<?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 Michael Hutson on Medium]]></title>
        <description><![CDATA[Stories by Michael Hutson on Medium]]></description>
        <link>https://medium.com/@michael-hutson?source=rss-d5a7ba3bf4fb------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*fTlTDKsCqFFve1a8Qnk2sA.jpeg</url>
            <title>Stories by Michael Hutson on Medium</title>
            <link>https://medium.com/@michael-hutson?source=rss-d5a7ba3bf4fb------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Fri, 20 Nov 2026 19:35:30 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@michael-hutson/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[AI Is Redefining Mid-Level Leadership by Default]]></title>
            <link>https://medium.com/@michael-hutson/ai-is-redefining-mid-level-leadership-by-default-fbb79cfa4a1b?source=rss-d5a7ba3bf4fb------2</link>
            <guid isPermaLink="false">https://medium.com/p/fbb79cfa4a1b</guid>
            <category><![CDATA[leadership]]></category>
            <category><![CDATA[productivity]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[future-of-work]]></category>
            <category><![CDATA[management]]></category>
            <dc:creator><![CDATA[Michael Hutson]]></dc:creator>
            <pubDate>Fri, 22 May 2026 21:16:00 GMT</pubDate>
            <atom:updated>2026-05-22T21:16:00.551Z</atom:updated>
            <content:encoded><![CDATA[<h3>The activity layer got absorbed, but the judgment layer didn’t. Why organizations are missing the shift.</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*5HMQeKsmrVZMTI8zy9PKGQ.png" /><figcaption>Generate with ChatGPT</figcaption></figure><p>Several technology leaders I’ve spoken with recently have quietly gotten back into building. Claude Code, Codex, spec-driven development. The tools made the work satisfying again. And they’ve all spent less time on the people management side without quite deciding to.</p><p>I caught the same pattern in myself.</p><p>Between contracts, I usually invest the time in professional development to keep nurturing my consulting practice. This stretch, I went deep on two things: how AI is reshaping organizations, and how AI-augmented development is changing the craft itself. Three or four hours a day with these tools. With Claude Code in particular, the work has been genuinely satisfying. It reminded me of the days when I was building applications from the ground up.</p><p>At first, I thought this was a personal pattern. The longer I sat with it, the more it looked like a signal about where the role itself is heading.</p><p><em>If Medium puts up the paywall, </em><a href="https://medium.com/@michael-hutson/fbb79cfa4a1b?source=friends_link&amp;sk=02fb5148087a5329ee8d41aed32dd111"><em>here’s the friend link.</em></a></p><h3>How AI is Redefining Management Roles</h3><blockquote><strong>How is AI changing middle management? </strong>Less than the headlines suggest, and more than the org charts show. The work managers can now produce themselves has expanded fast. The work that defined the role — reading signals, holding context, coaching through ambiguity — has not. The risk isn’t displacement. It’s drift.</blockquote><p>Mid-level management layers have been shrinking for years. Organizations have flattened. Spans of control have widened. The remaining managers have been asked to absorb more work, hold more context, and coordinate across more teams.</p><p>AI tooling arrived in the middle of that shift, not before it.</p><p>What changed is what managers can now produce themselves. Dashboards that used to require an analyst. Decks that used to require a designer. Code that used to require an engineer. For a manager already stretched thin, the tooling feels like relief.</p><p>The signal organizations are sending through what gets praised in performance conversations, through what gets celebrated as “AI fluency,” is that producing artifacts is now part of the manager’s job.</p><p>Deloitte’s most recent State of AI in the Enterprise reports that 84% of organizations have not redesigned jobs around AI capabilities.</p><p>The usual reading is that organizations need to move faster. I read it the other way, and I’ll flag that this is a reading, not a finding.</p><p>The jobs are being redesigned anyway. The tooling is doing it. So is the pressure. So is the absence of leadership clarity on what the role is supposed to be now. None of it on purpose.</p><h3>The Trap of AI Productivity in Leadership</h3><p>The work that defined mid-level leadership in the first place did not get faster with AI.</p><p>Reading team signals before they show up in metrics. Holding strategic context across initiatives that span four or five quarters. Coaching people through ambiguity, which mostly means sitting in it with them. Removing the structural friction no individual contributor can see. Knowing which team is quietly absorbing too much before the data confirms it.</p><p>None of this gets easier. The noise increases. The time pressure compresses. The signals get buried in dashboards. And because the output is invisible, a team that didn’t burn out, an investment that got caught before it locked in, it competes badly for attention.</p><p>In <em>Ask Better Questions</em>, I wrote about how leaders who default to solving problems absorb the thinking that should sit with the team. The same reflex is showing up one level higher, with a different mechanism. The instinct isn’t “let me solve the team’s problem.” It’s “let me build the thing myself.”</p><p><a href="https://medium.com/@michael-hutson/ask-better-questions-how-agile-leaders-move-decisions-closer-to-the-work-4eb5774f6b58">Ask Better Questions: How Agile Leaders Move Decisions Closer to the Work</a></p><p>Producing a dashboard on Saturday feels like progress. The Tuesday conversation with a team lead who has been struggling but hasn’t said so out loud does not feel like progress in the same way.</p><p>One produces an artifact. The other produces an outcome that might not be visible for months.</p><p>The activity layer got absorbed. The judgment layer didn’t. Most organizations haven’t noticed the difference because the dashboards still get built and the decks still get shipped.</p><h3>The Irreplaceable Value of Mid-Level Leaders</h3><p>Mid-level leaders sit at the one layer in the org with proximity to the work and a line of sight into the strategy. That’s it. Above them, the work blurs into abstraction. Below them, the strategy blurs into directives.</p><p>They know which team is about to lose its second senior engineer in six months because of a structural decision made two quarters ago that no one connected to the headcount risk. They read the silence in a retrospective and know what it means. They sit across from a VP and tell them, when it matters, that the strategy that was just announced won’t survive contact with how the work actually flows.</p><p>A model can produce the dashboard. It can’t make that call.</p><p>That capability gets built somewhere else. In the long, under-recognized work of growing people. In the slow work of keeping the operating model coherent while the strategy keeps moving. In the conversations where a manager tells a senior leader something they don’t want to hear.</p><p>The measurements exist for this work. They show up in retention numbers, in the team that still has its strongest people two years after the AI rollout, in the manager who can name three things their team is quietly carrying — but they’re lagging indicators on quarterly cycles, which means they get treated as soft when they’re actually the most reliable signal of whether the leadership capability is being built.</p><p>This is the work that is getting squeezed. Not eliminated. Not formally cut. Squeezed: by the calendar, by the pressure to produce, and by the satisfaction of producing.</p><p>Organizations focused on the software bits right now are disinvesting in this capability without noticing. The line items aren’t moving. The leadership development budget hasn’t been cut. The org charts still show the same titles. What has changed is what the role actually does day to day.</p><p>No one decided this is what should be happening. It is happening anyway.</p><h3>The Advantage Worth Investing In</h3><p>I caught this pattern in myself because I had the space to. Between contracts, with no deliverable pressure, the satisfaction of building was clear enough that I could see it for what it was. A real capability worth having, and a pull away from work that is harder to do and harder to see.</p><p>I’m not arguing managers should stop using AI tools. The tooling works. The productivity is real. The capability is worth building, and I plan to keep doing it.</p><p>A senior leader in an AI-augmented organization has more leverage than they have ever had. The temptation to spend that leverage on what’s most visible is going to keep growing.</p><p>More leverage isn’t the advantage. Knowing where to point it is, and so is the team capability that turns judgment into products customers actually want.</p><p>Both get built in the conversations that don’t produce a dashboard.</p><p><em>This piece is what happens when the tooling makes solving the problem yourself easier than asking the better question. The earlier piece, Ask Better Questions: How Agile Leaders Move Decisions Closer to the Work, looked at how delivery leaders become the constraint by solving problems instead of redistributing the thinking.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=fbb79cfa4a1b" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The Sterile Vibe: What Happens When AI Does 100% of the Work]]></title>
            <link>https://medium.com/@michael-hutson/the-sterile-vibe-what-happens-when-ai-does-100-of-the-work-8d268230fbec?source=rss-d5a7ba3bf4fb------2</link>
            <guid isPermaLink="false">https://medium.com/p/8d268230fbec</guid>
            <category><![CDATA[agentic-ai]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[product-management]]></category>
            <category><![CDATA[software-engineering]]></category>
            <category><![CDATA[future-of-work]]></category>
            <dc:creator><![CDATA[Michael Hutson]]></dc:creator>
            <pubDate>Mon, 04 May 2026 11:26:01 GMT</pubDate>
            <atom:updated>2026-05-04T11:26:01.247Z</atom:updated>
            <content:encoded><![CDATA[<h4>AI agents can execute a delivery playbook, but they can’t simulate years of “production scars.” That gap is the difference between a functional product and a successful one.</h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*MI08q_ttqHMOqxDA9ZNifw.png" /><figcaption>Generated with Gemini</figcaption></figure><p>School pickup. The gap before the bell is the only time parents actually get to talk to each other. I ended up next to a dad I hadn’t met before, a developer at one of the digital-first banks.</p><p>We got onto Claude Code. I told him I’d just built a diagnostic product solo over the previous couple of weeks. Not managing a team, building it. Writing the architecture, designing the experience, running the tests myself. What I’d come away with was the strangeness of building without anyone else in the room no engineer outside the codebase pushing back, no designer watching a user hesitate.</p><p>He nodded. Said his team had landed in the same place. AI can do a lot of the work, but it takes a team to ship something a customer actually connects with.</p><p>Then he gave me the line I drove home thinking about: <strong>“We’re all turning into builders, but we still have our own domain expertise to build together.”</strong></p><p>That’s the conversation I want to extend here.</p><p>If Medium puts up the paywall, <a href="https://medium.com/@michael-hutson/8d268230fbec?source=friends_link&amp;sk=21f8ad309f8a929fc31e1e8f89a9b11f">here’s the friend link.</a></p><p>I did this build deliberately. Over the past two years, I’ve led AI-driven delivery initiatives across financial services and watched how AI development tools were reshaping how teams work. Over the past several months, I’ve been using Claude Code across multiple projects, learning how AI-native development actually behaves in practice.</p><p>I had observations about what was changing. I wanted to pressure-test them from inside the keyboard, not from across the room.</p><p>What I found challenged the assumption I keep hearing in boardrooms: that AI agents can replace specialists. The agents didn’t fail. They performed exactly as designed. That turned out to be the problem.</p><h4>AI Agents Execute Like a Strong Delivery Team</h4><p>The build itself was technical, not casual. Development is where my career started before I moved into delivery leadership, and I went in with the full toolkit — VS Code, Git, GitHub, Claude Code as the AI layer.</p><p>The output held up. Clean, well-structured code. Test-driven development baked into the workflow. CI/CD pipelines, modular architecture, automated testing. The practices I’ve spent years coaching high-performing teams to follow, AI executed without resistance.</p><p>I incorporated agentic coding patterns that have emerged across the AI development community. Structured prompting, eval-driven validation, tool-use workflows. I had the delivery context to know which ones applied. The AI built the way a strong team builds.</p><p>But I was a solo builder. I didn’t have a cross-functional team. So I did what any delivery leader would do. I simulated one.</p><p>I created AI personas for every role that would normally sit across a product development team. Architect. UX designer. QA tester. Developer. Security lead. Risk officer.</p><p>Then I ran them the way I’d run a real team. I created feedback loops where each agent’s output became another agent’s input for review. The architect reviewed the developer’s structural decisions. The QA persona challenged edge cases. The security lens flagged risk assumptions.</p><p>The delivery leadership muscle transferred directly to orchestrating AI agents. How to structure feedback. How to sequence reviews. How to create productive interactions between roles. The same playbook I use to orchestrate teams turned out to orchestrate agents.</p><p>The output looked strong at every stage. Clean specs. Reasonable architecture. Well-structured code. Substantive feedback from each persona. It all felt like progress.</p><h4>The Gap Between Logic and Experience: Use Surfaces What Patterns Can’t</h4><p>No matter how good the discovery, the wireframes, or the agent feedback loops, the experience of using the live product revealed what the abstractions couldn’t.</p><p>Anyone who’s run a sprint demo knows the feeling. The stories move, the tests pass, the architecture holds up on paper — and the product you sit down with afterward is a different artifact.</p><p>AI widens this gap. Because every stage of AI output looks polished, the distance between “passes in development” and “lands in use” stays hidden until someone uses the thing. Each agent validated against patterns. None had ever sat across from a user.</p><p>I call it the <strong>sterile vibe</strong>. It’s functionally correct and logically sound, yet it feels like nobody ever actually used it. Here is what that looks like in practice:</p><ul><li><strong>The Flow:</strong> The diagnostic asked the right questions in the right order — but it felt like a form, not a conversation.</li><li><strong>The Hierarchy:</strong> Nothing in the visual layout told the user what mattered first; every element had the same mathematical weight.</li><li><strong>The Result:</strong> The product worked, but it didn’t lead. It lacked the “opinion” that a human designer brings to a journey.</li></ul><p>These aren’t bugs. They’re the absence of something AI agents can’t simulate. The professional instinct of having sat across from a real user, felt their hesitation, and designed around it.</p><h4>Agent vs Practitioner</h4><p>The table below maps what I observed across every role I simulated. In each case, the AI agent handled the activity with competence. In each case, the practitioner would have brought something the agent couldn’t.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*faEweiwNGty_kHJn3NEMaQ.jpeg" /></figure><p>AI agents validate against patterns. Practitioners validate against experience. The gap between them is where professional instinct lives, and that gap is the work.</p><h4>The Operating Model Mistake</h4><p>Organizations are making operating model decisions right now. Restructuring teams, reducing headcount, redefining roles. All based on the assumption that AI agents can replace specialists. If AI can write the code, design the architecture, and run the tests, why keep paying for the people who used to do those things?</p><p>Because AI absorbed the activity. It didn’t absorb the judgment.</p><p>The scaffolding work is gone. The line-by-line coding, the administration, the mechanical assembly. What’s left is harder to see and harder to automate: the discernment that shapes AI output into something a professional would stand behind. That work hasn’t disappeared. It’s become the entire job.</p><p>The data backs the pattern. A controlled METR study with experienced open-source developers found they were 19% slower using AI tools while believing they were 20% faster — the slowdown migrated into the validation work nobody is yet measuring. Stack Overflow’s 2025 survey shows AI usage above 84% while trust in the tools sits at 29%. That’s the data version of the sterile vibe: adoption is universal, validation is invisible, and the gap between them is sitting in someone’s lap.</p><h4>What Stays When the Activity Is Gone</h4><p>AI didn’t fail in this project. It performed exactly as designed. The agents weren’t broken; they were working within their limits.</p><p>Those limits are invisible during development. They only surface during validation, when a practitioner sits down with the live product and brings the one thing agents can’t simulate: years of doing this work under real conditions, with real users, under real pressure.</p><p>For organizations redefining their operating models right now, the question isn’t whether AI can do the work. It can. If we’re all turning into builders — as the dad at pickup put it — the question for the people redrawing the org chart is whether anyone is left who can tell when the build is done.</p><p>If this resonates with your team’s experience, feel free to drop this summary into your engineering or product channels.</p><p>If Medium puts up the paywall, <a href="https://medium.com/@michael-hutson/8d268230fbec?source=friends_link&amp;sk=21f8ad309f8a929fc31e1e8f89a9b11f">here’s the friend link.</a></p><p><strong>Key Takeaway:</strong> AI can execute a playbook, but it can’t simulate a <strong>“scar.”</strong> The gap between functional code and a successful product is the<strong> human judgment</strong> that tells us when a build is actually <em>done</em>.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=8d268230fbec" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Why Good Leaders Pick Bad Metrics]]></title>
            <link>https://medium.com/@michael-hutson/why-good-leaders-pick-bad-metrics-a3431b2867cd?source=rss-d5a7ba3bf4fb------2</link>
            <guid isPermaLink="false">https://medium.com/p/a3431b2867cd</guid>
            <category><![CDATA[productivity]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[leadership]]></category>
            <category><![CDATA[future-of-work]]></category>
            <category><![CDATA[engineering-mangement]]></category>
            <dc:creator><![CDATA[Michael Hutson]]></dc:creator>
            <pubDate>Fri, 24 Apr 2026 02:37:06 GMT</pubDate>
            <atom:updated>2026-04-24T02:37:06.428Z</atom:updated>
            <content:encoded><![CDATA[<h4>Campbell’s Law, token consumption, and the habit that breaks every measurement system</h4><p>Jensen Huang, CEO of NVIDIA, recently set a new high-water mark for the AI era: the <strong>$250,000 token threshold.</strong> A $500,000 engineer, he argued, should consume half their salary in AI tokens every year. Fall short, and he would be in his words “deeply alarmed.” It is a bold vision for a “10x amplified” workforce, but it serves as a perfect case study in a timeless organizational trap: the moment a diagnostic signal becomes a performance target, it ceases to be a signal at all.</p><p><em>If you’ve hit the Medium paywall, here’s a </em><a href="https://medium.com/@michael-hutson/a3431b2867cd?source=friends_link&amp;sk=ff442797635bccbb9f3398c22dd8c1c5"><em>complimentary link to read the full article</em></a></p><h4>The Distance Problem: Why Data Loses Its Soul</h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*NHoY8fopt-WOv_PZHf-R2Q.png" /><figcaption>The evolution of Engineering Metrics: How signals become Masks.</figcaption></figure><p>Campbell’s Law is simple: the more a metric is attached to high-stakes decisions, the more it gets defended, distorted, or hollowed out. Not because people are dishonest. Because when a number carries enough organizational weight, optimizing for the number becomes rational behaviour.</p><p>The moment token consumption determines compensation, engineers will consume tokens. Whether those tokens produce better decisions, better products, or better outcomes is a different question, and one the metric cannot answer.</p><p>By the time token consumption data reaches the desk of a CEO at a $3 trillion company, it has been filtered through HR dashboards, middle-management reviews, and automated compensation systems. Huang will see a green light on a dashboard. The numbers will look “right.” But the person least positioned to distinguish between a token that produced a breakthrough and one that produced “synthetic busywork” is the person setting the threshold from 30,000 feet.</p><p>The person least positioned to tell the difference between a token that produced insight and one that produced output is the person setting the threshold from the top.</p><h4>The Productivity Cycle: From Lines of Code to Tokens</h4><p>This is not a technology problem. The progression is consistent: lines of code, story points, token consumption. Each generation of leaders found a number that felt like it measured engineering productivity. Each one collapsed the moment engineers optimized for it rather than for outcomes.</p><p>The metric changed. The dynamic didn’t.</p><p>The same pattern appears everywhere a leader attaches a target to a number. Customer satisfaction scores climb because frontline teams learn which interactions to survey and which to let pass. Operational dashboards stay green because definitions shift before numbers do.</p><p>The scorecard looks healthy. The organization is performing for the scorecard.</p><p>I saw this play out while supporting retail operations for a large enterprise client through a consulting engagement. A senior manager told me, with no particular concern in his voice, that as long as he knew which metric needed to move, he could move it. I pushed back gently — asked what he meant, given that the client’s survey metrics had stayed largely unchanged for years. He didn’t hesitate. He explained, matter-of-factly, that stable metrics are the easiest to work. When the same questions get asked in the same way over a long enough period, you learn which levers move the numbers. Which interactions to prioritize before a survey window. Which customers to follow up with. Which friction points to smooth just enough to shift the score without fixing the underlying problem.</p><p>He wasn’t describing something illicit. He was describing institutional knowledge. The metric had been in place long enough that hitting it had become a discipline of its own — disconnected from whatever the original intention was. What started as a signal had become a performance. The number kept moving in the right direction. The executives kept seeing what they needed to see. And the original question the metric was designed to answer had long since stopped being asked.</p><h4><strong>Diagnostic vs. Performance</strong></h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*IsXAUyGQUOLBKbhtiFRPAg.png" /><figcaption>The Signal vs. Target Framework. Why high-stakes engineering productivity metrics often penalize expert engineers (Campbell’s Law).</figcaption></figure><p>A diagnostic metric is designed to surface truth. A performance metric is designed to drive behavior. These can start as the same number. The moment one carries a target, a compensation decision, or a slot on an executive report, it stops being the other. The pressure changes what the number is for — and eventually what it measures.</p><p>Token consumption, on its own, is a reasonable diagnostic signal. How deeply is this engineer engaging with AI tooling? Are they staying in familiar territory or pushing into what’s possible? Used as a proxy for development health, it could surface useful patterns.</p><p>But the moment it becomes a compensation threshold, engineers stop optimizing for judgment. They optimize for throughput. The ones who understand the system best may actually consume fewer tokens — not because they’re falling behind, but because they’ve learned which problems AI handles well and which it handles badly. Under Huang’s model, they’re the ones flagged.</p><h4><strong>What the Metric Cannot See</strong></h4><p>The engineers Huang should be most alarmed about are not the ones running low on tokens. They’re the ones running high without improving anything, generating output that looks productive, reviews clean, ships to production, and quietly accumulates technical debt that nobody will trace back to the prompt.</p><p>Token consumption cannot distinguish between an engineer who used AI to compress a three-day architecture decision into a focused afternoon, and one who used it to generate 4,000 lines of code the rest of the team will spend six months untangling.</p><p>The number looks the same. The outcomes don’t.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*y0WuOoz2EAxTk267G26Ulw.png" /><figcaption>A 4-quadrant framework for using engineering metrics diagnostically: Keep it diagnostic, open conversations, inspect output, and name the purpose.”</figcaption></figure><p>Huang’s instinct is not wrong. Whether engineers are genuinely engaging with AI tooling is worth understanding. Token consumption can surface at a team level, over time, as one signal among several. That’s a legitimate diagnostic use.</p><p>The mistake is the leap from signal to threshold. From observation to compensation instrument.</p><p>There is a more useful way to work with this data. It requires treating token consumption as what it actually is, a prompt for inquiry, not a verdict.</p><h4><strong>1. Keep it diagnostic</strong></h4><p>Token consumption tells you something about how deeply engineers are engaging with AI tooling and whether they are pushing into what is possible. That is worth knowing. It is not worth pricing.</p><h4><strong>2. Use it to open conversations, not close them.</strong></h4><p>If a team’s consumption is unexpectedly low, ask why. Are they under-equipped, under-trained, or working on problems where AI adds little value? Or have they done something harder — learned to use AI precisely enough that they are delivering the same output at lower cost? A team that has optimized token usage while continuing to produce value is not underperforming. They may be leading. The number creates the question. It doesn’t answer it.</p><h4><strong>3. Inspect the output of high consumers.</strong></h4><p>High token usage could mean an engineer compressed a three-day architecture decision into a focused afternoon. It could also mean they generated thousands of lines of code the rest of the team will spend months untangling. The metric looks identical in both cases. The outcome does not.</p><h4><strong>4. Name the metric’s purpose before it goes on a dashboard.</strong></h4><p>That conversation with the senior manager changed how I introduce metrics to delivery teams. I now name the distinction explicitly before anything gets tracked, this number is here to help us see, not to tell us how we’re doing. Without that clarity, a target-based system will penalize your best engineers, the ones who understand AI’s limitations well enough to use it precisely, and consequently consume fewer tokens. They are the ones Huang’s model would flag.</p><blockquote>A diagnostic metric and a performance metric can share the same number. They cannot share the same organizational weight.</blockquote><p>Most of what makes token consumption useful as a signal is not measurement discipline. It is leadership discipline. The questions in that list are not audit questions. They are the questions a leader asks when they see a number and resist the urge to treat it as an answer. The difference between a leader who uses metrics well and one who gets gamed by them is rarely analytical ability. It is the habit of staying curious one moment longer than feels necessary and coaching their team to do the same.</p><h4><strong>The Question Before the Target</strong></h4><p>Before any metric gets attached to a consequence, one question is worth asking. Not whether the number can be tracked. Whether tracking it will still tell you the truth once it matters.</p><p>Most metrics can survive observation. Very few survive accountability.</p><p>We’ve been here before. The metric always looks right until it matters. Then it tells you exactly what people wanted you to hear.</p><p>If you found these insights useful for your own practice, feel free to share this summary in your internal communities of practice or engineering channels.</p><blockquote><strong><em>TL;DR:</em></strong><em> When a diagnostic signal, like AI token consumption, is turned into a performance target, it stops measuring productivity and starts measuring “metric theater.” The further a number travels from the engineer to the executive, the more it loses its soul. Real leadership isn’t about hitting a threshold; it’s about staying curious enough to ask what the numbers actually produced. “</em><a href="https://medium.com/@michael-hutson/a3431b2867cd?source=friends_link&amp;sk=ff442797635bccbb9f3398c22dd8c1c5">https://medium.com/@michael-hutson/a3431b2867cd?source=friends_link&amp;sk=ff442797635bccbb9f3398c22dd8c1c5</a>&quot;</blockquote><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=a3431b2867cd" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Unlocking the Real ROI of AI: How to Turn Faster Coding into Faster Delivery]]></title>
            <link>https://medium.com/@michael-hutson/why-your-ai-investment-is-currently-rotting-in-a-queue-f8ec78bd8da7?source=rss-d5a7ba3bf4fb------2</link>
            <guid isPermaLink="false">https://medium.com/p/f8ec78bd8da7</guid>
            <category><![CDATA[software-development]]></category>
            <category><![CDATA[agile]]></category>
            <category><![CDATA[leadership]]></category>
            <category><![CDATA[productivity]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <dc:creator><![CDATA[Michael Hutson]]></dc:creator>
            <pubDate>Tue, 21 Apr 2026 04:12:16 GMT</pubDate>
            <atom:updated>2026-04-21T13:24:55.907Z</atom:updated>
            <content:encoded><![CDATA[<h4>AI Is Coding at 2026 Speeds, but Your Governance Is Stuck in 2019. Here Is Why Faster Code Isn’t Creating Faster Value.</h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*0UeXujNQjf0bP7W74W2K3A.png" /></figure><p>If Medium puts up the paywall, here’s the <a href="https://medium.com/@michael-hutson/f8ec78bd8da7?source=friends_link&amp;sk=cebd3707cad90adf57f4a00e5e005c93">friend link</a>. Let’s skip the velvet rope.</p><p>Building software used to feel like commissioning a bespoke suit.</p><p>A craftsman with years of training. Measurements, fittings, adjustments. A final product built entirely by hand, shaped around one specific set of requirements that rarely changed mid-process. It was skilled work, and it was slow by design.</p><p>Then came a series of shifts that fundamentally changed how software gets built. Agile challenged the assumption that you had to design everything before building anything — introducing iterative development, continuous feedback, and the idea that organizations could learn their way to the right product rather than plan their way there. CI/CD pipelines, DevOps practices, automated testing, and cloud infrastructure removed friction at every stage. And now AI-assisted development has compressed timelines further still. What once required a craftsman and a quarter can now be assembled, tested, and ready for release in days.</p><p><strong>The development side of the system has transformed. The organizational system around it largely has not.</strong></p><p>And that gap is where delivery value disappears.</p><h3>The Pattern I See in Every Engagement</h3><p>As an Agile delivery consultant, my role is to make visible what organizations cannot see themselves — analyzing existing delivery processes and surfacing the organizational systems that slow value across enterprise initiatives.</p><p>Early in almost every engagement, before I have reviewed a single dashboard or sat in a planning session, a pattern emerges.</p><p>The development team is moving. Tickets are closing. Code is shipping into review. There is visible momentum. And yet, somewhere downstream, the pace of actual delivery tells a different story. Features are not reaching production faster. Value is not reaching customers sooner. The roadmap is still slipping.</p><p>The first place I look is never the development team. It is the system around them.</p><p>The change advisory board meeting weekly. The security review running on its own timeline. The integration checkpoints, the approval layers, the operational readiness process designed for a release frequency the team is no longer running. That is where the constraint usually lives — not in how fast the team builds, but in how the organization is structured to absorb and release what the team produces.</p><p>AI has not changed this. In most organizations, it has made it more visible.</p><h3><strong>The Bottleneck: Goldratt’s Theory of Constraints</strong></h3><p>The Theory of Constraints, developed by Eliyahu Goldratt, makes a simple claim: a system performs at the pace of its slowest point. Improve every other step, and overall throughput barely moves. The constraint sets the ceiling, not the fastest step in the chain.</p><p>Developers generate code faster. Pull requests are created sooner. Task completion accelerates. And then the work reaches the gate.</p><ul><li>The <strong>Change Advisory Board </strong>meeting weekly.</li><li>The <strong>security reviews </strong>that running on their own timeline.</li><li>The <strong>operational readiness</strong> processes were designed for 2019.</li></ul><p>The work accelerates into the constraint. The constraint does not move.</p><h3>What the Data Shows</h3><p>The pattern I see in engagements has a number attached to it. Bain and Company researched this directly and found that <strong>coding and testing account for roughly 25 to 35 percent</strong> of the time from idea to launch. The rest — review, integration, approvals, and release coordination — remains largely untouched by AI.</p><blockquote>Speeding up a quarter of the system does not move the system.</blockquote><p>This is why measuring AI impact through task speed, output, or ticket velocity misses the point. Those metrics capture the part of the system AI already improved. Local acceleration is not the same as delivery improvement. If the governing bottleneck remains untouched, the backlog just forms somewhere else — and too often, that makes development look like the problem when it is not.</p><h3><strong>The Real Question: Is Your Governance AI-Ready?</strong></h3><p>The most useful question in any AI productivity conversation is not how much faster developers are moving.</p><p>It is whether the organizational systems surrounding development, the governance cadences, the approval structures, and the release processes were designed for the delivery environment that existed before AI arrived. In most organizations, they were. And most organizations have not yet asked whether they still fit.</p><p>AI is a powerful accelerator. But an accelerator pointed at the wrong part of the system produces faster queues, not faster delivery. The development side of the equation has transformed. The organizational side is still catching up.</p><p>That is where the conversation needs to go next.</p><h4><strong>Internal Share: How to Unlock the Real ROI of AI</strong></h4><p>If you found these insights useful for your own practice, feel free to share this summary in your internal communities of practice or engineering channels.</p><p>TL;DR for Slack/Teams: “AI is coding at 2026 speeds, but most organizations are still releasing value at 2019 governance speeds. The constraint was never development; it was the approval layers, security reviews, and operational readiness processes surrounding it. Teams are shipping faster into the same CAB cadences and release gates that existed before AI arrived. Key takeaway: Local acceleration is not the same as delivery improvement — speeding up a quarter of the system does not move the system. You can read the full breakdown here: [<a href="https://medium.com/@michael-hutson/f8ec78bd8da7?source=friends_link&amp;sk=cebd3707cad90adf57f4a00e5e005c93">https://medium.com/@michael-hutson/f8ec78bd8da7?source=friends_link&amp;sk=cebd3707cad90adf57f4a00e5e005c93</a>]</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=f8ec78bd8da7" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[How to Build a Learning Climate for the AI Era]]></title>
            <link>https://medium.com/@michael-hutson/how-to-build-a-learning-climate-for-the-ai-era-7d90d3c607ec?source=rss-d5a7ba3bf4fb------2</link>
            <guid isPermaLink="false">https://medium.com/p/7d90d3c607ec</guid>
            <category><![CDATA[engineering-management]]></category>
            <category><![CDATA[teamdesign]]></category>
            <category><![CDATA[culture-learning]]></category>
            <category><![CDATA[future-of-work]]></category>
            <category><![CDATA[agile-leadership]]></category>
            <dc:creator><![CDATA[Michael Hutson]]></dc:creator>
            <pubDate>Thu, 16 Apr 2026 17:10:54 GMT</pubDate>
            <atom:updated>2026-04-17T21:10:27.131Z</atom:updated>
            <content:encoded><![CDATA[<h3>Why your AI strategy is failing the 90%: A delivery leader’s guide to building teams that actually get smarter.</h3><h4>As AI accelerates output, the ‘Apprenticeship Gap’ is widening. Learn 5 tactical moves to protect your team’s judgment and prevent a productivity plateau.</h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*j-0VIjFV3gPgBxjv" /><figcaption>Photo by <a href="https://unsplash.com/@bluext?utm_source=medium&amp;utm_medium=referral">blue sky</a> on <a href="https://unsplash.com/?utm_source=medium&amp;utm_medium=referral">Unsplash</a></figcaption></figure><p><em>If Medium puts up the paywall, here’s the </em><a href="https://medium.com/@michael-hutson/how-to-build-a-learning-climate-for-the-ai-era-7d90d3c607ec?source=friends_link&amp;sk=6756369c644ec3236eca77aeb1d0dcbd"><em>friend link</em></a><em>. Let’s skip the velvet rope.</em></p><h3>The 10% Trap</h3><p>Organizations are currently rushing to fund AI tools and “adoption support.” They buy the licenses, host a few webinars, and consider the box checked.</p><p>But there is a massive gap between <strong>access to a platform</strong> and the <strong>capability to use it.</strong></p><p>In my delivery practice, I’ve seen how easily focus narrows under pressure. We obsess over the product in front of us because that’s what keeps delivery on track. But moving a team from good to great doesn’t happen in a training manual, it happens in the “breathing room” between tasks.</p><p><strong>That space is the first thing to disappear when the pressure mounts.</strong></p><h3>The 70–20–10 Reality</h3><p>Research from the Center for Creative Leadership highlights why a “training-first” approach fails.</p><p>Their <strong>70–20–10 model</strong> shows that:</p><ul><li><strong>70%</strong> of capability comes from on-the-job experience.</li><li><strong>20%</strong> comes from coaching and peer interaction.</li><li><strong>10%</strong> comes from formal coursework.</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/720/1*BJNNKVoK6xgxPvh1eqhptg.jpeg" /><figcaption>Generate with Claude AI</figcaption></figure><p>If that ratio holds, then <strong>90% of your team’s development happens in the environment itself.</strong> It doesn’t happen on a platform. It happens in the team, through the work.</p><p>The distinction is simple: A training program addresses the 10%. A learning climate protects the 90%.</p><p>Most organizations invest heavily in the first and leave the second to chance. In the AI era, that is a gamble you can no longer afford to take.</p><h3>Why This Shows Up in Delivery</h3><p>Teams that keep getting better share habits that have nothing to do with velocity. They surface problems early. They treat mistakes as information. They run postmortems that lead to structural changes, not just action items.</p><p><strong>Teams that plateau look different:</strong></p><ul><li><strong>Green Metrics, Red Reality:</strong> They hit their numbers but repeat the same failures.</li><li><strong>Polite Retros:</strong> Meetings are unproductive because nobody wants to “break the vibe.”</li><li><strong>Silence as a Signal:</strong> Problems only surface when they become crises.</li></ul><p>The quietest teams are rarely the healthiest. One famous study on psychological safety found that high-performing teams reported <strong>more</strong> errors than low-performing ones because they learned from them rather than concealing them.</p><h3>AI is Compressing the “Apprenticeship Path”</h3><p>This is where the wheels usually come off. AI is accelerating output, but it is also narrowing the “apprenticeship path” the slow, relationship-based learning where juniors build judgment by watching seniors navigate ambiguity.</p><blockquote><strong><em>AI can generate the answer. It cannot teach someone how to know when the answer is wrong.</em></strong></blockquote><p>When that learning path narrows and the team has no climate to compensate, judgment erodes. You end up with high velocity but low resilience.</p><h3>You Don’t Need Permission to Start</h3><p>A learning climate isn’t declared by an executive; it is designed by the leaders closest to the work. You don’t need a multi-year “transformation” to start. You need to make these five deliberate moves:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/720/1*b3o2DCFYtMEVOEWxtoOGHQ.jpeg" /><figcaption>Generated with Claude AI</figcaption></figure><h4>1. Protect Team Continuity</h4><p>Learning compounds when people stay together long enough to develop shared context. In <a href="https://teamtopologies.com/">Team Topologies</a>, Skelton and Pais argue that long-lived teams outperform project groups because capability accumulates rather than resetting with every new “resource” allocation.</p><p><strong>The Move:</strong> Push back when a high-performing team is about to be broken up. Every reassignment resets the learning clock.</p><h4>2. Control the Experience of Failure</h4><p>Creating safety across a company is a long game, but you can create it in your next standup. When someone admits a mistake, do you thank them for the data, or do you ask “why” in a tone that implies blame?</p><p><strong>The Move:</strong> Make your incident reviews the place where honesty is met with curiosity instead of consequences.</p><h4>3. Ask One Question to Make Learning Visible</h4><p>Learning has no seat at the table unless you pull up a chair. Add this single question to your next sprint review:</p><p>“What did we learn this sprint that we didn’t know when we planned it?”</p><p><strong>The Move:</strong> If the team can’t answer, your climate is absent. Normalizing this question makes learning an explicit part of delivery, not an extracurricular.</p><h4>4. Create “Focus Blocks” for Deep Work</h4><p>In the AI era, the temptation is to fill the time saved by tools with more shallow tasks. This kills the reflection necessary for integration. Learning doesn’t happen in 10-minute gaps between meetings.</p><p><strong>The Move:</strong> Implement a “No-Meeting Wednesday” or a daily 2-hour “Deep Work” block. Frame it to leadership as a <strong>Delivery Acceleration Zone</strong>.</p><h4>5. Normalize “Show and Tells” (The Peer 20%)</h4><p>Most teams keep their AI “hacks” or workflow improvements to themselves, leaving the 20% peer-learning gap wide open.</p><p><strong>The Move:</strong> Dedicate 15 minutes a week to an informal “Low-Stakes Demo.” No slides, no prep — just a team member sharing one small prompt, bug, or process tweak they discovered.</p><h3>The Cost of the “Default” Climate</h3><p>Every organization has a learning climate. If you haven’t designed it intentionally, it likely emerged by default, shaped by deadline pressure and risk aversion.</p><p>In an AI-driven environment, a “default” climate is a dangerous one. When output speeds up, but reflection slows down, you aren’t actually becoming more efficient you’re just accelerating toward a plateau.</p><h3>Executive Summary for Internal Sharing</h3><p>If you found these insights useful for your own practice, feel free to share this summary in your internal communities of practice or engineering channels.</p><blockquote><strong><em>TL;DR for Slack/Teams:</em></strong><em> “Organizations are currently obsessing over the 10% (AI tools and training) while neglecting the 90% (the environment where capability actually grows). This guide outlines 5 tactical moves to protect our </em><strong><em>Learning Climate</em></strong><em> and prevent the ‘Apprenticeship Gap’ as AI accelerates our output. Essential reading for anyone managing delivery teams right now.</em></blockquote><blockquote><strong><em>Full Article: [</em></strong><a href="https://medium.com/@michael-hutson/how-to-build-a-learning-climate-for-the-ai-era-7d90d3c607ec?source=friends_link&amp;sk=6756369c644ec3236eca77aeb1d0dcbd">https://medium.com/@michael-hutson/how-to-build-a-learning-climate-for-the-ai-era-7d90d3c607ec?source=friends_link&amp;sk=6756369c644ec3236eca77aeb1d0dcbd</a>]</blockquote><h3><strong>References</strong></h3><p>Center for Creative Leadership — 70–20–10 Model for Learning and Development</p><p>Edmondson, A. C. — <em>The Fearless Organization: Creating Psychological Safety in the Workplace for Learning, Innovation, and Growth</em></p><p>DORA — State of DevOps / State of AI-assisted Software Development Research (dora.dev)</p><p>Skelton, M. &amp; Pais, M. — <em>Team Topologies: Organizing Business and Technology Teams for Fast Flow</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=7d90d3c607ec" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Your AI Tools Opted In for You]]></title>
            <link>https://medium.com/@michael-hutson/your-ai-tools-opted-in-for-you-4a54258ce7ff?source=rss-d5a7ba3bf4fb------2</link>
            <guid isPermaLink="false">https://medium.com/p/4a54258ce7ff</guid>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[leadership]]></category>
            <category><![CDATA[engineering-management]]></category>
            <category><![CDATA[software-engineering]]></category>
            <category><![CDATA[agile]]></category>
            <dc:creator><![CDATA[Michael Hutson]]></dc:creator>
            <pubDate>Mon, 13 Apr 2026 22:16:01 GMT</pubDate>
            <atom:updated>2026-04-17T20:39:25.780Z</atom:updated>
            <content:encoded><![CDATA[<h4>The shift from opt-in to opt-out has already happened. Agents are running in your tools right now.</h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*Y8-GCW8ZD9ekUAI0" /><figcaption>Photo by <a href="https://unsplash.com/@borisview?utm_source=medium&amp;utm_medium=referral">boris misevic</a> on <a href="https://unsplash.com/?utm_source=medium&amp;utm_medium=referral">Unsplash</a></figcaption></figure><p><em>If Medium puts up the paywall, here’s the </em><a href="https://medium.com/@michael-hutson/your-ai-tools-opted-in-for-you-4a54258ce7ff?source=friends_link&amp;sk=ff93cc67129376a468e0672ceda39a2f"><em>friend link</em></a><em>. Let’s skip the velvet rope.</em></p><p>Sometime this month, an agent started making decisions inside your delivery system. You were not consulted. You may not have been told.</p><p>Atlassian automatically activated Rovo Dev — its agentic AI for software development — in all Premium and Enterprise Jira environments. Not a prompt to evaluate. Not an opt-in. A notification that it was on, with an option to disable it if you acted quickly. Rovo Dev is not a passive assistant. It operates inside your planning layer. It interprets tickets, decomposes work, opens pull requests. The defaults it shipped with determine what gets surfaced to your team, what gets treated as complete, and what enters the workflow at all. Your organization didn’t configure those defaults. Atlassian did.</p><p>That is not an accusation. It is a description of something most delivery leaders have not yet named.</p><h3>What happened this month</h3><p>Atlassian is not alone. In the same window, GitHub changed its Copilot training data policy for Free, Pro, and Pro+ users — effective April 24. Interaction data, including inputs, outputs, and the context your developers provide when prompting, now trains GitHub’s models unless individual users opt out. Copilot Business and Enterprise users are exempt. Everyone else is not, and the opt-out is per account, not per organization. A delivery leader cannot fix this with a policy. They have to fix it with a conversation, developer by developer.</p><p>Slack pushed more than 30 AI capabilities into Slackbot across its platform in a single update<strong>.</strong> Slackbot now synthesizes threads, surfaces recommendations, and summarizes meetings. The coordination layer your team uses to stay aligned is now being interpreted by an agent operating under default settings you did not choose.</p><p>Microsoft, meanwhile, launched Agent 365 — positioned explicitly as “the control plane for AI agents” — into public preview this month, with general availability on May 1. It is bundled into Microsoft 365 E7 at $15 per user per month. The framing matters: Microsoft is selling governance infrastructure for a problem they are acknowledging directly. An enterprise governance product at that price point does not exist unless the problem is real, widespread, and recognized at the buying level.</p><p>The pattern across all four is the same. <strong>Vendors are not waiting for organizational deployment decisions. They are deploying by default and offering controls afterward. The organization that believes it controls when agents arrive has already been passed by.</strong></p><h3>What that actually means</h3><p>Your organization evaluated the tools. It did not evaluate what the tools decided on your behalf.</p><p>When delivery leaders hear “agents in your tools,” the reflex is to reach for the settings review. Check what’s enabled. Verify permissions. Update the acceptable use policy. That is the right response to a feature rollout. It is not the right response to a delegation shift.</p><p><strong>Settings control features. Defaults control judgment. Those are not the same review.</strong></p><p>The judgment being affected is not abstract. It maps directly to specific decisions inside the delivery system:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*c1BFEIoNy5FeGN2c2eTikg.png" /><figcaption>Generated with Claude</figcaption></figure><p>These are delegation decisions. When Rovo Dev determines what surfaces to your team’s attention, that is a judgment call about priority and risk. When Slackbot synthesizes a meeting and your team treats the summary as the shared record, that is a judgment call about what the conversation meant. The agent made it. Your organization didn’t authorize it. Most delivery leaders haven’t noticed the difference.</p><p>The organizations most exposed are not the ones where agents are doing things wrong. They are the ones where agents are doing things right — automatically, without a deliberate handoff.</p><p>When Rovo Dev correctly completes a task a team member was responsible for, accountability distributes quietly. Nobody announced the handoff. Nobody updated the accountability model. The tool looks like it’s working. Over time, delivery leaders find they have less visibility into who owns what, not because the system is broken, but because the system is completing work on behalf of people who never formally transferred it.</p><p>The loss is invisible until it isn’t. The signals that never reached the system don’t show up in throughput metrics. They show up in the dependency that surfaced too late, the risk nobody named, the conversation that should have happened before the sprint.</p><blockquote><em>Before you move on from this section — when did your organization last review not what your AI tools do, but what they decide? The settings review and the delegation review are not the same conversation. Most organizations have only had one of them.</em></blockquote><h3>What to act on now</h3><p>The GitHub training data deadline is the most immediate. April 24 is three weeks away. Developers on Free, Pro, or Pro+ Copilot who have not opted out by then are contributing interaction data to GitHub’s training corpus under the new policy. The coverage of this has focused on code. The more important exposure may be context: the architectural decisions, naming conventions, and design rationale that developers embed in prompts. That content is harder to audit and harder to recover than a code snippet. Delivery leaders who do not know which Copilot tiers their developers use do not have enough information to act.</p><p>Atlassian Rovo requires a different kind of review. Awareness that it is active is not governance. A governance position for Rovo Dev means a defined scope — what tasks it can handle autonomously — an escalation path, and an accountability assignment. Without those three things, the agent is operating in your planning layer under defaults that no one in your organization selected.</p><p>On Agent 365: its commercial existence is itself a signal. Microsoft is not building governance infrastructure for a theoretical future problem. They are building it because enterprises are already running agents without governance and are beginning to recognize that. If your organization is in that position, the arrival of Agent 365 is not an opportunity to wait for a cleaner solution. It is a confirmation that the gap is real and that the cost of addressing it is now priced.</p><h3>Three questions your organization should be able to answer</h3><p>Most delivery organizations have done the adoption review. They evaluated the tools, approved the rollout, updated the acceptable use policy. That review is complete.</p><p>The delegation review is a different exercise.</p><p><strong>Which decisions in your delivery system are now being shaped by agent defaults rather than by your team </strong>— which dependencies are being surfaced or filtered, which planning recommendations are being treated as shared direction, which thread summaries are becoming the official record of what was agreed?</p><p><strong>Where in the workflow is agent judgment invisible — not because it is hidden, but because no one has looked?</strong></p><p><strong>What does your organization still want a human to own explicitly — triage, escalation, the dependency that requires a conversation rather than a ticket </strong>— and have you said so anywhere the tools can reflect it?</p><p>If you cannot answer those three questions for the agents currently running in your delivery environment, your organization has done the adoption review and skipped the delegation review. Those are not the same thing. Right now, in most organizations, only one of them has happened.</p><h3>The model that changed</h3><p>The operating model did not change when your organization adopted AI tools. It changed when the vendor defaults changed — quietly, inside tools that were already in use, without a deployment decision from you.</p><p>That is a harder thing to govern than a tool you chose. There is no moment of adoption to point back to, no implementation decision to revisit. The change is already inside the workflow. The question now is whether your organization has examined what it inherited, and whether the judgment currently living in those defaults is the judgment you would have authorized if you had been asked.</p><p><strong>You adopted the tools. You did not authorize the delegation. That distinction is now the most important governance question in your delivery system — and most organizations have not asked it yet.</strong></p><p>If you found these insights useful for your own practice, feel free to share this summary in your internal communities of practice or engineering channels.</p><h3>TL;DR for Slack/Teams</h3><blockquote><em>“We didn’t just adopt AI; we inherited a new operating model. With vendors like Atlassian and GitHub moving to ‘opt-out’ defaults, agents are already making decisions inside our delivery system without a formal handoff.”</em></blockquote><blockquote><strong><em>Key takeaway:</em></strong><em> </em><strong><em>Settings control features, but defaults control judgment.</em></strong><em> We need to stop doing ‘Settings Reviews’ and start doing ‘Delegation Reviews’ to define what a human must still own.</em></blockquote><blockquote><strong><em>Read the full breakdown here:</em></strong><em> [</em><a href="https://medium.com/@michael-hutson/your-ai-tools-opted-in-for-you-4a54258ce7ff?source=friends_link&amp;sk=ff93cc67129376a468e0672ceda39a2f">https://medium.com/@michael-hutson/your-ai-tools-opted-in-for-you-4a54258ce7ff?source=friends_link&amp;sk=ff93cc67129376a468e0672ceda39a2f</a><em>]</em></blockquote><p><em>In the Flow of Work publishes weekly. If this issue was useful, share it with a delivery leader who is navigating the same questions.</em></p><p><em>I started this series because I kept having the same conversation with delivery leaders, Agile coaches, and programme managers who could see AI changing their organizations but were not finding many places to think through what it really meant.</em></p><p>Sources</p><p><strong>Atlassian</strong> — Rovo Dev automatic activation for Premium and Enterprise accounts, April 2026. Atlassian Support documentation and Atlassian blog.</p><p><strong>GitHub</strong> — “GitHub Copilot — Changes to data use for Free, Pro, and Pro+ users,” GitHub Blog, effective April 24, 2026. Additional coverage: The Register, March 26, 2026; InfoQ.</p><p><strong>Salesforce / Slack</strong> — Slackbot 3.0 AI capabilities rollout, March 31, 2026. TNW (The Next Web) and Slack product announcement.</p><p><strong>Microsoft</strong> — Agent 365 public preview and general availability, April–May 2026. Microsoft 365 Blog; NexusTek and LevelUpM365 roadmap analysis.</p><p><strong>QCon London 2026</strong> — State-of-the-field practitioner reporting on coding agent deployments, March 16–19, 2026. InfoQ (multiple articles), The Register (March 18, 2026).</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=4a54258ce7ff" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The Coordination Trap: Why AI Updates Aren’t the Same as Shared Understanding]]></title>
            <link>https://medium.com/@michael-hutson/when-the-noticing-gets-automated-5545c2a32929?source=rss-d5a7ba3bf4fb------2</link>
            <guid isPermaLink="false">https://medium.com/p/5545c2a32929</guid>
            <category><![CDATA[software-development]]></category>
            <category><![CDATA[agile]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[future-of-work]]></category>
            <category><![CDATA[leadership]]></category>
            <dc:creator><![CDATA[Michael Hutson]]></dc:creator>
            <pubDate>Mon, 06 Apr 2026 15:06:47 GMT</pubDate>
            <atom:updated>2026-04-17T20:42:55.443Z</atom:updated>
            <content:encoded><![CDATA[<h4>AI agents are removing the friction where human judgment used to be required. Here is how to avoid the “Coordination Debt” they leave behind</h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/720/1*2d-U2EuTSU71QJpj1Cvgow.jpeg" /><figcaption>Generated with NotebookLM</figcaption></figure><p><em>If Medium puts up the paywall, here’s the </em><a href="https://medium.com/@michael-hutson/when-the-noticing-gets-automated-5545c2a32929?source=friends_link&amp;sk=a57fe2e02db0ac4820d6a8da896dc3c3"><em>friend link</em></a><em>. Let’s skip the velvet rope.</em></p><h3>In the Flow of Work</h3><p>A briefing for delivery leaders navigating AI-driven change across Agile, flow, and execution.</p><p><strong>A practitioner’s take</strong></p><p><em>I started this series because I kept having the same conversation with delivery leaders, Agile coaches, and programme managers who could see AI changing their organizations but were not finding many places to think through what it really meant.</em></p><p>Something changed in March, but it did not arrive like a transformation. It arrived like a feature being turned on.</p><p>In the span of four weeks, GitHub put Copilot coding agent for Jira into public preview, letting teams invoke an agent directly from a Jira work item and open a pull request without leaving the workflow. Linear launched an agent operating inside the roadmap environment, able to synthesize context, make recommendations, and take action across issues and code. Slack repositioned Slackbot as a personal AI agent across the workspace, drawing on messages, files, channels, and connected tools. Atlassian, meanwhile, restructured to fund deeper AI investment across its platform. None of this required an organizational redesign. It required a configuration.</p><p>From a delivery perspective, that matters because the distance between planning, coordination, and execution is shrinking inside the tools themselves. The same systems that once tracked work are starting to interpret it, sequence it, summarize it, and initiate action on it. Agents are not just reducing coordination effort. They are removing some of the recurring moments where human judgment used to be structurally required. The risk is not lower visibility inside the tools. It is that organizations mistake tool-generated updates for shared understanding and stop noticing what never entered the system in the first place.</p><h3>What changed in the workflow</h3><p>Delivery leaders did not own all coordination work. They never did. What they often held was something more specific: the cross-team acts that kept coordination legible. Surfacing the dependency two teams had not named. Forcing unresolved ambiguity into the open before it reached a sprint. Not owning the whole layer: maintaining its legibility, across boundaries, under pressure, before things broke.</p><p>The planning session, the dependency review, the board walkthrough were not valuable because they were rituals. They were valuable because they created structured inspection points: recurring moments where someone had to ask: what matters here, what is at risk, what is not yet visible enough to manage. When agents take the trigger function, the judgment does not automatically move with it.</p><p>Agents are now operating in that same space. A Jira agent interprets the ticket and decomposes the work. Linear Agent acts within the roadmap it understands. Slackbot synthesises signals across channels and surfaces a summary. The cycle time data from Anthropic’s 2026 Agentic Coding Trends Report shows the structural shift clearly: traditional SDLC stages that took weeks now take days. The throughput gains are real.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*NBiRdhEJRb8M4VNd2TG3nA.png" /><figcaption>Adapted from Anthropic 2026 Agentic Coding Trends Report</figcaption></figure><p><em>Traditional SDLC vs. agentic SDLC: cycle times compress. So do the moments where human judgment is structurally required.</em> <em>Adapted from Anthropic 2026 Agentic Coding Trends Report, Trend 1</em></p><p>But there is a limit the throughput data does not capture.</p><p>Not all meaningful delivery signals originate inside the system. Some emerge through a tense conversation between two leads that does not make it into Jira. Through stakeholder hesitation in the room that is never formalized. Through a regulatory shift or competitor move that arrives through a conversation before it arrives through a tool. Through a team’s quiet loss of confidence that is felt before it is documented.</p><p>Agents improve system-visible coordination. They do not automatically capture unreconciled reality.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/720/1*X9tCYxXQsmhoMWosjrmbgw.jpeg" /><figcaption>Generated with NotebookLM</figcaption></figure><h3>What that changes for leaders</h3><p>The throughput gains arrive first. Anthropic’s research shows AI-assisted teams often attempt significantly more work, not just move faster. DORA’s 2025 research points in the same direction: flow metrics improve first.</p><p>What follows is less visible, and it arrives later. The prompts that surfaced judgment disappear with the friction. The dependency that got named because someone was sequencing two teams before a planning session. The blocker that emerged because a delivery lead reviewed the board and noticed something misaligned. These do not get automated. They get missed. The coordination legibility that remains is tool-generated: agent-written summaries, auto-populated fields, PR descriptions produced by the agent that wrote the code. It looks like legibility. It is not the same.</p><p>Many organizations are introducing agent participation into planning and execution flows faster than they are explicitly redesigning who owns the judgment that sits around those flows. The assumption that someone is doing the noticing persists.</p><p>Over the next six to twelve months, delivery organizations that have not named who holds the judgment work agents are not doing will accumulate coordination debt that is invisible until it is not. Dependencies untracked. Risks unnamed. Blockers surfaced too late. Throughput and flow metrics will look healthy. Rework, change failure rate, and downstream surprise will tell a different story. DORA’s data is clear: AI amplifies the existing delivery system. If the judgment infrastructure was part of that system, its absence gets amplified too.</p><p>Anthropic’s research on how developers work with coding agents makes the pattern concrete: engineers use AI in roughly 60% of their work but can fully delegate only 0–20% of tasks. Full delegation is limited to work that is easily verifiable. What is true for individual engineers holds for delivery systems: the judgment that cannot be easily verified cannot be safely delegated.</p><blockquote>Full delegation is limited to work that is easily verifiable. The judgment that cannot be easily verified cannot be safely delegated.</blockquote><p>Executive expectations of faster delivery, reasonable given the genuine throughput gains, compound this. The tool’s performance is treated as permission to reduce the overhead around it. Not all of that overhead was coordination. Some of it was the mechanism.</p><h3>What to watch now</h3><p>Three diagnostics, not a governance checklist.</p><h4>1. Where does ambiguity become shared understanding?</h4><p>Map the moments, not the tools. Where does uncertainty get surfaced, tested, and made visible enough to manage? Which of those moments are now agent-mediated? Which still depend on human judgment? Where is judgment assumed to be happening but no longer actively exercised?</p><h4>2. What are your throughput metrics hiding?</h4><p>Throughput is a leading indicator in an agentic delivery system. Rework, change failure rate, and downstream quality are lagging ones. If flow looks strong while downstream surprises are rising, the DORA amplifier dynamic is already in play.</p><h4>3. Who is noticing what the system does not show?</h4><p>Someone still needs the mandate and context to interpret signals that have not yet reached the formal workflow. Stakeholder hesitation. Executive appetite shifting. A concern raised informally that has not yet become a ticket. That boundary between system-visible work and lived organizational reality needs an owner. It cannot stay assumed.</p><h3>What needs explicit ownership</h3><p>The question is not whether to use agents in your delivery workflow. That decision has already been made, in most cases by configuration rather than by explicit organizational conversation. The more important question is whether your organization has a named answer to this: who is responsible for the boundary between what the system shows and what the organization is actually experiencing?</p><p>Most organizations do not have a named answer. That is not a failure. It is the normal pace of tooling change: workflow design decisions embedded in configuration before anyone framed them as design decisions. The task now is to make the inherited model explicit, before the gaps in it become visible as problems.</p><p>Name the person or function that holds the judgment boundary. Not a governance title: a working accountability. Who is responsible for surfacing the dependency that did not make it into Jira, the stakeholder hesitation that has not been formalized, the concern that is being carried informally from one meeting to the next? If you cannot answer that in ten seconds, the role is assumed but unowned. That gap will not appear in your throughput metrics. It will appear in your coordination debt.</p><h3>What agents cannot hold</h3><p>Throughput and legibility are not the same thing.</p><p>Tool-generated summaries, agent-written PR descriptions, and automated planning updates create the appearance of coordination. The information is there. The structure is there. What is not guaranteed is that the right things are visible: that the dependency nobody named, the risk nobody surfaced, the stakeholder concern that never made it into Jira, the direction that changed in a conversation before it changed in the system, are being noticed by someone with the context to act on them.</p><p>The biggest risk is not that delivery becomes less legible. It is that organizations continue believing they are well-coordinated right up until they are not. The delivery leaders who understand that distinction, between the appearance of coordination and the act of it, are the ones who will know what to watch.</p><h3>Internal Share: Why AI Updates Aren’t Shared Understanding</h3><p><strong>TL;DR for Slack/Teams:</strong></p><p><em>AI agents are removing the friction from our workflow, but they’re also removing the recurring moments where human judgment used to be structurally required. We are at risk of accumulating </em><strong><em>Coordination Debt</em></strong><em> — where our tools look perfectly updated, but our teams are no longer aligned on the reality of the work.</em></p><blockquote><strong><em>Key Takeaway:</em></strong><em> </em><strong><em>Throughput and legibility are not the same thing.</em></strong><em> We must explicitly name who owns the ‘Noticing’ work that exists outside of our Jira tickets and Slack summaries.</em></blockquote><blockquote><strong><em>Read the full breakdown here:</em></strong><em> [</em><a href="https://medium.com/@michael-hutson/when-the-noticing-gets-automated-5545c2a32929?source=friends_link&amp;sk=a57fe2e02db0ac4820d6a8da896dc3c3">https://medium.com/@michael-hutson/when-the-noticing-gets-automated-5545c2a32929?source=friends_link&amp;sk=a57fe2e02db0ac4820d6a8da896dc3c3</a><em>]</em></blockquote><p><em>In the Flow of Work publishes weekly. If this issue was useful, share it with a delivery leader who is navigating the same questions.</em></p><p><em>Michael Hutson is a delivery and Agile practitioner at Beacon Hill Enterprise.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=5545c2a32929" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[AI Is Killing Some Versions of Agile, Not Agile Itself]]></title>
            <link>https://medium.com/@michael-hutson/ai-is-killing-some-versions-of-agile-not-agile-itself-4a5681776d78?source=rss-d5a7ba3bf4fb------2</link>
            <guid isPermaLink="false">https://medium.com/p/4a5681776d78</guid>
            <category><![CDATA[agile]]></category>
            <category><![CDATA[leadership]]></category>
            <category><![CDATA[software-development]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[future-of-work]]></category>
            <dc:creator><![CDATA[Michael Hutson]]></dc:creator>
            <pubDate>Fri, 03 Apr 2026 20:56:01 GMT</pubDate>
            <atom:updated>2026-04-17T20:45:25.764Z</atom:updated>
            <content:encoded><![CDATA[<h4><em>Agility existed before the manifesto. Every era has reshaped it. This one may go deeper.</em></h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*8H3hOKpjRiCMuJv2" /><figcaption>Photo by <a href="https://unsplash.com/@indiratjokorda?utm_source=medium&amp;utm_medium=referral">Indira Tjokorda</a> on <a href="https://unsplash.com/?utm_source=medium&amp;utm_medium=referral">Unsplash</a></figcaption></figure><p><em>If Medium puts up the paywall, here’s the </em><a href="https://medium.com/@michael-hutson/ai-is-killing-some-versions-of-agile-not-agile-itself-4a5681776d78?source=friends_link&amp;sk=61d9f38de6693bfa404440fbb34c8f54"><em>friend link</em></a><em>. Let’s skip the velvet rope.</em></p><blockquote>The death of Agile has been announced so many times that the pattern matters more than the claim.</blockquote><p>Every few years, the case gets rebuilt using the language of the moment. First it was process failure. Then bureaucracy. Then scaling. Then ceremony without substance. Now it is AI. The trigger changes. The frustration underneath it does not. A new environment emerges, the inherited version of Agile starts to strain, and people mistake that strain for proof that Agile itself has run out of road.</p><p>That is why the history matters now. AI is not the first force to make inherited versions of Agile look strained. It is simply the latest, and perhaps the most revealing. Once you see how often Agile has been reshaped by the environment around it, the current moment looks less like an ending and more like the next reinterpretation.</p><p>That does not mean AI is just another routine turn in Agile’s history. It may prove more discontinuous than earlier shifts because it changes not only how work is organized, but the cost, speed, and authorship of the work itself.</p><p>When ChatGPT moved into the mainstream and large language models became part of everyday business language, the usual “Agile is dead” commentary followed close behind. I noticed it, but I was more interested in watching what would actually change than reacting to the first wave of noise.</p><p>What became visible over time was not the end of Agile, but a change in the conditions around it. AI was beginning to reshape how work is coordinated, how learning happens, how judgment gets exercised, and how visible the flow of work remains to the people responsible for steering it. That is the lens I have been trying to name.</p><p>It is also possible for Agile’s core challenge to remain relevant even as some of the roles, rituals, and industry structures built around Agile change dramatically or decline. That distinction matters in this moment.</p><h3>Before 2001, agility already existed. It just did not share a single name.</h3><p>Different software communities were pushing against the same frustrations: too much planning, too little learning; too many handoffs, too little collaboration; too much certainty assumed upfront in work that was never that certain to begin with. The dominant model still favored predict and control. Requirements were expected to stabilize early. Work moved through phases. Documentation stood in for understanding.</p><p>But underneath that model, teams were already working differently. They were shortening feedback loops, building in increments, staying closer to customers, and learning through delivery instead of trying to think everything through before the work began. They did not all use the same language. They did not all follow the same method. But they were reacting to the same reality: important work changed while it was being done.</p><p>The Agile Manifesto did not invent that instinct. It recognized it. It gave scattered practices a common frame. The manifesto was not the beginning of adaptation. It was the moment adaptation became legible.</p><h4><strong>How Agile Kept Changing</strong></h4><p>From 2001 to 2005, the manifesto gave emerging practices a shared language. What had been scattered across methods and communities now had a recognizable philosophy behind it. Agile became a credible alternative to heavyweight delivery, even if it was still mostly contained within software teams.</p><p>From 2006 to 2010, Agile became easier to adopt and easier to reduce. Scrum made it operationally portable. Sprints, standups, retrospectives, reviews, Product Owners, Scrum Masters, and backlogs gave organizations a structure they could teach and replicate. But what made Agile portable also made it easier to perform. Many teams slowly shifted from asking what the values demanded in context to asking whether they were doing Agile correctly.</p><p>That pattern had a texture most practitioners felt before they could name it.</p><p>I lived through that shift. Colocation was almost a requirement — physical boards anchored to shared office spaces, business and technology in the same room for the first time. When it worked, it worked because the convergence was real. Developers, testers, and business stakeholders shared the same context, felt the same pressure, and made decisions together. I remember standing in front of a task board with a product owner who had never seen work represented that way before — watching her move a card from in-progress to done was the first time she felt connected to what the team was actually doing. But proximity alone didn’t create that. Some of the most siloed teams I saw were also the most colocated. And as the method spread, the ceremonies started scaling past their useful size. Standups that once took ten minutes stretched to thirty. The ritual survived. The signal inside it didn’t.</p><blockquote>The ritual survived. The signal inside it didn’t.</blockquote><p>From 2011 to 2015, Agile stopped being just a team method and became an organizational ambition. Once leaders saw what smaller teams could achieve, they wanted to scale it across programs, portfolios, and departments. That is where the tension sharpened. Agile had emerged as a challenge to heavyweight delivery, but now it had to live alongside budgeting cycles, governance forums, reporting structures, and executive oversight. Its influence widened. So did the gap between what it meant in principle and how it was often applied in practice.</p><p>That gap showed up differently depending on where you sat in the organization.</p><p>This was the era of scaling frameworks, certification explosions, and enterprise tooling spreading across program layers. And to be fair, those frameworks were solving a real problem. Coordinating multiple Agile teams without collapsing into chaos required something. Scrum of scrums, release trains, PI planning — these were genuine attempts at a genuine challenge. But the solution introduced a coordination layer that started generating its own overhead. Dependency boards, governance forums, quarterly planning cycles that looked more like waterfall than anything the manifesto had imagined. For many organizations, the framework became the point. Agile had arrived in the enterprise. It had also started to calcify there.</p><blockquote>Agile had arrived in the enterprise. It had also started to calcify there.</blockquote><p>From 2016 to 2020, the limits of ritual-heavy Agile became harder to ignore. Teams could run the ceremonies and still struggle to move value cleanly through the system. The deeper question was no longer whether a team was doing Agile. It was whether the broader delivery system was helping work learn, move, and improve. Product thinking, DevOps, CI/CD, and faster release pipelines shifted the center of gravity from ceremony to flow.</p><p>The engineering progress was real. The organizational lag was just as real.</p><p>Development maturity genuinely advanced in this period. DevOps pipelines, CI/CD, faster release cadences — the engineering discipline finally caught up to what Agile had been asking for. Team Topologies gave the product operating model a vocabulary that resonated. But most organizations were still being measured against project milestones. Leaders were adopting the language of products and value streams while their governance structures still rewarded delivery against a fixed scope and date. That gap — between the model people were talking about and the model they were actually being held to — defined this era as much as the technical progress did.</p><p>Seen together, those phases tell a consistent story. The manifesto stayed relatively stable. The interpretation did not. Every era preserved part of Agile and distorted part of Agile. Every era pulled the values through its own management assumptions, technical constraints, and organizational needs.</p><p>Then the floor moved.</p><p>The AI era arrived inside the work itself, not as a methodology. I watched estimation break down in real time — a team spent thirty minutes debating whether a feature was five or eight story points before an AI built it in two minutes, followed by three days of review, validation, and integration work the Fibonacci sequence had no number for. Ceremonies designed for human-paced work started feeling misaligned when AI could produce a draft user story, test cases, and a code scaffold before the refinement session had started. It became whether the process was still describing the work. And underneath that, something harder to name: decisions were forming before conversations happened. Accountability was moving, and it wasn’t always clear where it had gone.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*-pwNw44a6cVFmTkGrGpcMQ.png" /><figcaption>Generated with NotebookLM</figcaption></figure><h4><strong>AI Changes the Conditions Again</strong></h4><p>AI did not arrive as a clean new methodology. It arrived inside the work itself. In code generation. In testing. In summarization. In research. In planning support. In customer insight. In workflow automation. It began changing not just how fast work could move, but how much of that movement remained visible to the people responsible for guiding it.</p><p>That is why the current “Agile is dead” conversation feels familiar. Once again, the environment changed. Once again, inherited practices started to strain. And once again, many people took that strain as proof that Agile itself had become irrelevant.</p><p>But AI is not simply making teams faster. It is changing the conditions under which Agile’s values have to operate.</p><p>In practice, that shows up in specific ways. Code can now be generated faster than review capacity scales with it. Requirements can be summarized, decomposed, and reframed before human negotiation has really happened. Work can appear to move faster while becoming harder to inspect, explain, or challenge at the points that matter most.</p><p>A team can now operate with a very different division of labor. Humans increasingly define intent, constraints, architecture, and acceptance boundaries. Agents increasingly execute against them. An engineer sets the plan, and agents produce large portions of the implementation. That is not just faster delivery. It is a different model of work.</p><p>The implications run across every role. Quality work shifts from manually authoring every test case toward defining quality boundaries and letting agents generate, run, and adapt tests from the plan itself. Product Owners and product leads use AI to synthesize customer signals, shape candidate priorities, and influence what enters the future product. And if human judgment, technical coherence, and accountability do not keep pace, teams risk becoming faster at producing output than at understanding its implications.</p><p>Working software is easier to generate, but not necessarily easier to understand. Customer feedback can be synthesized at scale, but that does not automatically produce better judgment. Responding to change can become more predictive and more automated, but also more opaque. And individuals and interactions do not become less important just because more work passes through tools, prompts, and agents. In many cases, they become more important, because the shared understanding that used to emerge through visible human exchanges can now thin out without anyone noticing right away.</p><p>That is the real pressure AI places on Agile. It is not exposing whether teams can move quickly. It is exposing whether organizations still know how learning, coordination, trust, and accountability are supposed to work when more of the workflow is shaped by machines.</p><p>This also raises the bar on technical excellence, not lowers it. If AI makes generation easier, then architecture, testing discipline, review quality, observability, and engineering judgment matter even more. The faster a team can produce code, stories, or test assets, the more important it becomes to know whether the system is coherent, maintainable, and safe. AI does not remove the need for craft. It makes weak craft easier to hide until the consequences compound.</p><blockquote>AI does not remove the need for craft. It makes weak craft easier to hide until the consequences compound.</blockquote><p>This transition is also being shaped by economics. As generation gets cheaper and faster, organizations will question which forms of human work remain differentiating, which become supervisory, and which are reduced or redesigned altogether.</p><p>“If Agile ceremonies are straining, it’s often because the underlying environment hasn’t adapted. I’ve broken down how to design that environment in How to Build a Learning Climate for the AI Era</p><p><a href="https://medium.com/@michael-hutson/how-to-build-a-learning-climate-for-the-ai-era-7d90d3c607ec">How to Build a Learning Climate for the AI Era</a></p><h4><strong>What Survives</strong></h4><p>If that is true, then the important question is not whether AI kills Agile.</p><p>It is what remains essential when another layer of adaptation becomes necessary.</p><p>The answer is not the ceremonies. Not the branding. Not the vocabulary. Not even the specific methods that happened to dominate one phase of the story.</p><p><strong>What survives is the underlying challenge Agile was always trying to answer: how do people work well when uncertainty is real, learning matters, and the path cannot be fully known in advance?</strong></p><p>That question existed before the manifesto. It existed through every reinterpretation that followed. And it exists now, perhaps more sharply than before, because AI is making it possible to produce more output with less visible understanding.</p><p>What survives is not only a way of building software. It is an ongoing negotiation over the human role inside software development itself.</p><p>What does it mean to be a software engineer when code generation becomes commonplace?</p><p>What does it mean to be a tester when validation is increasingly automated?</p><p>What does it mean to be a Product Owner, a Scrum Master, or any equivalent role when coordination, synthesis, and parts of decision support are increasingly handled by intelligent systems?</p><p>Those are not side questions. They are part of the next adaptation.</p><p>Some of what AI is now pressuring may not be essential Agile at all. It may be accumulated coordination overhead, role inflation, or ritualized compensation for weak systems. If those layers recede, that may not mark the death of Agile. It may mark the stripping away of what Agile learned to carry.</p><p>The future is not simply about whether humans stay in the loop. It is about what kind of loop they are in. Where judgment sits. Where accountability sits. Where learning sits. And how teams adapt together as those boundaries move.</p><p>The next phase of Agile will not be announced all at once. It will emerge through practice, just as earlier phases did. But it will not emerge in a vacuum. It will be shaped by who gets to define the new boundaries between automation, judgment, accountability, and cost.</p><h4><strong>What We Carry Forward, and What We Don’t</strong></h4><p>AI may kill some inherited versions of Agile. The ceremony-heavy ones. The reporting-heavy ones. The ones that treated process as a substitute for judgment.</p><p>But that is not the same as killing Agile.</p><p>If anything, this moment is forcing a harder question: what was essential, and what were we simply carrying forward because it once made sense under different conditions?</p><p>That was always the deeper challenge. And it is still the challenge now.</p><p>AI may kill some versions of Agile.</p><p>Not Agile itself.</p><h3>Internal Share: Is AI Killing Agile or Just the Ceremony?</h3><p><strong>TL;DR for Slack/Teams:</strong></p><blockquote><em>“AI isn’t killing Agile; it’s killing the version of Agile that relies on human-paced ceremonies to mask a lack of technical craft. As AI automates generation, our biggest risk is producing output faster than our shared understanding can keep up.”</em></blockquote><blockquote><strong><em>Key Takeaway:</em></strong><em> </em><strong><em>AI does not remove the need for craft; it makes weak craft easier to hide.</em></strong><em> We must shift our focus from ‘performing’ rituals to hardening our architecture and engineering judgment.</em></blockquote><blockquote><strong><em>Read the full breakdown here:</em></strong><em> [</em><a href="https://medium.com/@michael-hutson/ai-is-killing-some-versions-of-agile-not-agile-itself-4a5681776d78?source=friends_link&amp;sk=61d9f38de6693bfa404440fbb34c8f54">https://medium.com/@michael-hutson/ai-is-killing-some-versions-of-agile-not-agile-itself-4a5681776d78?source=friends_link&amp;sk=61d9f38de6693bfa404440fbb34c8f54</a><em>]</em></blockquote><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=4a5681776d78" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Agentic AI Will Make Delivery Harder To Read Before It Makes It Easier To Govern]]></title>
            <link>https://medium.com/@michael-hutson/agentic-ai-will-make-delivery-harder-to-read-before-it-makes-it-easier-to-govern-378c7ac2a740?source=rss-d5a7ba3bf4fb------2</link>
            <guid isPermaLink="false">https://medium.com/p/378c7ac2a740</guid>
            <category><![CDATA[agile]]></category>
            <category><![CDATA[digital-transformation]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[engineering-mangement]]></category>
            <category><![CDATA[leadership]]></category>
            <dc:creator><![CDATA[Michael Hutson]]></dc:creator>
            <pubDate>Thu, 26 Mar 2026 14:08:03 GMT</pubDate>
            <atom:updated>2026-04-17T20:53:30.478Z</atom:updated>
            <content:encoded><![CDATA[<p><em>Workflow legibility, the ability to reconstruct how work moved, is weakening in most delivery organizations right now. Most aren’t measuring it yet.</em></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*3-_fttejtmi7U6-A" /><figcaption>Photo by <a href="https://unsplash.com/@efekurnaz?utm_source=medium&amp;utm_medium=referral">Efe Kurnaz</a> on <a href="https://unsplash.com/?utm_source=medium&amp;utm_medium=referral">Unsplash</a></figcaption></figure><p><em>If Medium puts up the paywall, here’s the </em><a href="https://medium.com/@michael-hutson/agentic-ai-will-make-delivery-harder-to-read-before-it-makes-it-easier-to-govern-378c7ac2a740?source=friends_link&amp;sk=665820f8494db8782c1205b8b1e5f7b6"><em>friend link</em></a><em>. Let’s skip the velvet rope.</em></p><h3>In the Flow of Work</h3><p><em>A briefing for delivery leaders navigating AI-driven change across Agile, flow, and execution.</em></p><p><strong>A practitioner’s take</strong></p><p>I started this series because I kept having the same conversation, with delivery leaders, Agile coaches, Program managers, who were watching AI change their organizations and weren’t finding anywhere to think it through honestly. Not capability announcements. Not productivity benchmarks. The harder questions: what happens to the people on these teams? Where does a delivery leader add value when the nature of the work is changing? What does good leadership look like when the technology is moving faster than the frameworks?</p><p>In the Flow of Work is my attempt to hold that space. A practitioner’s lens on what AI-driven change actually means for delivery, for individuals, for teams, and for the leaders accountable for both.</p><h3>Agentic AI will make delivery harder to read before it makes it easier to govern</h3><p><em>Workflow legibility, the ability to reconstruct how work moved, is weakening in most delivery organizations right now. Most aren’t measuring it yet.</em></p><p>The delivery risk most organizations are not measuring yet is not autonomy. It is legibility.</p><p>More AI-generated work is moving through delivery systems right now. More agents are acting across tools and environments. And the systems built to give leaders visibility into how work moves, planning conversations, ticket history, review ownership, named decision-makers, were designed for human-generated work.</p><p>They are starting to show the strain.</p><p>The industry keeps telling the story of agentic AI as a story of capability. Agents are getting better. Coding tools are getting faster. Control features are being added. All true.</p><p>But that is not the most useful lens for a delivery leader.</p><p>The more important shift is this: AI is starting to do meaningful work inside delivery flows, while the systems around it are being redesigned for identity, permissions, monitoring, evaluation, and control.</p><p><strong>Agents are being treated as governed actors.</strong> OpenAI’s Frontier, its new platform for deploying AI in enterprise workflows, frames AI coworkers as entities with their own identities, explicit permissions, and boundaries. Jellyfish, an engineering intelligence platform that tracks software delivery data across enterprise teams, shows 64% of companies in its study now generate a majority of their code with AI assistance, with top-quartile adopters producing 2x the pull request throughput of low adopters.</p><p><strong>Control is moving into the execution layer itself.</strong> GitHub is pushing secret scanning directly into AI coding flows before code reaches commit or pull request. OpenAI is also integrating Promptfoo, an open-source tool for testing and evaluating AI model outputs, into Frontier for security testing, evaluation, reporting, and traceability, while separately publishing how it monitors its own internal coding agents for misalignment in real deployments.</p><p><strong>Read one by one, these look like product and governance updates.</strong></p><p><strong>Read together, they point to a different operational shift.</strong></p><p>The first major delivery failure mode of agentic AI will not be bad outputs or runaway autonomy. It will be loss of workflow legibility, where more work gets done but leaders understand less clearly how it moved, where control sat, and why delivery risk changed.</p><p>Workflow legibility is the ability to reconstruct how work moved, where decisions were made, what AI touched, and who owned the control path.</p><p>For most delivery organizations, that used to come almost for free. Human-generated work was never perfectly traceable, but it was usually legible enough. Planning conversations, ticket history, design reviews, pull requests, Slack threads, and named decision-makers gave leaders enough context to understand how delivery actually happened.</p><p>That assumption weakens when meaningful parts of the work are generated, transformed, or executed by non-human actors.</p><h3>What changed</h3><p>Three recent signals matter because they reinforce the same structural change.</p><p><strong>Agents have moved from features to governed actors.</strong> Frontier is explicitly about AI coworkers operating within identities, permissions, and boundaries. Once an agent can act across tools and systems, “the team is using AI” stops being precise enough. You now have non-human actors inside the workflow with defined access and scope.</p><blockquote>As Addy Osmani, Engineering Lead at Google, put it: <em>“Accountability has to be structural, not just cultural. Someone technical must own that code path, full stop. That owner reviews, approves, and carries the pager.”</em></blockquote><p><strong>AI-assisted delivery is past the experimental stage.</strong> Jellyfish, an engineering intelligence platform that tracks software delivery data across enterprise teams, shows AI-assisted code is already a material part of delivery in many organizations, with top-quartile adopters producing 2x the pull request throughput of low adopters. The 2025 DORA State of AI-Assisted Software Development report sharpens the picture: AI coding assistants boost individual output dramatically, 21% more tasks completed, 98% more pull requests merged, but organizational delivery metrics stay flat. More change is moving faster. That does not make the flow easier to understand, govern, or compare. It makes it harder.</p><p><strong>Governance is moving into the execution path itself.</strong> GitHub’s MCP server now scans for exposed secrets before code reaches commit or pull request. OpenAI is moving Promptfoo’s testing and evaluation capabilities into Frontier. OpenAI’s internal monitoring work shows the same pattern from another angle: once coding agents operate in real environments, behavior observation becomes part of the operating model, not just a policy appendix.</p><p>This is why the usual reading is too shallow. The point is not just that governance is getting more serious. The point is that delivery is becoming harder to interpret while it is happening.</p><p>That is a different management problem</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/900/1*OCXxOvYsyDhUbVossD0gCA.jpeg" /><figcaption>image generated with Claude</figcaption></figure><h3>What it means</h3><p>Most organizations will respond to this with familiar moves: policy, oversight, platform guardrails, maybe a steering group.</p><p>That will not be enough if the underlying problem is that delivery is becoming harder to see clearly as it moves.</p><p>The issue is not simply that AI needs governance. The issue is that delivery systems built around human-readable work are starting to absorb non-human actors that increase output while weakening visibility into how the system actually operates.</p><p>That changes several things at once, and they compound.</p><p>First, review capacity becomes a pacing constraint. When AI-assisted output rises, the review burden rises with it. Someone has to own that. Right now, in most organizations, no one explicitly does.</p><p>Second, decision rights blur. Once agents operate with permissions across tools and environments, the question of who approved what, and who can reverse it, stops having a clean answer across delivery, engineering, platform, and security.</p><p>Third, and this is the one most organizations are not yet treating seriously: shared metrics start comparing unlike systems. Two teams can post similar numbers while operating with fundamentally different levels of AI assistance, review rigor, and traceability. The portfolio view looks coherent. The underlying reality is not.</p><p>A delivery dashboard can keep looking clean while the underlying work systems diverge. One team may be using AI heavily in generation, triage, and workflow execution. Another may not. One team may have strong review ownership and explicit exception paths. Another may still be relying on informal judgment. If both still roll up into the same delivery metrics, the portfolio view gets noisier before anyone can prove something is broken.</p><p>That is where the institutional risk starts to bite.</p><p>You may still be reporting confidence while your system is becoming less explainable to the people accountable for it. Forecasts can still look stable. Throughput can still look healthy. Status can still sound under control. But the causal path underneath those signals is getting harder to reconstruct.</p><p>That makes diagnosis slower. Escalation murkier. Comparison weaker. Governance more performative than real.</p><p>Loss of workflow legibility rarely announces itself as failure on day one. It arrives first as a weaker ability to explain why a team looks fast, why a release feels riskier than expected, or why a problem took longer to isolate than it should have.</p><p>The consequence is not coming. It is arriving. Delivery organizations that do not redesign visibility, approval, and exception handling around AI-generated and agent-executed work will find their forecasts noisier, their escalations slower, and their cross-team metrics less trustworthy, not as a future risk but as a current condition that is quietly worsening.</p><blockquote><em>When did you last ask not whether your delivery metrics look healthy, but whether the people reading them could still explain what is underneath them?</em></blockquote><p>The industry keeps asking whether AI will replace people.</p><p>That is the wrong question for delivery leaders right now.</p><p>The right question is whether your function can still see clearly enough to govern what is already happening.</p><h3>What delivery leaders should watch</h3><p><strong>Watch for review load separating from delivery ownership.</strong> If AI-assisted output rises but no one clearly owns the review burden, the flow is speeding up while accountability is thinning out.</p><p><strong>Watch for agent permissions expanding faster than escalation paths.</strong> Once agents have identities, permissions, and boundaries, someone has to own approvals, reversals, and exceptions. If delivery, engineering, platform, and security each assume another group has that covered, decision-right confusion is already forming.</p><p><strong>Watch for shared metrics hiding different work systems.</strong> If teams have different levels of AI assistance, different review rigor, and different traceability standards, common roll-up metrics become less trustworthy. Cycle time, throughput, and delivery confidence can start comparing unlike systems.</p><p><strong>Watch for traceability becoming necessary to explain normal work.</strong> Observability is often framed as safety or compliance. For delivery leaders, it is increasingly a coordination mechanism. If you cannot reconstruct how work moved, why an action was taken, or where an agent intervened, diagnosis gets harder long before a major failure occurs.</p><h3>Practical takeaway</h3><p><strong>Treat workflow legibility as a delivery control, not a reporting nice-to-have.</strong></p><p>That means redesigning three things before the system forces the issue.</p><ol><li>Who owns the review when AI-assisted output rises?</li><li>Who owns approval, reversal, and exception handling when agents act across tools and environments?</li><li>How are teams compared when they are no longer operating under the same delivery conditions?</li></ol><p>Do not start only with “where can AI help us move faster?” Start with “where could AI-generated or agent-executed work become harder to trace, approve, compare, or reverse?” That is the better leadership question because it gets closer to how delivery risk will actually appear.</p><h3>Closing thought</h3><p>The next AI delivery question is not whether the system can do more.</p><p>It is whether your function can still explain how the work moved, who owned the decision path, and where the risk accumulated once AI entered the flow.</p><p>More output is coming. The question is whether visibility travels with it.</p><h3>Internal Share: The Legibility Crisis in Agentic Delivery</h3><h4>TL;DR for Slack/Teams</h4><blockquote><em>“We are gaining delivery speed but losing </em><strong><em>Workflow Legibility</em></strong><em>. As agents act across our tools, it’s becoming harder to reconstruct how decisions were made and who actually owns the control path.”</em></blockquote><blockquote><strong><em>Key Takeaway:</em></strong><em> </em><strong><em>Legibility is a delivery control, not a reporting nice-to-have.</em></strong><em> We must redesign our review and approval processes to ensure visibility travels at the same speed as our new AI output.</em></blockquote><blockquote><strong><em>Read the full breakdown here:</em></strong><em> [</em><a href="https://medium.com/@michael-hutson/agentic-ai-will-make-delivery-harder-to-read-before-it-makes-it-easier-to-govern-378c7ac2a740?source=friends_link&amp;sk=665820f8494db8782c1205b8b1e5f7b6">https://medium.com/@michael-hutson/agentic-ai-will-make-delivery-harder-to-read-before-it-makes-it-easier-to-govern-378c7ac2a740?source=friends_link&amp;sk=665820f8494db8782c1205b8b1e5f7b6</a><em>]</em></blockquote><h4>About this series</h4><p><em>I use </em><strong><em>&quot;In the Flow of Work&quot;</em></strong><em> as a way to learn through the pace of AI change in the public. AI has been useful for gathering signals, spotting patterns, and helping me stay close to what is shifting across delivery, Agile, and software work.</em></p><p><em>But the harder part is still interpretation, understanding what those changes actually mean for how work is organized, governed, and executed. This series is my way of doing that thinking out loud and sharing it with others navigating the same shift.</em></p><p><em>By delivery leaders I mean the people accountable for how work moves, programme managers, engineering managers, Agile coaches, transformation leads, product operations leaders, and senior managers responsible for execution at scale.</em></p><h3>Sources</h3><p><strong>OpenAI</strong>, <em>Introducing OpenAI Frontier</em>, useful for showing how leading vendors are framing agents as governed actors with identities, permissions, and boundaries.</p><p><strong>Jellyfish</strong>, <em>Jellyfish Reveals AI’s Real Impact on Engineering Teams</em>, adds real-world evidence that AI-assisted code generation is already affecting software delivery throughput.</p><p><strong>GitHub</strong>, <em>Secret scanning in AI coding agents via the GitHub MCP Server</em>, shows governance moving upstream into the execution path itself, not just sitting in policy or post hoc review.</p><p><strong>OpenAI</strong>, <em>OpenAI to acquire Promptfoo</em>, supports the point that evaluation, testing, reporting, and traceability are becoming part of the operating layer for agents.</p><p><strong>OpenAI</strong>, <em>How we monitor internal coding agents for misalignment</em>, grounds the argument that observability is becoming part of normal operational management for agentic systems.</p><p><strong>DORA</strong>, <em>2025 State of AI-Assisted Software Development</em>, provides empirical data on the gap between individual AI output gains and flat organizational delivery metrics.</p><p><strong>Addy Osmani</strong>, <em>Code Review in the Age of AI</em>, grounds the review ownership argument with a practitioner perspective on structural accountability.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=378c7ac2a740" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Your Team Is Getting More Done. Your Organization Is Losing the Thread]]></title>
            <link>https://medium.com/@michael-hutson/your-team-is-getting-more-done-your-organization-is-losing-the-thread-6ce3e3ca0671?source=rss-d5a7ba3bf4fb------2</link>
            <guid isPermaLink="false">https://medium.com/p/6ce3e3ca0671</guid>
            <category><![CDATA[knowledge-management]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[leadership]]></category>
            <category><![CDATA[future-of-work]]></category>
            <category><![CDATA[organization-design]]></category>
            <dc:creator><![CDATA[Michael Hutson]]></dc:creator>
            <pubDate>Tue, 17 Mar 2026 12:01:01 GMT</pubDate>
            <atom:updated>2026-04-17T20:58:20.254Z</atom:updated>
            <content:encoded><![CDATA[<h4>What leaders can design for before AI-first behavior erodes the shared understanding their organizations depend on.</h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*tMdDD08yt0Sl5nvY" /><figcaption>Photo by <a href="https://unsplash.com/@silverkblack?utm_source=medium&amp;utm_medium=referral">Vitaly Gariev</a> on <a href="https://unsplash.com/?utm_source=medium&amp;utm_medium=referral">Unsplash</a></figcaption></figure><p><em>If Medium puts up the paywall, here’s the </em><a href="https://medium.com/@michael-hutson/your-team-is-getting-more-done-your-organization-is-losing-the-thread-6ce3e3ca0671"><em>friend link</em></a><em>. Let’s skip the velvet rope.</em></p><p>It was an ordinary standup. Mid-sprint, new capability in build. An engineer gave his update. A colleague offered to help. The engineer said he was fine — he’d dig into it himself first, maybe reach out if he got stuck.</p><p>We moved on.</p><p>Over the following weeks the outputs stayed solid. Deadlines were met. But the Slack threads that used to run deep — someone asking why something worked a certain way, a colleague answering with context that went three layers past the question — those were quieter.</p><p>What I was watching wasn’t a performance issue. It was a transfer failure. And the standup was the only place it surfaced — because it was the last place where a human question had to travel through a human to get answered.</p><p><em>“What I was watching wasn’t a performance issue. It was a transfer failure.”</em></p><h3>The Same Pattern, At Every Level</h3><p>We ran a monthly operational review. Level one support was offshore through a third-party vendor. Level two was internal. Level three was the platform ops team. The platform was relatively stable, so the meeting existed to maintain the connection between teams that didn’t share day-to-day visibility — pipeline items, incidents, what was building under the surface.</p><p>When AI tools became available, we used them. We databound all incidents and let the model surface what we couldn’t see reviewing one ticket at a time — priority trends, recurring patterns, customer sentiment signals embedded in the logs. The output was better by every measurable standard. Real-time. Comprehensive. The model gave us in minutes what the meeting had previously spent an hour reconstructing.</p><p>The meeting simplified. And a question formed in the back of my mind: did we still need it?</p><p>We kept it. Not because the prioritization work required it — the model handled that better than we ever had. We kept it because something in the meeting wasn’t in the data. Conversational patterns the L1 team only surfaced verbally. Marketplace shifts heard informally. Roadmap updates not yet in ServiceNow. Incidents outside our domain that someone flagged because they’d connected a dot we hadn’t.</p><p>None of that lived in the incident logs. It lived in the people — and the meeting was the mechanism that brought it into contact.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/720/1*v2GGpDR0CWjPgqNIvOFysA.png" /><figcaption>Generated with Claude AI</figcaption></figure><h3>What the Data Doesn’t Show</h3><p>Here is what AI does well. It makes volume cheap. Ask it a question and it answers faster, more comprehensively, more patiently than most colleagues. Documented processes, structured data, procedural knowledge — AI handles all of it better than the mechanisms it replaced.</p><p><strong>Volume is not the same as correctness</strong></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/720/1*XwTLjdV0QtTMK8Vi_k16rQ.png" /><figcaption>Generated with Claude AI</figcaption></figure><p>Organizational correctness is something different. Understanding why a decision matters given what happened two quarters ago. Knowing the downstream implication your team can’t see because it lives in a different function. Sensing the political constraint that isn’t in the brief. Hearing a challenge that AI would never raise because it lacks the lived context to know the challenge exists.</p><p><em>“The model gave us better data. It couldn’t give us what the meeting was actually producing.”</em></p><p>When a person asks a colleague rather than a model, they may get a slower answer. They also get something the output can’t contain — a better framing of the question, a dependency they hadn’t known existed, a perspective from a part of the organization they couldn’t see, and a relationship marginally stronger than before. None of that shows up in the output. All of it compounds over time.</p><p>Before AI, the friction of needing each other was unavoidable. You needed a colleague answering to move forward. You needed the support manager’s read to understand the queue. The interactions that produced shared understanding were not designed. They were compelled.</p><blockquote><em>The interactions that produced shared understanding were not designed. They were compelled.</em></blockquote><p>Not all of it was valuable. Some of those conversations were inefficient, politically charged, or simply wrong. The workaround nobody wrote down sometimes existed because the official process was broken and nobody fixed it. AI replacing that friction is a genuine improvement.</p><p>AI removes the need. Individuals can move forward alone — faster, with better answers, without the pause of not knowing who to ask. That is a genuine improvement at the individual level.</p><p>The byproduct disappears with the friction. In the delivery environments I’ve worked in, the pattern is consistent: when people stop needing to ask each other things, the informal exchanges that carried tacit knowledge stop happening. Shared context stops accumulating across teams. Organizational coherence — the thing that makes a group of capable individuals an organization rather than a collection of them — degrades. Each step in that chain is the result of a reasonable individual decision. None of it registers on a dashboard.</p><p>Some knowledge doesn’t exist anywhere in a system until a person puts it there. Not the documented process — AI retrieves that well. The workaround everyone uses and nobody has written down. The approach that failed eighteen months ago for reasons that didn’t make it into the post-mortem. The analyst who asks AI gets an answer. The analyst who asks the person who built the last three models gets the answer and the context. One produces correct output. The other produces organizationally correct output. At scale, those diverge in ways that don’t show up until something breaks in an unexpected direction.</p><p>And it is not only new employees. Experienced people, mid-tenure, are reaching for Copilot or ChatGPT for micro-problems that would have been short conversations twelve months ago. Each one not discussed is a transfer that doesn’t occur, a relationship that doesn’t strengthen, a piece of common context that doesn’t form.</p><h3>How We Got Here</h3><p>This didn’t begin with AI. The shift to remote work reduced synchronous exchange first. Conference calls, in-person conversations, the sidebar after the meeting — those declined before any model was deployed.</p><p>Organizations adapted and called it resilience.</p><p>AI arrived into that environment and provided something that looked like a solution. By end of 2026, Gartner projects 40% of enterprise applications will embed task-specific AI agents — up from less than 5% in 2025. [2] Instead of reaching out to a colleague you couldn’t easily find, you could ask the model and get an answer immediately. The reduction in synchronous engagement that had felt like a loss started to feel like an upgrade.</p><p>What changed wasn’t the amount of work getting done. What changed was the ambient human exchange that had been keeping organizational knowledge current and shared. And because no dashboard measures that exchange — because its output is common context rather than completed tasks — the loss registered nowhere.</p><p>The organization looked fully staffed and capable. The shared understanding was thinning anyway.</p><h3>What Leaders Can Do About It</h3><p>The first three are grounded in what the research on knowledge transfer actually supports. The fourth addresses tooling. The fifth is the one that closes the gap between the diagnosis and the prescription.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/720/1*K17uoSzSlARceq5uCtu4Wg.png" /><figcaption>Generated with Claude AI</figcaption></figure><h4>Make mentorship a formal commitment</h4><p>Informal mentorship produces inconsistent results — and in an AI-augmented environment, it produces almost none. Make it a formal delivery commitment: a named senior person, a defined relationship, time protected in the sprint. Not a development perk assigned at onboarding and forgotten by week three. Without this, the unwritten playbook never has a channel to move through.</p><h4>Keep the informal time on the calendar</h4><p>The conditions where informal learning happens are exactly the conditions efficiency pressure eliminates first<strong>.</strong> Team lunches, casual debriefs, working alongside someone rather than just receiving their output — none of these appear on a dashboard, and no system will protect them. The organizations that let these rituals collapse discover the cost when someone with fifteen years of institutional memory hands in their notice.</p><h4>Surface the questions before they disappear</h4><p>If team members are routing questions to AI before they reach a senior colleague, the senior colleague never gets the signal that a gap exists. Run structured check-ins at thirty, sixty, and ninety days. The question isn’t “how’s it going” but “explain the why behind a decision you made this week.” Not the what — the why. The gap tends to become visible around day forty-five. If they can’t answer it, the output metrics have been hiding a capability gap.</p><h4>Change the default on your AI tools</h4><p>This is the intervention most leaders haven’t thought about yet and the one with the most leverage. When organizations deploy AI tools, the default routes every question to the model. Nobody decided employees should consult AI before colleagues for process decisions, architectural choices, or stakeholder context. It happened because that’s how the tool shipped. Leaders with authority over tooling decisions can change that default — flagging categories of questions that should prompt a colleague first. Some questions are better answered by the system that knows everything. Some are better answered by the person who knows your organization.</p><h4>Audit the rituals AI made look redundant</h4><p>When AI handles the stated job of a meeting better than the meeting did, the meeting feels optional. That’s the moment to ask what the second job was. The operational review didn’t survive because the model couldn’t prioritize incidents — it survived because someone noticed the prioritization wasn’t the only thing happening in the room. For every recurring ritual that has quietly simplified, moved to async, or stopped entirely since AI tools arrived, ask: what was it producing that nobody named? Shared awareness. Weak-signal exchange. The sense of how things are going across teams that don’t share a dashboard. That’s what needs protecting, not the meeting itself, but the job the meeting was doing that nobody put on the agenda.</p><h3>The CEO Building Frontier AI Spends 40% of His Time on This</h3><p>Dario Amodei, CEO of Anthropic, said on the Dwarkesh Podcast earlier this year that he spends a third to 40% of his time making sure the culture of Anthropic is good. His concern was specific: as the company grows, he has watched other AI labs develop what he called decoherence. Teams losing the thread of why the work matters.</p><p>His response was not a framework. It was his own time — biweekly all-hands, active Slack presence, direct communication he described as having value impossible to achieve passing things down six levels. [1]</p><p>He is doing for Anthropic’s culture what the operational review was doing for my support teams — maintaining a shared picture of the organization that no system will maintain on its own. Deliberately. Continuously. Because he knows that if he stops, it stops.</p><p>The CEO of one of the world’s leading AI labs has concluded that organizational coherence requires more direct human attention as AI takes on more of the work. Not less. More. His context is a fast-growing research organization where culture is the product. But the signal is worth taking seriously regardless of your industry: the person closest to what AI can do has decided that human connection is the thing it cannot replace. That is not sentiment. It is a diagnosis from the inside.</p><p><em>AI will keep getting better at what your organization has already written down. The organizations that pull ahead will protect what hasn’t been written down yet and design deliberately to keep it moving.</em></p><p><em>What is your organization’s shared picture of itself right now and how would you know if it was getting thinner?</em></p><h3>Internal Share: Why Efficiency is Hiding a Knowledge Crisis</h3><h4><strong>TL;DR for Slack/Teams:</strong></h4><blockquote><em>AI is making volume cheap, but it’s also removing the ‘productive friction’ of needing our colleagues to move forward. We are getting more done individually, but we are losing the ambient human exchange that keeps our organizational context alive.</em></blockquote><blockquote><strong><em>Key Takeaway: Volume is not the same as correctness.</em></strong><em> We must move from ‘compelled’ interactions to ‘designed’ ones to protect the tacit knowledge that no AI model can replicate.</em></blockquote><blockquote><strong><em>Read the full breakdown here: </em></strong>[<a href="https://medium.com/@michael-hutson/your-team-is-getting-more-done-your-organization-is-losing-the-thread-6ce3e3ca0671?source=friends_link&amp;sk=baa98a193337c2f5993694df10be676a">https://medium.com/@michael-hutson/your-team-is-getting-more-done-your-organization-is-losing-the-thread-6ce3e3ca0671?source=friends_link&amp;sk=baa98a193337c2f5993694df10be676a</a>]</blockquote><h4>Recommended Reading</h4><p>While organizational decoherence erodes our culture, it also creates a functional “coordination debt” that hides in our automated workflows. This piece explores the mechanical side of that loss.</p><p><a href="https://medium.com/@michael-hutson/when-the-noticing-gets-automated-5545c2a32929">The Coordination Trap: Why AI Updates Aren’t the Same as Shared Understanding</a></p><p><strong>Sources</strong></p><p>[1] Dario Amodei, Dwarkesh Podcast, published February 13, 2026. Direct quote: “I probably spend a third, maybe 40%, of my time making sure the culture of Anthropic is good.” Also reported by Fortune, “Anthropic’s Dario Amodei says he spends up to 40% of his time on company culture, not products” (February 26, 2026).</p><p>[2] Gartner, “Gartner Predicts Up to 40% of Enterprise Applications Will Include Conversational AI by End of 2026, Up from Less Than 5% in 2025” (August 26, 2025).</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=6ce3e3ca0671" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>