-and-delaying-projects-globally.html">https://www.globenewswire.com/news-release/2026/11/12/3401317/0/en/ai-errors-hitting-business-finances-and-delaying-projects-globally.html</a></li><li>Google, <em>Site Reliability Engineering</em>, Chapter 3, “Embracing Risk”: <a href="https://sre.google/sre-book/embracing-risk/">https://sre.google/sre-book/embracing-risk/</a></li><li>NIST, AI 600–1, “Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile” (July 2024): <a href="https://nvlpubs.nist.gov/nistpubs/ai/NIST.AI.600-1.pdf">https://nvlpubs.nist.gov/nistpubs/ai/NIST.AI.600-1.pdf</a></li><li>EY, “Managing hallucination risk in LLM deployments at the EY organization” (Jan 2026): <a href="https://www.ey.com/content/dam/ey-unified-site/ey-com/en-gl/technical/documents/ey-gl-managing-hallucination-risk-in-llm-deployments-01-26.pdf">https://www.ey.com/content/dam/ey-unified-site/ey-com/en-gl/technical/documents/ey-gl-managing-hallucination-risk-in-llm-deployments-01-26.pdf</a></li><li>Nylas, “54% of Companies Report an AI Agent Error in the Past 90 Days, Nylas Survey Finds” (Oct 20, 2026): <a href="https://www.nylas.com/blog/pr/ai-agent-mistakes-survey-state-of-agentic-ai-h2-2026/">https://www.nylas.com/blog/pr/ai-agent-mistakes-survey-state-of-agentic-ai-h2-2026/</a></li><li>ILYVEN: <a href="https://ilyven.co/">https://ilyven.co</a></li></ol><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=10a55fa62fa0" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Organizations That Remember Why]]></title>
            <link>https://medium.com/@carla_1448/organizations-that-remember-why-2792ccbb36f0?source=rss-cf24ffb761af------2</link>
            <guid isPermaLink="false">https://medium.com/p/2792ccbb36f0</guid>
            <category><![CDATA[decision-making]]></category>
            <category><![CDATA[leadership]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[knowledge-management]]></category>
            <category><![CDATA[enterprise-technology]]></category>
            <dc:creator><![CDATA[Carla Stone]]></dc:creator>
            <pubDate>Sun, 15 Nov 2026 13:07:51 GMT</pubDate>
            <atom:updated>2026-11-15T13:07:51.685Z</atom:updated>
            <content:encoded><![CDATA[<h4>AI agents are finally getting memory. The bigger opportunity is institutions that keep the reasons behind their choices, and can tell when those reasons have expired.</h4><p><em>By Carla Stone, Founder of ILYVEN</em></p><p>Earlier this month, Salesforce AI Research described a problem every operator will recognize. In March, a customer tells an AI agent their office is in Bangalore. In June, they mention it has moved to Pune. In July, the agent answers as though nothing changed. As the researchers put it, “a reply built on an older fact reads exactly like a reply built on a current one” (<a href="https://www.salesforce.com/blog/agentforce-memory-update/">Salesforce</a>).</p><p>Their fix was a policy change: treat an update as a correction to something the agent already knows, not as a new fact filed next to the old one. On their evaluation, the updated policy recalled 1.8 times as many stated facts and 2.2 times as many facts that change over time.</p><p>That is real progress, and it points at something bigger than software. Machines are learning to remember. Most organizations still haven’t.</p><h3>What organizations forget</h3><p>Companies are good at keeping files. They are much worse at keeping understanding.</p><p>In a Panopto survey of 1,001 U.S. employees, people reported spending an average of 5.3 hours a week waiting for information or guidance that only an in-house expert could provide (<a href="https://www.panopto.com/blog/how-much-time-is-lost-to-knowledge-sharing-inefficiencies-at-work/">Panopto</a>). In Newforma’s 2025 survey of architecture, engineering and construction firms, 40% said they worry about losing institutional knowledge as veteran employees retire (<a href="https://www.newforma.com/labor-shortages-in-aec-the-hidden-impact-and-how-smart-firms-are-adapting/">Newforma</a>).</p><p>I spent more than a decade in construction, in operations, estimating and subcontractor coordination. The costly problems I saw rarely came from missing information. They came from information that was scattered and disconnected, and most of all from decisions whose reasons lived only in someone’s head. When that person moved to another job, or simply forgot, the decision stayed in place and the reasoning disappeared.</p><p>That is the version of forgetting that hurts. Not losing the document, but losing the why.</p><h3>Remembering what versus remembering why</h3><p>Software engineering already has a habit worth borrowing. AWS describes architectural decision records as documents that capture “the decision, the context, and considerations that led to the decision.” When a team changes course, the new record supersedes the old one instead of erasing it (<a href="https://docs.aws.amazon.com/prescriptive-guidance/latest/architectural-decision-records/adr-process.html">AWS</a>). AWS also names the cost of skipping the why: when people don’t understand why a decision was made, it “results in the same topic being discussed multiple times” (<a href="https://docs.aws.amazon.com/prescriptive-guidance/latest/architectural-decision-records/introduction.html">AWS</a>).</p><p>Every organization knows that meeting. The question comes back, nobody remembers how it was settled, and the team spends an hour rebuilding an argument it already had.</p><p><strong>AI’s real problem isn’t intelligence. It’s reliability: it forgets why decisions were made.</strong> The same is true of companies. An organization that remembers what it decided but not why is one personnel change away from repeating its mistakes, or undoing its best calls.</p><h3>Memory you can trust</h3><p>As AI systems start to carry memory for us, a new discipline is forming around it. In a November 11 post on agentic memory, the Cloud Security Alliance argued that “a memory write is a state change,” and that each entry should carry its source, its creation time and its approval state. One line from that piece belongs on every leadership team’s wall: “Relevance is not authority” (<a href="https://cloudsecurityalliance.org/blog/2026/11/11/agentic-memory-and-the-next-security-boundary">Cloud Security Alliance</a>).</p><p>A remembered fact that is close to the question is not the same as a fact that is true, current and approved. That applies to an agent’s memory, and it applies to the stories organizations tell themselves about why things are the way they are.</p><h3>Memory that knows when to let go</h3><p>There is a fair objection here. Institutional memory can also be a trap. Harvard Business Review once ran a piece titled “When Organizational Memory Stands in the Way,” arguing that accumulated memory can hold a company back (<a href="https://hbr.org/2013/03/when-organizational-memory-sta">HBR</a>).</p><p>I agree, and I think it makes the case for keeping the why. “We’ve always done it this way” is what memory sounds like when the reason has been lost. If you keep the reason, you can test it. If the conditions that justified a decision have changed, you can see that, retire the decision on purpose and move on. Remembering why is what makes it safe to forget.</p><h3>What a remembering organization looks like</h3><p>This is a vision, not a description of any product on the market today. But I think the organizations that lead the next decade will share a few traits:</p><ol><li><strong>Decisions travel with their reasons.</strong> The approval, the inputs and the trade-offs are kept together, so anyone can reconstruct a choice without detective work.</li><li><strong>Every important assumption has an owner.</strong> When a fact underneath a decision changes, someone is responsible for asking whether the decision still holds.</li><li><strong>Corrections replace, they don’t pile up.</strong> Like Salesforce’s memory policy, an update supersedes the old fact visibly instead of sitting beside it.</li><li><strong>Memory has provenance.</strong> People and AI systems can see where a remembered fact came from and whether it was ever approved.</li><li><strong>Old decisions expire deliberately.</strong> Retiring a choice is a decision too, and it gets recorded like one.</li></ol><h3>What leaders can do now</h3><p>You do not need new software to start. Pick one recurring decision your team keeps relitigating. Write down what was decided, why, what it depended on and who owns it. Review it when one of those dependencies changes. Then ask the same of any AI tool you bring in: when it remembers something, can it show you where that memory came from and whether it is still current?</p><h3>Where ILYVEN fits</h3><p>ILYVEN is early. Our vision is a more reliable intelligence layer, starting in construction, that connects information, preserves the reasoning behind decisions, identifies emerging risks and helps organizations understand the consequences of changes before they become costly mistakes. That is the direction we are building toward, not a released capability.</p><p>Intelligence will keep getting cheaper. Memory of why will become the scarce thing. The organizations that keep it will make fewer of the same mistakes twice.</p><h3>Thoughts from the Woman in the Red Dress</h3><blockquote>I have watched good decisions get undone for no better reason than that nobody remembered why they were made. That is not a technology failure first; it is a leadership one. I want to help build organizations, and AI, that can look back at a choice and still explain it.</blockquote><p><strong>Carla Stone</strong></p><p>Founder, ILYVEN</p><p><a href="https://ilyven.co/">ilyven.co</a></p><p><strong>Sources (verified loading Nov 15, 2026):</strong></p><ol><li>Salesforce, Itai Asseo and Egor Pakhomov, “Double recall: Inside the Agentforce Memory update” (Nov 3, 2026): <a href="https://www.salesforce.com/blog/agentforce-memory-update/">https://www.salesforce.com/blog/agentforce-memory-update/</a></li><li>Panopto, “How much time is lost to knowledge sharing inefficiencies at work?” (Oct 16, 2019): <a href="https://www.panopto.com/blog/how-much-time-is-lost-to-knowledge-sharing-inefficiencies-at-work/">https://www.panopto.com/blog/how-much-time-is-lost-to-knowledge-sharing-inefficiencies-at-work/</a></li><li>Newforma, “Labor Shortages in AEC: The Hidden Impact and How Smart Firms Are Adapting”: <a href="https://www.newforma.com/labor-shortages-in-aec-the-hidden-impact-and-how-smart-firms-are-adapting/">https://www.newforma.com/labor-shortages-in-aec-the-hidden-impact-and-how-smart-firms-are-adapting/</a></li><li>AWS Prescriptive Guidance, “Architectural decision record process”: <a href="https://docs.aws.amazon.com/prescriptive-guidance/latest/architectural-decision-records/adr-process.html">https://docs.aws.amazon.com/prescriptive-guidance/latest/architectural-decision-records/adr-process.html</a></li><li>AWS Prescriptive Guidance, “Using architectural decision records to streamline decision-making during development”: <a href="https://docs.aws.amazon.com/prescriptive-guidance/latest/architectural-decision-records/introduction.html">https://docs.aws.amazon.com/prescriptive-guidance/latest/architectural-decision-records/introduction.html</a></li><li>Cloud Security Alliance, Chandra Inguva, “Agentic Memory and the Next Security Boundary” (Nov 11, 2026): <a href="https://cloudsecurityalliance.org/blog/2026/11/11/agentic-memory-and-the-next-security-boundary">https://cloudsecurityalliance.org/blog/2026/11/11/agentic-memory-and-the-next-security-boundary</a></li><li>Harvard Business Review, Vijay Govindarajan and Srikanth Srinivas, “When Organizational Memory Stands in the Way” (Mar 11, 2013): <a href="https://hbr.org/2013/03/when-organizational-memory-sta">https://hbr.org/2013/03/when-organizational-memory-sta</a></li><li>ILYVEN: <a href="https://ilyven.co/">https://ilyven.co</a></li></ol><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=2792ccbb36f0" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Three Frameworks, One Question: Can You Show Why?]]></title>
            <link>https://medium.com/@carla_1448/three-frameworks-one-question-can-you-show-why-ce81573606eb?source=rss-cf24ffb761af------2</link>
            <guid isPermaLink="false">https://medium.com/p/ce81573606eb</guid>
            <category><![CDATA[decision-intelligence]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[ai-governance]]></category>
            <category><![CDATA[risk-management]]></category>
            <category><![CDATA[eu-ai-act]]></category>
            <dc:creator><![CDATA[Carla Stone]]></dc:creator>
            <pubDate>Sat, 14 Nov 2026 13:31:10 GMT</pubDate>
            <atom:updated>2026-11-14T13:31:10.504Z</atom:updated>
            <content:encoded><![CDATA[<p>NIST’s AI RMF, the EU AI Act and ISO/IEC 42001 look different on paper. For operators, they are converging on the same demand: prove what your AI did, and why it was allowed to.</p><p><em>By Carla Stone, Founder of ILYVEN</em></p><h4>The three pillars</h4><p>NIST AI RMF: the shared vocabulary</p><p>The NIST AI Risk Management Framework was released on January 26, 2023, and is “intended for voluntary use” (NIST). It organizes AI risk work into four functions: Govern, Map, Measure and Manage. A Generative AI Profile followed in July 2024, and in April 2026 NIST released a concept note for a profile on trustworthy AI in critical infrastructure. The framework itself is still version 1.0, and NIST says “it is being revised as part of the White House AI Action Plan.”</p><p>For operators, the AI RMF is less a rulebook than a common language. When a risk team, an auditor and a buyer need to talk about AI without talking past each other, this is usually the vocabulary they reach for. It certifies nothing.</p><p>The EU AI Act: the law, on a new clock</p><p>The EU AI Act is binding, and it applies in stages. Prohibitions and AI literacy duties began on February 2, 2025. Obligations for general-purpose AI models followed on August 2, 2025. Transparency requirements and enforcement started on August 2, 2026. After the Digital Omnibus on AI, the obligations for stand-alone high-risk systems listed in Annex III now apply from December 2, 2027, and those for high-risk AI embedded in regulated products from August 2, 2028 (AI Act Service Desk).</p><p>A later deadline is not a softer requirement. Article 12 still says high-risk AI systems “shall technically allow for the automatic recording of events (logs) over the lifetime of the system” (Article 12). The extra time is preparation time, and record-keeping is the kind of thing that takes longest to retrofit.</p><p>ISO/IEC 42001: the certificate</p><p>ISO/IEC 42001, published in 2023, describes itself as “the world’s first AI management system standard,” and it covers organizations that build AI and those that simply use it (ISO). Unlike the other two, it can be audited and certified.</p><p>That is where the market momentum is. BSI announced its 50th ISO/IEC 42001 certificate on October 21, 2026, less than two years after it issued its first (BSI). The ISO42k directory, which counts only publicly announced certifications, lists 240 organizations across 30 countries, with 129 certified in 2026 so far. Enterprise software companies make up 45% of the directory, and the site itself notes that vendors selling to enterprises have a commercial reason to publicize (ISO42k).</p><h4>What the market is actually buying</h4><p>Put those signals together and a pattern shows up.</p><p>Governance is moving from policy to evidence. A published AI policy used to be enough to answer a questionnaire. Now buyers, auditors and regulators ask to see records.</p><p>Certification is becoming a procurement artifact. The share of software vendors among public 42001 announcements suggests many companies certify because their customers ask, not because a law requires it.</p><p>Timelines move. Direction does not. The EU pushed dates back, but it kept the logging, oversight and documentation duties. New York added incident reporting. The trend is toward more proof, not less.</p><h4>The question underneath all three</h4><p>Strip away the acronyms and every one of these frameworks asks a version of the same thing: can you show what your AI did, and why it was allowed to?</p><p>Logs capture events. Management systems capture policies. Risk frameworks capture intentions. Each is useful. None of them, on its own, keeps the reasoning attached to a decision once the facts underneath it start to change.</p><p>That is the gap I keep coming back to. AI’s real problem isn’t intelligence. It’s reliability: it forgets why decisions were made. In a sense, governance frameworks are society’s attempt to force memory onto systems that do not have it.</p><h4>What operators can do now</h4><ol><li>Pick a vocabulary. Use NIST’s Govern, Map, Measure and Manage to describe where AI touches your work, even if no law requires it yet.</li><li>Inventory decisions, not just models. List the places AI informs or makes a decision, and name a person who owns each one.</li><li>Keep the “why” with the “what.” Store the inputs, the approval and the reasoning together, so a decision can be reconstructed later without detective work.</li><li>Check your calendar. If you sell AI-enabled products or services into Europe, December 2, 2027 and August 2, 2028 are planning dates, not distant ones.</li><li>Ask vendors for evidence, not badges. A certificate shows a management system exists. Ask how a specific decision made with their system could be explained after the fact.</li></ol><h4>Where ILYVEN fits</h4><p>ILYVEN is early. Our vision is a more reliable intelligence layer, starting in construction, that connects information, preserves the reasoning behind decisions, identifies emerging risks and helps organizations understand the consequences of changes before they become costly mistakes. That is the direction we are building toward, not a released capability, and none of these frameworks endorses any product, ours included.</p><p>The rulebooks will keep changing. The question they share will not. Organizations that can show why will have an easier time with every framework that comes next.</p><p>— — —</p><h4>Thoughts from the Woman in the Red Dress</h4><p><em>Every framework in this piece is trying to answer a question construction has always struggled with: why did we do it that way? Regulation can require a record. It cannot make anyone keep the reasoning alive, and that part is on the people building these systems. I intend to be one of them.</em></p><p>Carla Stone</p><p>Founder, ILYVEN</p><p>ilyven.co</p><p>— — —</p><p>Sources (verified loading Nov 14, 2026):</p><ol><li>TechCrunch, Rebecca Bellan, “OpenAI, Anthropic face new oversight as New York’s AI safety law takes effect” (Nov 10, 2026): <a href="https://techcrunch.com/2026/11/10/openai-anthropic-face-new-oversight-as-new-yorks-ai-safety-law-takes-effect/">https://techcrunch.com/2026/11/10/openai-anthropic-face-new-oversight-as-new-yorks-ai-safety-law-takes-effect/</a></li><li>New York State Department of Financial Services, “Information for Frontier Developers” (RAISE Act): <a href="https://www.dfs.ny.gov/digit/Frontier-Developers">https://www.dfs.ny.gov/digit/Frontier-Developers</a></li><li>NIST, “AI Risk Management Framework”: <a href="https://www.nist.gov/itl/ai-risk-management-framework">https://www.nist.gov/itl/ai-risk-management-framework</a></li><li>European Commission AI Act Service Desk, “When does the AI Act go into effect? What is its timeline for implementation?”: <a href="https://ai-act-service-desk.ec.europa.eu/en/ai-act/faq/when-does-ai-act-go-effect-what-its-timeline-implementation">https://ai-act-service-desk.ec.europa.eu/en/ai-act/faq/when-does-ai-act-go-effect-what-its-timeline-implementation</a></li><li>EU Artificial Intelligence Act, “Article 12: Record-Keeping”: <a href="https://artificialintelligenceact.eu/article/12/">https://artificialintelligenceact.eu/article/12/</a></li><li>ISO, “ISO/IEC 42001:2023 — AI management systems”: <a href="https://www.iso.org/standard/42001">https://www.iso.org/standard/42001</a></li><li>BSI, “BSI awards its 50th ISO/IEC 42001 certificate” (Oct 21, 2026): <a href="https://www.bsigroup.com/en-GB/insights-and-media/media-centre/press-releases/2026/november/bsi-awards-its-50th-iso-iec-42001-certificate/">https://www.bsigroup.com/en-GB/insights-and-media/media-centre/press-releases/2026/november/bsi-awards-its-50th-iso-iec-42001-certificate/</a></li><li>ISO42k, “State of ISO 42001 Certification” (accessed Nov 14, 2026): <a href="https://iso42k.com/state-of-iso-42001">https://iso42k.com/state-of-iso-42001</a></li><li>ILYVEN: <a href="https://ilyven.co/">https://ilyven.co</a></li></ol><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=ce81573606eb" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Accountability Is a Job, Not a Name on a Form]]></title>
            <link>https://medium.com/@carla_1448/accountability-is-a-job-not-a-name-on-a-form-ff1978ecaaa3?source=rss-cf24ffb761af------2</link>
            <guid isPermaLink="false">https://medium.com/p/ff1978ecaaa3</guid>
            <category><![CDATA[decision-intelligence]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[construction]]></category>
            <category><![CDATA[leadership]]></category>
            <category><![CDATA[ai-governance]]></category>
            <dc:creator><![CDATA[Carla Stone]]></dc:creator>
            <pubDate>Fri, 13 Nov 2026 13:15:39 GMT</pubDate>
            <atom:updated>2026-11-13T13:15:39.475Z</atom:updated>
            <content:encoded><![CDATA[<h3>What job sites teach about ownership, sign-off and the authority to stop work, and why software keeps forgetting it as AI agents start making real decisions.</h3><p><em>By Carla Stone, Founder of ILYVEN</em></p><p>This week, three very different publications landed on the same uncomfortable idea. On November 9, Dataiku opened an essay on agentic AI with a single line: “A named owner is not the same thing as accountability” (<a href="https://www.dataiku.com/blog/building-authority-chain-agentic-ai">Dataiku</a>). On November 12, the law firm Mayer Brown wrote that with AI agents, “what must be governed is not merely what the AI agent does, but what the AI agent is authorized to do on the company’s behalf” (<a href="https://www.mayerbrown.com/en/insights/publications/2026/11/contracting-to-address--agentic-ai-governance">Mayer Brown</a>). And on November 13, a Forbes Technology Council piece ended with a sentence I would hang in any project trailer: “Human oversight means little if the human cannot change the outcome” (<a href="https://www.forbes.com/councils/forbestechcouncil/2026/11/13/why-human-oversight-of-ai-needs-more-than-a-human-in-the-loop/">Forbes</a>).</p><p>None of that would surprise anyone who has run work in the field. Job sites learned a long time ago what AI governance is learning now: <strong>accountability is a job, not a name on a form.</strong></p><h3>Where I learned this</h3><p>I entered construction over a decade ago and co-founded a residential construction company in 2018 that grew to about $10 million in annual gross sales by 2024. I worked in operations, estimating, subcontractor coordination and workflows. The costly problems I saw rarely came from missing information. They came from scattered, disconnected information, and very often from the gaps between the people who were supposed to own it.</p><p>I don’t think construction has accountability figured out. But the physical world is unforgiving in a useful way. Work you can walk through tends to force questions that software lets you skip. Here are four of them.</p><h3>Lesson one: a name is where accountability starts, not where it ends</h3><p>On a well-run project, scopes are assigned, inspections are signed and someone is responsible for each piece of work. But everyone in the trade knows the difference between the name on the paperwork and the person who actually owns the outcome. Ownership means you know what the work is for, you know what can go wrong, and you answer for it when it does.</p><p>Software tends to collapse all of that into a field. Dataiku describes the gap precisely: “A name on a platform tells you whom to contact; it does not establish what that person owns, what they can decide, or when responsibility passes to someone else.”</p><p><strong>An owner field answers “who do I email?” Accountability answers “who knows why, and who can do something about it?”</strong></p><h3>Lesson two: if you can’t stop the work, you don’t own it</h3><p>Strong safety cultures give people the authority to stop work when something looks wrong. The point is not the rule itself. The point is that the authority is real, the person knows they have it, and nobody gets punished for using it.</p><p>AI oversight is arriving at the same place. The Forbes piece warns that an organization “can assign a reviewer, document an escalation path and install an override control” without answering whether the reviewer can “recognize a consequential error and intervene when it matters.” Its advice reads like a superintendent’s: “Define which decisions people own, when they should challenge an automated output and who can override, escalate or stop a process.”</p><p>A reviewer who can only approve is not providing oversight. They are providing a signature.</p><h3>Lesson three: accountability dies in the handoff</h3><p>On a job site, work moves from trade to trade, crew to crew and phase to phase. Each handoff is a moment where one party’s assumptions become someone else’s starting conditions. When a handoff is clean, the next crew knows what they are inheriting. When it isn’t, problems get buried under the next layer of work and surface later, when they cost more to fix.</p><p>Dataiku makes the same observation about AI systems: “The most serious accountability failures often occur not within one role, but between roles.” Its prescription is one every project manager would recognize: “The goal is not to proliferate names on a document, but to eliminate unowned handoffs.”</p><p>AI agents multiply handoffs. A model provider, a configuration team, a business owner, an approver and an operations team can each shape a single outcome without any one of them seeing it whole.</p><h3>Lesson four: on a job site, you can still ask why</h3><p>Here is the lesson I think software forgets most. On a project, when something looks off, you can usually find the person who made the call and ask why. The reason lives in someone’s head, and that is fragile, but it lives somewhere.</p><p>AI systems are getting very good at recording what happened. Mayer Brown recommends “auditable records and logs that allow for reconstruction of agent actions and verification of whether the agent acted within its authorized action profile.” That is necessary. But a log that proves an action was authorized does not tell you whether the reason behind the authorization still holds. I wrote yesterday about how <a href="https://medium.com/@carla_1448/approvals-are-promises-about-the-future-3b69f7531d16">approvals are promises about the future</a>; accountability is the person, or the system, that keeps that promise honest.</p><p>This is the pattern I keep coming back to. <strong>AI’s real problem isn’t intelligence. It’s reliability, because it forgets why decisions were made.</strong> Reconstructing what an agent did is an audit. Reconstructing why it was allowed to is governance.</p><h3>Four job-site habits worth bringing to AI governance</h3><p>None of these require new technology to start.</p><p><strong>1. Pair every owner with a reason.</strong> Record not just who owns an AI system or decision, but what it is for and what would make it wrong.</p><p><strong>2. Make stop authority real.</strong> Name who can pause or override, confirm they know it and make sure they have the information and time to use it.</p><p><strong>3. Treat handoffs as decisions.</strong> When responsibility moves between teams, vendors or agents, write down what is being handed over, including the assumptions.</p><p><strong>4. Keep the “why” findable.</strong> If the person who made the call is unavailable, someone should still be able to see the reasoning behind it.</p><h3>Demonstrated, in development and vision</h3><p>To keep the lines clear: what is demonstrated is the problem. Practitioners, lawyers and governance teams are now publicly distinguishing named owners from real accountability. What is in active development across the industry is the machinery: authority chains, oversight thresholds, override controls and audit logs. What remains a vision is a reliable way to keep the reasoning behind decisions attached to the people and systems accountable for them, so ownership doesn’t quietly expire at the handoff.</p><h3>Where ILYVEN fits</h3><p>I founded ILYVEN (<a href="https://ilyven.co/">ilyven.co</a>) once AI made a long-held vision feasible: a more reliable intelligence layer, starting in construction, that connects information, preserves the reasoning behind decisions, identifies emerging risks and helps organizations understand the consequences of changes before they become costly mistakes. We are early. This is the direction we are building toward, not a finished product, and nothing here describes a released capability.</p><p>A job site will tell you who signed. The best ones can also tell you why, and who can still stop the work. Software should be able to say the same.</p><h3>Thoughts from the Woman in the Red Dress</h3><blockquote>The people I trusted most on a project were never the ones with the biggest titles. They were the ones who could tell you why something was done and were willing to stop the work when it didn’t add up. That is the standard I want to hold AI to.</blockquote><p><strong>Carla Stone</strong></p><p>Founder, ILYVEN</p><p><a href="https://ilyven.co/">ilyven.co</a></p><p><strong>Sources (verified loading Nov 13, 2026):</strong></p><ol><li>Dataiku Blog, Faye Murray and Jacob Beswick, “Who is actually accountable? Building a real authority chain for agentic AI” (Nov 9, 2026): <a href="https://www.dataiku.com/blog/building-authority-chain-agentic-ai">https://www.dataiku.com/blog/building-authority-chain-agentic-ai</a></li><li>Mayer Brown, Lei Shen, Arsen Kourinian, Anjani D. Nadadur and Evan Han, “Contracting to Address Agentic AI Governance” (Nov 12, 2026): <a href="https://www.mayerbrown.com/en/insights/publications/2026/11/contracting-to-address--agentic-ai-governance">https://www.mayerbrown.com/en/insights/publications/2026/11/contracting-to-address--agentic-ai-governance</a></li><li>Forbes Technology Council, Alma Jones, “Why Human Oversight Of AI Needs More Than A Human In The Loop” (Nov 13, 2026): <a href="https://www.forbes.com/councils/forbestechcouncil/2026/11/13/why-human-oversight-of-ai-needs-more-than-a-human-in-the-loop/">https://www.forbes.com/councils/forbestechcouncil/2026/11/13/why-human-oversight-of-ai-needs-more-than-a-human-in-the-loop/</a></li><li>Carla Stone, “Approvals Are Promises About the Future” (Medium, Nov 12, 2026): <a href="https://medium.com/@carla_1448/approvals-are-promises-about-the-future-3b69f7531d16">https://medium.com/@carla_1448/approvals-are-promises-about-the-future-3b69f7531d16</a></li><li>ILYVEN: <a href="https://ilyven.co/">https://ilyven.co</a></li></ol><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=ff1978ecaaa3" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Approvals Are Promises About the Future]]></title>
            <link>https://medium.com/@carla_1448/approvals-are-promises-about-the-future-3b69f7531d16?source=rss-cf24ffb761af------2</link>
            <guid isPermaLink="false">https://medium.com/p/3b69f7531d16</guid>
            <category><![CDATA[ai-agent]]></category>
            <category><![CDATA[ai-governance]]></category>
            <category><![CDATA[decision-intelligence]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[risk-management]]></category>
            <dc:creator><![CDATA[Carla Stone]]></dc:creator>
            <pubDate>Thu, 12 Nov 2026 13:21:39 GMT</pubDate>
            <atom:updated>2026-11-12T13:21:39.152Z</atom:updated>
            <content:encoded><![CDATA[<h3>Every sign-off assumes the facts behind it will hold. As AI agents start acting on approvals given in advance, the real question is no longer who signed, but whether the reason they signed is still true.</h3><p><em>By Carla Stone, Founder of ILYVEN</em></p><p>In the past ten days, three very different voices said the same thing about approvals. On November 2, the Hong Kong Monetary Authority published a discussion paper on agentic AI governance in banking, built around a principle it calls intent-anchored accountability: “An agent should operate only in pursuit of an approved business intent, within a defined delegation of authority” (<a href="https://www.hkma.gov.hk/media/eng/doc/key-information/press-release/2026/20261102e3a1.pdf">HKMA</a>). On November 4, AWS described how its Smart Approvals feature handles agent actions, noting that each automatic approval “is single-use, applies only to the exact action and parameters reviewed, and grants no ongoing permission” (<a href="https://aws.amazon.com/blogs/machine-learning/smart-approvals-in-amazon-quick-human-oversight-where-it-matters/">AWS</a>). On November 5, Singapore’s Senior Minister of State Tan Kiat How told a government technology conference: “Recommending a payment is different from making one” (<a href="https://world.medianama.com/explained-how-singapore-plans-human-approval-for-higher-risk-government-ai-agents/">MediaNama</a>).</p><p>All three are wrestling with the same quiet truth. An approval is not a statement about the present. It is a promise about the future.</p><p>When someone signs off on a budget, a submittal or an AI agent, they are saying: this will still be right when it is carried out. Usually that happens later, by someone else, under conditions the approver never saw.</p><h3>Agents are making the promise explicit</h3><p>For most of business history, the future part of an approval stayed implicit. AI agents force it into the open, because they act on approvals given in advance.</p><p>The Agent Payments Protocol (AP2) specification describes this plainly. In its “Human Not Present” mode, “The User sees and approves a set of constraints over what closed Checkout and Payment would meet their intent.” The specification even recommends that these advance approvals expire as soon as practical, at “the smallest value that will allow the Shopping Agent to complete the assigned task” (<a href="https://ap2-protocol.org/ap2/specification/">AP2</a>).</p><p>That is what every approval has always been: permission now for an action later, inside limits, in service of an intent. Agents just make the gap impossible to ignore.</p><h3>Every approval has two halves: the bounds and the bet</h3><p>Look closely at any approval and you find two different things.</p><p><strong>The bounds</strong> are what is allowed: an amount, a payee, a scope, a date range, a list of tools. Bounds are concrete, and software is getting very good at enforcing them.</p><p><strong>The bet</strong> is why the approval made sense: the facts the approver assumed would hold. The price was firm. The schedule had room. The data was current. The bet is what the approver was actually promising, and it is almost never written down.</p><p>AWS states the limit of bounds with unusual honesty: “These enforcement boundaries limit what an approval authorizes. They don’t establish that the classifier’s judgment was correct” (<a href="https://aws.amazon.com/blogs/machine-learning/smart-approvals-in-amazon-quick-human-oversight-where-it-matters/">AWS</a>). The same is true of human approvals. An action can sit inside every bound and still break the promise, because the bet underneath it no longer holds.</p><p><strong>Controls check whether an action fits the approval. Nothing checks whether the approval still fits the world.</strong></p><h3>Approvals drift without anyone deciding</h3><p>The HKMA paper names the uncomfortable part. Routine changes can expand an agent’s permissions and toolsets, and approval steps can weaken under volume. “A deployment may also move along this spectrum without a deliberate decision,” the paper says. “Its current position therefore may not reflect where it will be in a year” (<a href="https://www.hkma.gov.hk/media/eng/doc/key-information/press-release/2026/20261102e3a1.pdf">HKMA</a>).</p><p>Its answer is to tie authority back to intent. When models, tools, data sources or permissions change, institutions should assess “whether a proposed change affects the approved business intent, delegated authority, risk assessment or materiality of the deployment, and whether re-approval, reclassification or revalidation is required.” The paper also refuses to treat a signature as proof of oversight: “A formal approval step is not, by itself, meaningful human oversight.”</p><p>That is a regulator writing about banks. It is also one of the clearest descriptions of decision intelligence I have read this month. An approval is only as good as the intent it was anchored to, and intent is the bet, written down.</p><h3>Who keeps the promise after the signature?</h3><p>Last week I wrote about <a href="https://medium.com/@carla_1448/the-half-life-of-a-decision-84bacd3ca7d7">the half-life of a decision</a>, how the evidence under a decision ages. This is the other side of that problem. A promise has two parties. The approver makes it once. Everyone downstream relies on it, every day, for as long as the work continues.</p><p>I entered construction over a decade ago and co-founded a residential construction company in 2018 that reached about $10 million in annual gross sales by 2024. My work spanned operations, estimating, subcontractor coordination and workflows. The costly problems I saw rarely came from missing information. They came from scattered, disconnected information.</p><p>An approval on a project is rarely the end of a conversation. It is the start of everyone else’s work. Purchasing orders against it. Crews schedule around it. Budgets assume it. When the facts behind it move, the people relying on it are often the last to find out, because the reason lived with whoever signed, not with the approval.</p><p>That is the pattern I keep coming back to. <strong>AI’s real problem isn’t intelligence. It’s reliability, because it forgets why decisions were made.</strong> An agent acting on an old approval is not disobeying anyone. It is keeping a promise whose reasons have quietly expired.</p><h3>Five ways to write approvals as promises</h3><p>None of this requires new technology to start.</p><p><strong>1. Write the “because” next to the stamp.</strong> “Approved” says what. “Approved because the quoted price holds through delivery” says what the approver is promising.</p><p><strong>2. Separate the bounds from the bet.</strong> Record the limits that can be enforced and the assumptions that can’t. Most risk intelligence lives in the second list.</p><p><strong>3. Name who is relying on it.</strong> A promise without a known audience can’t be renegotiated when it breaks.</p><p><strong>4. Decide in advance what breaks the promise.</strong> Not just a date, but the conditions: a price change, a new model version, a revised drawing, a wider data scope.</p><p><strong>5. Give the promise a keeper.</strong> Accountability should not end at the signature. Someone should own the question “is this still true?” for as long as others are acting on it.</p><h3>Demonstrated, in development and vision</h3><p>To keep the lines clear: what is demonstrated is the problem. Regulators and platform providers now publicly separate authority from judgment and call for re-approval when conditions change. What is in active development across the industry is the machinery of bounds: scoped permissions, expiring mandates and audit logs. What remains a vision is a reliable way to keep the bet attached to the approval, so the people and agents relying on it learn when a fact underneath it changes.</p><h3>Where ILYVEN fits</h3><p>I founded ILYVEN (<a href="https://ilyven.co/">ilyven.co</a>) once AI made a long-held vision feasible: a more reliable intelligence layer, starting in construction, that connects information, preserves the reasoning behind decisions, identifies emerging risks and helps organizations understand the consequences of changes before they become costly mistakes. We are early. This is the direction we are building toward, not a finished product, and nothing here describes a released capability.</p><p>A signature tells you someone agreed. Only the reason tells you whether they would still agree today.</p><h3>Thoughts from the Woman in the Red Dress</h3><blockquote>I used to think of a signature as the end of a decision. Running projects taught me it was closer to a promise, and the people downstream were the ones who felt it when the promise broke. If AI agents are going to act on our approvals, I want them to carry our reasons along with our permission.</blockquote><p><strong>Carla Stone</strong></p><p>Founder, ILYVEN</p><p><a href="https://ilyven.co/">ilyven.co</a></p><p><strong>Sources (verified loading Nov 12, 2026):</strong></p><ol><li>Hong Kong Monetary Authority, “Discussion Paper on Agentic A.I. Governance in the Banking Sector” (November 2026; published Nov 2, 2026): <a href="https://www.hkma.gov.hk/media/eng/doc/key-information/press-release/2026/20261102e3a1.pdf">https://www.hkma.gov.hk/media/eng/doc/key-information/press-release/2026/20261102e3a1.pdf</a></li><li>AWS Machine Learning Blog, Steve Vinsik, “Smart Approvals in AWS Quick: Human oversight where it matters” (Nov 4, 2026): <a href="https://aws.amazon.com/blogs/machine-learning/smart-approvals-in-amazon-quick-human-oversight-where-it-matters/">https://aws.amazon.com/blogs/machine-learning/smart-approvals-in-amazon-quick-human-oversight-where-it-matters/</a></li><li>MediaNama, Nikhil Pahwa, “Explained: How Singapore plans human approval for higher-risk government AI agents” (Nov 8, 2026; remarks delivered Nov 5, 2026 at GovTech STACK Conference 2026): <a href="https://world.medianama.com/explained-how-singapore-plans-human-approval-for-higher-risk-government-ai-agents/">https://world.medianama.com/explained-how-singapore-plans-human-approval-for-higher-risk-government-ai-agents/</a></li><li>Agent Payments Protocol (AP2) Documentation, “Agent Payments Protocol” specification: <a href="https://ap2-protocol.org/ap2/specification/">https://ap2-protocol.org/ap2/specification/</a></li><li>Carla Stone, “The Half-Life of a Decision” (Medium, Nov 5, 2026): <a href="https://medium.com/@carla_1448/the-half-life-of-a-decision-84bacd3ca7d7">https://medium.com/@carla_1448/the-half-life-of-a-decision-84bacd3ca7d7</a></li><li>ILYVEN: <a href="https://ilyven.co/">https://ilyven.co</a></li></ol><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=3b69f7531d16" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The Missing Layer Between Your Software and Your Decisions]]></title>
            <link>https://medium.com/@carla_1448/the-missing-layer-between-your-software-and-your-decisions-6cb65250cff0?source=rss-cf24ffb761af------2</link>
            <guid isPermaLink="false">https://medium.com/p/6cb65250cff0</guid>
            <category><![CDATA[ai-agent]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[ai-governance]]></category>
            <category><![CDATA[enterprise-technology]]></category>
            <category><![CDATA[decision-making]]></category>
            <dc:creator><![CDATA[Carla Stone]]></dc:creator>
            <pubDate>Wed, 11 Nov 2026 13:26:43 GMT</pubDate>
            <atom:updated>2026-11-11T13:26:43.901Z</atom:updated>
            <content:encoded><![CDATA[<h4>Enterprise AI is racing to give agents shared context. The layer that matters most connects tools, people and AI around the decisions themselves, and remembers why each one was made.</h4><p><em>By Carla Stone, Founder of ILYVEN</em></p><p>In the past week, enterprise AI kept circling the same missing piece. At DataHub’s CONTEXT 2026 summit on November 4, Optiver presented an evaluation in which an agent found documentation describing a known issue that had been fixed three days earlier, followed it correctly “and produced exactly the wrong answer” (<a href="https://datahub.com/blog/context-2026-highlights/">DataHub</a>). On November 10, IBM and Microsoft announced Imagine Enterprise, an agentic platform built in part to “connect enterprise knowledge, data, and context across the organization” so that “people and agents can operate with richer enterprise knowledge and context behind them” (<a href="https://newsroom.ibm.com/2026-11-10-ibm-and-microsoft-announce-a-new-ai-agent-platform-designed-for-industry-specific-transformation">IBM Newsroom</a>).</p><p>Different companies, different products, same diagnosis. Between the software an organization runs and the decisions it makes, something is missing.</p><p>I want to be precise about what that something is, because I think the industry is close to the right answer and at risk of stopping one layer short.</p><h3>The connection problem is real and measurable</h3><p>Salesforce’s 2026 Connectivity Benchmark Report, a survey of 1,050 enterprise IT leaders, found the average number of applications grew from 897 to 957 year over year, with only 27% of them integrated together. Half of AI agents “operate in isolated silos,” and 86% of IT leaders are concerned that agents will introduce more complexity than value without proper integration (<a href="https://www.salesforce.com/news/stories/connectivity-report-announcement-2026/">Salesforce</a>).</p><p>Connection alone does not close the gap. In a Teradata survey of 1,000 senior technology and data leaders, 77% said 20% or less of their enterprise data is sufficiently described and contextualized for agents to use. Leaders ranked data lacking “the necessary metadata, context, and relationships” (43%) nearly level with data fragmented across systems (42%) (<a href="https://www.teradata.com/insights/ai-and-machine-learning/disconnected-data-stifling-roi-from-agentic-ai">Teradata</a>).</p><p>The trust consequences follow. In a global survey of 300 data and technology executives by MIT Technology Review Insights with Google Cloud, 51% said they trust the accuracy and relevance of their AI agents’ decisions. Among “data laggards,” organizations giving AI access to 30% or less of their data, it was 22% (<a href="https://cloud.google.com/transform/scaling-ai-agents-trustworthy-data-foundation-mit-review-insights-report">Google Cloud</a>).</p><p>These are vendor-sponsored surveys, so I read them as a direction, not a measurement. The direction is consistent: the tools exist, the data exists, and the connective tissue between them is thin.</p><h3>Context is not the same as reasoning</h3><p>Much of the context work on display this month centers on one question: what does this data mean? Which metric definition is authoritative, which table is current, who owns it. That is necessary work, and the teams doing it are reporting real progress.</p><p>But a decision is a different object than a dataset. A decision has an owner, a moment in time, a set of facts it relied on, alternatives that were rejected and conditions that would make it wrong. Very little of that lives inside any single application. It lives in a meeting, an email thread, a phone call and someone’s judgment.</p><p>The Optiver example shows why that matters. The documentation was not missing. It was outdated, and nothing connected the fix to the document that described the problem. Nothing told the agent that the reason behind the instruction no longer held.</p><p>AI’s real problem isn’t intelligence. It’s reliability, because it forgets why decisions were made.</p><p><strong>Connecting your software tells AI what you have. Connecting your decisions tells it what you meant.</strong></p><h3>What the missing layer has to do</h3><p>If the gap sits between software and decisions, the layer that fills it needs four properties.</p><p><strong>1. It has to persist.</strong> It should outlive the chat session, the project and the employee who made the call. Memory that resets with every conversation is a cache, not institutional memory.</p><p><strong>2. It has to span tools, not belong to one.</strong> With hundreds of applications in a typical enterprise, the reasoning behind a single decision rarely sits in one of them. A layer owned by one app sees one slice.</p><p><strong>3. It has to include people, not just agents.</strong> At the same summit, Google Cloud’s Prajakta Damle described wanting “a layer that can observe your environment, learn from both your data and your systems and the people working on these tasks” (<a href="https://datahub.com/blog/context-2026-highlights/">DataHub</a>). The exceptions, trade-offs and approvals people make are where most of the why actually lives.</p><p><strong>4. It has to understand consequences.</strong> When a fact changes, the layer should surface which decisions rested on it, before someone builds on them.</p><h3>Why this is an AI governance question</h3><p>Much of AI governance today focuses on what agents are allowed to do: permissions, controls, observability. That work is essential. But governance also has to answer questions after the fact. Why was this done? Based on what? Approved by whom? Is that reason still true?</p><p>Without a persistent record that connects decisions to their reasons, governance can show that an agent was authorized. It cannot show whether the agent was right.</p><p><strong>Permission tells you an agent could act. Only the reasoning tells you whether it should have.</strong></p><h3>What construction taught me about the gap</h3><p>I entered construction over a decade ago and later co-founded a residential construction company in 2018 that reached about $10 million in annual gross sales by 2024. My work spanned operations, estimating, subcontractor coordination and workflows. The costly problems I saw rarely came from missing information. They came from information that was scattered and disconnected from the decisions that depended on it.</p><p>A construction project is a good model of the missing layer because it runs on many tools and many people at once. Estimates, schedules, drawings, contracts and invoices each live somewhere. The reasons connecting them, why a scope was cut or a sequence changed, usually lived with whoever made the call. When that person moved on, the reason moved with them.</p><h3>Demonstrated, in development and vision</h3><p>To keep the lines clear: what is demonstrated is the problem. Surveys from several directions point to fragmented systems, thin context and limited trust in agent decisions. What is in active development across the industry is shared context for agents, governance controls and standards for connecting tools. What remains a vision, and an open question, is a persistent layer that connects tools, people and AI around decisions and their reasons, and flags the consequences of a change before it becomes costly.</p><h3>Where ILYVEN fits</h3><p>I founded ILYVEN (<a href="https://ilyven.co/">ilyven.co</a>) once AI made a long-held vision feasible: a more reliable intelligence layer, starting in construction, that connects information, preserves the reasoning behind decisions, identifies emerging risks and helps organizations understand the consequences of changes before they become costly mistakes. We are early. This is the direction we are building toward, not a finished product, and nothing here describes a released capability.</p><p>The next shift in enterprise AI will not be another application. It will be the layer between them that remembers why.</p><h3>Thoughts from the Woman in the Red Dress</h3><blockquote><em>I’m encouraged that so many teams are now saying out loud that agents need context. My worry is that we will stop at teaching AI what our data means and never teach it why we decided what we did. That second part is where most of the expensive problems I saw in construction came from, and it’s the part I can’t stop thinking about.</em></blockquote><p><strong>Carla Stone</strong></p><p>Founder, ILYVEN</p><p><a href="https://ilyven.co/">ilyven.co</a></p><p><strong>Sources (verified loading Nov 11, 2026):</strong></p><ol><li>DataHub, “CONTEXT 2026 Highlights: Context Management in Practice” (Nov 5, 2026; summit held Nov 4, 2026): <a href="https://datahub.com/blog/context-2026-highlights/">https://datahub.com/blog/context-2026-highlights/</a></li><li>IBM Newsroom, “IBM and Microsoft Announce a New AI Agent Platform Designed for Industry-Specific Transformation” (Nov 10, 2026): <a href="https://newsroom.ibm.com/2026-11-10-ibm-and-microsoft-announce-a-new-ai-agent-platform-designed-for-industry-specific-transformation">https://newsroom.ibm.com/2026-11-10-ibm-and-microsoft-announce-a-new-ai-agent-platform-designed-for-industry-specific-transformation</a></li><li>Salesforce, “Multi-Agent Adoption to Surge 67% by 2027 as Enterprises Race Toward Agentic Transformation” (2026 Connectivity Benchmark Report, Feb 5, 2026): <a href="https://www.salesforce.com/news/stories/connectivity-report-announcement-2026/">https://www.salesforce.com/news/stories/connectivity-report-announcement-2026/</a></li><li>Teradata, “Disconnected Data Is Stifling ROI From Agentic AI” (Oct 11, 2026): <a href="https://www.teradata.com/insights/ai-and-machine-learning/disconnected-data-stifling-roi-from-agentic-ai">https://www.teradata.com/insights/ai-and-machine-learning/disconnected-data-stifling-roi-from-agentic-ai</a></li><li>Google Cloud Blog, “Scaling AI agents with trustworthy data: New findings from MIT Technology Review Insights” (Aug 12, 2026): <a href="https://cloud.google.com/transform/scaling-ai-agents-trustworthy-data-foundation-mit-review-insights-report">https://cloud.google.com/transform/scaling-ai-agents-trustworthy-data-foundation-mit-review-insights-report</a></li><li>ILYVEN: <a href="https://ilyven.co/">https://ilyven.co</a></li></ol><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=6cb65250cff0" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Revision 7 of 12: The True Cost of Document Churn]]></title>
            <link>https://medium.com/@carla_1448/revision-7-of-12-the-true-cost-of-document-churn-ff6f6967dc2e?source=rss-cf24ffb761af------2</link>
            <guid isPermaLink="false">https://medium.com/p/ff6f6967dc2e</guid>
            <category><![CDATA[risk-management]]></category>
            <category><![CDATA[data-quality]]></category>
            <category><![CDATA[construction-management]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[construction]]></category>
            <dc:creator><![CDATA[Carla Stone]]></dc:creator>
            <pubDate>Tue, 10 Nov 2026 13:18:23 GMT</pubDate>
            <atom:updated>2026-11-10T13:18:23.481Z</atom:updated>
            <content:encoded><![CDATA[<h3>Drawings and specs are supposed to change. Rework starts when a revision moves faster than the reason behind it, and nobody can say which version a decision was built on.</h3><p><em>By Carla Stone, Founder of ILYVEN</em></p><p>Every set of construction documents is a moving target. A sheet goes out at revision 1, picks up a cloud at revision 3 and by revision 7 of an eventual 12 may look only loosely like the drawing someone priced months ago. That is normal; design is iterative.</p><p>The danger is not that documents change. It is that each revision quietly makes some earlier decision out of date, and most projects cannot see which ones. I call that document churn: a flow of revisions, addenda and re-issued specs that outpaces a team’s ability to track what each change touched.</p><h3>What the research says about design-driven rework</h3><p>In a study published in 1987, the Construction Industry Institute collected deviation data from nine fast-track industrial projects. Rework, repair and replacement averaged 12.4% of total installed project cost. Design deviations, defined as “design changes, errors and omissions,” made up 78% of the deviations and 79% of their cost, about 9.5% of total installed cost. The researchers called those numbers conservative because they excluded schedule impacts (<a href="https://www.construction-institute.org/costs-of-quality-deviations-in-design-and-construction">CII, SD-29</a>).</p><p>A 2014 McGraw Hill Construction SmartMarket Report, produced with the AIA Large Firm Roundtable, adds an owner’s-eye view. Only 7% of owners surveyed believed perfect construction documents are possible. On average, owners said they expect to pay 3% to 5% in added cost because of design errors and omissions. The report also summarizes a 2012 review of more than 25 papers covering 359 projects: direct rework costs from design errors and omissions alone ran 0.5% to 2.6% of construction cost. Including all causes, direct rework costs ran 4.03% to 6.05%, and adding indirect costs brought the estimate to 7.25% to 10.89% (<a href="https://www.agc.org/sites/default/files/Managing%20Uncertainty%20and%20Expectations%20in%20Building%20Design%20and%20Constructio%20%20%20.pdf">McGraw Hill Construction SmartMarket Report</a>). Owner-driven program or design changes ranked as the most impactful cause on cost “by a significant margin.”</p><p>The most recent numbers come from a 2026 survey of 688 construction professionals in the UK and Ireland, commissioned by Procore and run by Dodge. Respondents estimated that 28% of project time is lost to rework, “partly down to a lack of real-time visibility and teams working from outdated documentation and drawings,” and that 18% of project time goes to searching for information (<a href="https://bdcmagazine.com/2026/07/uk-construction-teams-lose-eight-working-weeks-a-year-searching-for-project-information/">BDC Magazine</a>). These are vendor-commissioned respondent estimates, not audited costs, so I read them as a direction, not a measurement.</p><p>Across decades and methods, the pattern holds. Change is expected and priced in. What goes unpriced is work done against information that has already moved on.</p><h3>Why churn is a reliability problem, not a document problem</h3><p>Most firms already have document control, and the best-run projects follow standards that take it seriously. UK guidance on ISO 19650 says it is important “to keep track of the changes between previous and current revisions and versions” and “equally important to also keep track of which revision and version is shared with whom.” It also uses status codes to make clear what a document “should be used for, and by extension, what it should not be used for” — for example, a document still under review should not be used for “procuring materials, agreeing contract costs or constructing the works” (<a href="https://ukbimframework.org/wp-content/uploads/2020/02/ISO_19650_Guidance_Part_2_Processes_for_Project_Delivery_ThirdEdition.pdf">UK BIM Framework, ISO 19650 Guidance Part 2</a>).</p><p>That is the right instinct. But a revision log records what changed in the document, not what was built on the previous version. The estimate was priced from revision 4. The subcontractor’s scope letter references revision 5. The submittal was approved against revision 6. Revision 7 arrives. The document system knows it exists. It does not know that three decisions now rest on something superseded.</p><p><strong>A revision cloud tells you what changed. It cannot tell you what that change just broke.</strong></p><h3>What this looks like on a project</h3><p>I entered construction over a decade ago and later co-founded a residential construction company that reached about $10 million in annual gross sales by 2024. Working in operations, estimating, subcontractor coordination and workflows, I saw costly problems come less from missing information than from scattered, disconnected information. Revisions show that clearly: the newest drawing is almost always available somewhere. The question is whether the decisions downstream of it know it exists.</p><p>Consider a typical scenario. A spec section for an exterior wall is revised to change the air barrier product. The revision is issued and logged correctly. But the takeoff used the earlier version, the installer’s submittal was approved against it and the crew is working from a printed set. Nobody did anything wrong in isolation. The rework comes later, when someone notices the installed product does not match the current spec and the team has to reconstruct which version each decision relied on. That reconstruction is slow and often contested.</p><h3>Where AI fits, and where it falls short</h3><p>AI is getting good at reading documents. It can compare two revisions and summarize what moved faster than a person can. Useful, but not the hard part.</p><p>AI’s real problem isn’t intelligence. It’s reliability, because it forgets why decisions were made. A model reading the current set can tell you what revision 7 says. It cannot tell you why the detail changed or which decisions depended on revision 5, unless that history was captured and connected in the first place.</p><p><strong>Knowing the latest version is easy. Knowing what was decided on the old one is where the money is.</strong></p><h3>Four questions for every revision</h3><p><strong>1. Which decisions rest on the version being replaced?</strong> Estimates, scope letters, submittals and field work all point to a specific revision, whether or not anyone wrote it down.</p><p><strong>2. Why did it change?</strong> Without a recorded reason, everyone downstream has to guess whether the change matters to them.</p><p><strong>3. Who needs to know, and did they acknowledge it?</strong> Issuing a revision is not the same as reaching the people building from it.</p><p><strong>4. Can you show which version each decision used?</strong> This is evidence traceability. When a dispute comes, “we priced revision 4, issued on this date” settles in one line what memory cannot settle in a month.</p><h3>Demonstrated, in development and vision</h3><p>To keep the lines clear: what is demonstrated is the problem. Decades of research link design changes and outdated documents to a meaningful share of rework, and ISO 19650 already treats revision control as core discipline. What remains open is tying each revision to the decisions that depended on it, so construction risk management sees the impact of a change before the rework, not after.</p><h3>Where ILYVEN fits</h3><p>I founded ILYVEN (<a href="https://ilyven.co/">ilyven.co</a>) once AI made a long-held vision feasible: a more reliable intelligence layer, starting in construction, that connects information, preserves the reasoning behind decisions, identifies emerging risks and helps organizations understand the consequences of changes before they become costly mistakes. We are early. This is the direction we are building toward, not a finished product, and nothing here describes a released capability.</p><p>Revision 7 will always come. Enterprise data reliability means that when it does, the project remembers what revisions 1 through 6 decided.</p><h3>Thoughts from the Woman in the Red Dress</h3><blockquote>Drawings are supposed to change. What worries me is a team that cannot say which drawing a decision came from. That is the gap I care about most, and I think it is where AI either earns trust in construction or loses it.</blockquote><p><strong>Carla Stone</strong></p><p>Founder, ILYVEN</p><p><a href="https://ilyven.co/">ilyven.co</a></p><p><strong>Sources (verified loading Nov 10, 2026):</strong></p><ol><li>Construction Industry Institute, “Costs of Quality Deviations in Design and Construction,” Publication SD-29 (Aug 1987): <a href="https://www.construction-institute.org/costs-of-quality-deviations-in-design-and-construction">https://www.construction-institute.org/costs-of-quality-deviations-in-design-and-construction</a></li><li>McGraw Hill Construction, “Managing Uncertainty and Expectations in Building Design and Construction,” SmartMarket Report, with the AIA Large Firm Roundtable (2014), hosted by AGC: <a href="https://www.agc.org/sites/default/files/Managing%20Uncertainty%20and%20Expectations%20in%20Building%20Design%20and%20Constructio%20%20%20.pdf">https://www.agc.org/sites/default/files/Managing%20Uncertainty%20and%20Expectations%20in%20Building%20Design%20and%20Constructio%20%20%20.pdf</a></li><li>BDC Magazine, “UK Construction teams lose eight working weeks a year searching for project information” (Jul 31, 2026): <a href="https://bdcmagazine.com/2026/07/uk-construction-teams-lose-eight-working-weeks-a-year-searching-for-project-information/">https://bdcmagazine.com/2026/07/uk-construction-teams-lose-eight-working-weeks-a-year-searching-for-project-information/</a></li><li>UK BIM Framework, “Information management according to BS EN ISO 19650: Guidance Part 2: Processes for Project Delivery,” Third Edition (Jan 2020): <a href="https://ukbimframework.org/wp-content/uploads/2020/02/ISO_19650_Guidance_Part_2_Processes_for_Project_Delivery_ThirdEdition.pdf">https://ukbimframework.org/wp-content/uploads/2020/02/ISO_19650_Guidance_Part_2_Processes_for_Project_Delivery_ThirdEdition.pdf</a></li><li>ILYVEN: <a href="https://ilyven.co/">https://ilyven.co</a></li></ol><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=ff6f6967dc2e" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The Citation Is the Product]]></title>
            <link>https://medium.com/@carla_1448/the-citation-is-the-product-4495d55dbdfa?source=rss-cf24ffb761af------2</link>
            <guid isPermaLink="false">https://medium.com/p/4495d55dbdfa</guid>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[technology]]></category>
            <category><![CDATA[explainable-ai]]></category>
            <category><![CDATA[trustworthy-ai]]></category>
            <category><![CDATA[ai-governance]]></category>
            <dc:creator><![CDATA[Carla Stone]]></dc:creator>
            <pubDate>Mon, 09 Nov 2026 13:14:50 GMT</pubDate>
            <atom:updated>2026-11-09T13:14:50.199Z</atom:updated>
            <content:encoded><![CDATA[<h3>In enterprise settings, a fluent answer is only a draft. The answer someone else can inspect, trace and challenge is the one you can actually rely on.</h3><p><em>By Carla Stone, Founder of ILYVEN</em></p><p>Ask an enterprise AI assistant a question and you get a polished paragraph. Ask who will answer for that paragraph six months from now, and the polish stops helping.</p><p>That is the gap I want to look at today. Most conversations about AI quality focus on how good the answer sounds. In regulated, contractual and high-consequence work, the better question is what the answer leaves behind: a source someone can open, a version someone can check and a record of who accepted it. In those settings, the citation is not decoration on the product. It is the product.</p><h3>The courts have already written the standard</h3><p>The clearest public example comes from the legal profession. On June 3, 2026, the U.S. Court of Appeals for the Ninth Circuit sanctioned two attorneys over briefs containing multiple nonexistent cases, misattributed quotations and gross misrepresentations of real cases. The court stressed that it was not punishing the use of generative AI itself: “the rules are not violated at the point of research and drafting, but at the point of signing and filing” (<a href="https://cdn.ca9.uscourts.gov/datastore/opinions/2026/06/03/24-4790.pdf">Ninth Circuit, <em>Malkeet LNU v. Blanche</em></a>).</p><p>The order also repeats a line from a California appellate decision: no filing “should contain any citations — whether provided by generative AI or any other source — that the attorney responsible for submitting the pleading has not personally read and verified.” It also found a second failure: a lack of candor about where the errors came from.</p><p>An American Bar Association article from March 2026 surveys how courts are enforcing this. It quotes one federal court: “the signing attorney is the final auditor for all legal and factual claims contained in their motions,” and the certification “attaches at filing and applies regardless of workflow, staffing model, or internal delegation” (<a href="https://www.americanbar.org/groups/litigation/resources/newsletters/professional-liability/generative-ai-ethical-lawyer-standards-different-dilemmas-part-2/">ABA</a>).</p><p>Read that as a design requirement. A human is accountable for the output, so that human must be able to inspect the evidence behind it. A citation that cannot be opened and checked is a claim that only sounds like evidence.</p><h3>Explainability is an audit property</h3><p>NIST’s AI Risk Management Framework lists “explainable and interpretable” among the characteristics of trustworthy AI, and it is specific about why that matters operationally: “Explainable systems can be debugged and monitored more easily, and they lend themselves to more thorough documentation, audit, and governance.” It also separates three ideas that often get blurred. “Transparency can answer the question of ‘what happened’ in the system. Explainability can answer the question of ‘how’ a decision was made in the system. Interpretability can answer the question of ‘why’ a decision was made” (<a href="https://nvlpubs.nist.gov/nistpubs/ai/NIST.AI.100-1.pdf">NIST AI RMF 1.0</a>).</p><p>Eloquence answers none of the three. A smooth paragraph does not tell you what the system looked at, how it reached the conclusion or why it applies to you.</p><h3>Regulators are writing it down as logging</h3><p>The EU AI Act points the same way. Article 12 says high-risk AI systems “shall technically allow for the automatic recording of events (logs) over the lifetime of the system,” and that logging should enable “a level of traceability of the functioning” appropriate to the system’s purpose. For one category of high-risk systems listed in Annex III, the minimum logs include the reference database the input was checked against, the input data that led to a match and “the identification of the natural persons involved in the verification of the results” (<a href="https://artificialintelligenceact.eu/article/12/">EU AI Act, Article 12</a>). The page I read lists application from December 2, 2027 for Annex III systems, so check current timelines before planning around any date.</p><p>Regulators are asking for a trail, not better prose: what was consulted, what matched and who verified.</p><h3>What this looks like on a project</h3><p>I entered construction over a decade ago and later co-founded a residential construction company that reached about $10 million in annual gross sales by 2024. Working in operations, estimating and subcontractor coordination, I saw costly problems come less from missing information than from scattered, disconnected information. When something went wrong, the first question was never “who sounded sure?” It was “show me the document.”</p><p>Consider a typical scenario. A project manager asks an AI assistant whether a subcontractor’s scope includes a particular line item. The assistant answers clearly: no. The manager acts on it and a change order dispute follows. Months later, someone asks what that answer rested on. If the only artifact is a chat transcript, nobody can say which document, which revision or which email thread the answer came from, who reviewed it or whether the underlying facts changed afterward. The answer may have been right. It just cannot be defended.</p><p>That is why I keep saying AI’s real problem isn’t intelligence. It’s reliability, because it forgets why decisions were made. A model can be intelligent in the moment and still leave no inspectable reason behind.</p><p><strong>An answer you cannot trace is an answer you cannot defend.</strong></p><h3>What a citation has to carry to count as evidence</h3><p>This is my own view, not an established standard. For a sourced answer to hold up as audit evidence, I would want it to carry at least four things:</p><p><strong>1. A specific pointer.</strong> Not “per the contract,” but which document, which section and which version.</p><p><strong>2. A date.</strong> Both when the source was current and when the answer was produced, so a later change can be detected.</p><p><strong>3. A visible link between claim and source.</strong> A reviewer should be able to see which statement rests on which passage, rather than a bibliography appended to a paragraph.</p><p><strong>4. A record of acceptance.</strong> Who reviewed the answer, who relied on it and what decision followed.</p><p>The person who signs is the final auditor. Tools should make that audit fast, not impossible.</p><h3>Demonstrated, in development and vision</h3><p>To keep the lines clear: what is demonstrated is the accountability standard: courts are sanctioning unverified citations today, and NIST and EU rules describe documentation, audit and traceability as part of trustworthy AI. What remains an open problem is building systems where inspectable evidence is the default rather than something a reviewer reconstructs by hand.</p><h3>Where ILYVEN fits</h3><p>I founded ILYVEN (<a href="https://ilyven.co/">ilyven.co</a>) with a vision of a more reliable intelligence layer, starting in construction, that connects information, preserves the reasoning behind decisions, identifies emerging risks and helps organizations understand the consequences of changes before they become costly mistakes. We are early. This is the direction we are building toward, not a finished product, and nothing here describes a released capability.</p><p>The measure I keep returning to is simple. When an AI-assisted answer is challenged, can someone who was not in the room open the evidence, see what it relied on and judge it for themselves? Fluent answers will keep getting cheaper. Defensible ones will not.</p><h3>Thoughts from the Woman in the Red Dress</h3><blockquote>On a job site, nobody argues about whose summary sounded better. They ask for the document. I think enterprise AI is heading to the same place, and the teams that build for that question early will earn more trust than the ones with the smoothest answers.</blockquote><p><strong>Carla Stone</strong></p><p>Founder, ILYVEN</p><p><a href="https://ilyven.co/">ilyven.co</a></p><p><strong>Sources (verified loading Nov 9, 2026):</strong></p><ol><li>U.S. Court of Appeals for the Ninth Circuit, Malkeet LNU v. Blanche, №24–4790, Order (June 3, 2026): <a href="https://cdn.ca9.uscourts.gov/datastore/opinions/2026/06/03/24-4790.pdf">https://cdn.ca9.uscourts.gov/datastore/opinions/2026/06/03/24-4790.pdf</a></li><li>American Bar Association, “GenAI and the Ethical Lawyer: Same Standards, Different Dilemmas Part II — How Courts Are Enforcing the Duty of Candor” (Mar 10, 2026): <a href="https://www.americanbar.org/groups/litigation/resources/newsletters/professional-liability/generative-ai-ethical-lawyer-standards-different-dilemmas-part-2/">https://www.americanbar.org/groups/litigation/resources/newsletters/professional-liability/generative-ai-ethical-lawyer-standards-different-dilemmas-part-2/</a></li><li>NIST, “Artificial Intelligence Risk Management Framework (AI RMF 1.0),” NIST AI 100–1 (Jan 2023): <a href="https://nvlpubs.nist.gov/nistpubs/ai/NIST.AI.100-1.pdf">https://nvlpubs.nist.gov/nistpubs/ai/NIST.AI.100-1.pdf</a></li><li>EU Artificial Intelligence Act, Article 12: Record-Keeping (artificialintelligenceact.eu): <a href="https://artificialintelligenceact.eu/article/12/">https://artificialintelligenceact.eu/article/12/</a></li><li>ILYVEN: <a href="https://ilyven.co/">https://ilyven.co</a></li></ol><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=4495d55dbdfa" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>