<?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 Stephanie Muxfeld on Medium]]></title>
        <description><![CDATA[Stories by Stephanie Muxfeld on Medium]]></description>
        <link>https://medium.com/@stephanie-muxfeld?source=rss-697b881c9471------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*A2Kb1H9mA5Wfu0bZ6YAq3Q.jpeg</url>
            <title>Stories by Stephanie Muxfeld on Medium</title>
            <link>https://medium.com/@stephanie-muxfeld?source=rss-697b881c9471------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Wed, 18 Nov 2026 17:54:40 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@stephanie-muxfeld/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[One Person Doing Three Jobs With One Title]]></title>
            <link>https://stephanie-muxfeld.medium.com/one-person-doing-three-jobs-with-one-title-b288076d0c48?source=rss-697b881c9471------2</link>
            <guid isPermaLink="false">https://medium.com/p/b288076d0c48</guid>
            <category><![CDATA[future-of-work]]></category>
            <category><![CDATA[product-strategy]]></category>
            <category><![CDATA[leadership]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[product-management]]></category>
            <dc:creator><![CDATA[Stephanie Muxfeld]]></dc:creator>
            <pubDate>Mon, 16 Nov 2026 12:31:01 GMT</pubDate>
            <atom:updated>2026-11-16T12:31:01.963Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*12ecsZqbyOicSgpqdFcG7g.png" /></figure><p><strong>Signal</strong></p><p>Product management has never had a tightly bounded job description. From the earliest lean startup playbooks, the title absorbed whatever the team around it couldn’t cover: a product manager (PM) who ran customer interviews because no researcher had been staffed, who sketched a flow because no designer was free that sprint, who wrote the business case because finance hadn’t been looped in yet. That elasticity was part of the job’s appeal, and part of its risk, long before anyone mentioned artificial intelligence (AI).</p><p>What has shifted is not the elasticity itself. It is the rate at which new capability becomes available for a single person to absorb, and the quiet way organizations have started to treat that capability as coverage rather than as a gap still waiting on a decision.</p><p>A product manager with an AI assistant can now produce a rough wireframe in minutes, work that used to require a design partner’s time, even a junior one. A product manager can turn twenty interview transcripts into a synthesis document that used to take a researcher most of a week, and have a workable draft before lunch. A product manager can write an engineering ticket detailed enough, because a tool helped think through the edge cases, that an engineer no longer asks for an architecture conversation before starting. Each of these looks, on its own, like a reasonable productivity gain. Stacked across one role, they add up to a single title doing what a PM, a researcher, and a junior designer used to split three ways.</p><p>Call this scope expansion by accretion. Nobody convenes a meeting to formally redesign the role to include research synthesis and interface design. The scope arrives task by task, each addition modest enough that no one names it as a change to the job. It gets accepted the way most incremental additions get accepted: the work got done, and questioning why one person did work that used to take three felt like second guessing a result nobody had complained about yet.</p><p>This incentive compounds under budget pressure. A hiring freeze rarely arrives with an explicit instruction to skip research or design. It arrives as an instruction to do more with the team already in place, and a PM with a capable AI tool is the person best positioned, in the moment, to look like they’re meeting that instruction without anyone having to admit a role went unfilled. The capability makes the coping mechanism look like a solution instead of a stopgap, which is exactly what makes it easy to leave in place quarter after quarter.</p><p>This is a close cousin of the pattern this arc opened with two weeks ago, a team that shrinks while the title stays fixed. The product version has its own signature, though. It is rarely a headcount cut that triggers it. More often it is a research or design role that goes unfilled after someone leaves, or a design system maturing to the point where a PM can self serve a few screens without waiting on a design review, or a hiring freeze that quietly becomes permanent. AI did not start this drift. Product scope has been migrating toward the PM for years, long before generative tools existed, because the role sits at the center of the org chart and absorbs ambiguity by default. What AI changed was the capability bottleneck that used to force a conversation. Before, a PM who wanted to sketch a flow without a designer ran into the limits of their own drawing ability, and that limit was itself a kind of forcing function, making the gap visible and usually prompting someone to ask whether design needed to be staffed on the project. Now the tool closes the gap quietly enough that the forcing function never fires.</p><p>The effect shows up differently depending on which function gets absorbed. When a PM absorbs research, the loss is usually rigor: AI-assisted synthesis is fast, but a trained researcher is listening for things a transcript summary doesn’t surface on its own, a contradiction between what a user says and what they do, a pattern across many sessions that’s invisible in any single one. When a PM absorbs design, the loss is usually consistency: a PM producing wireframes without a design partner’s eye on the system tends to solve the immediate screen well and the pattern library badly, one more one-off component in a product that already has too many. When a PM absorbs delivery coordination, from writing tickets to running the ceremonies a delivery lead used to own, the loss is usually time and attention for the part of the job that was always the point: deciding what to build and why.</p><p>None of these losses are visible on a roadmap. They show up later, in a release that needed a design review it never got, in a research finding that got missed because nobody was listening for the contradiction, in a backlog that drifted because the person writing tickets stopped having room to think about priority. The title stays the same the whole time. So does the org chart. The scope is the only thing that moved, and it moved without anyone deciding it should.</p><p><strong>Steph’s Perspective</strong></p><p>I worked with a product organization last year where one PM, the lead on a mid-sized consumer app, had quietly become the de facto owner of product research, interaction design, and half of delivery coordination for her squad. None of that was in her job description. None of it was a decision anyone made out loud. It built up over about eighteen months, one gap at a time: a researcher left and wasn’t replaced because the organization was watching headcount, a junior designer moved to a different team and the backfill kept slipping a quarter, and the delivery lead role got folded into engineering management after a reorganization that technically preserved every title on paper while changing what almost every person next to that title actually did day to day.</p><p>She was, by every account I heard from her peers and her own manager, doing all three jobs well. She ran interviews herself and fed the transcripts into a summarizer that gave her a synthesis document in an afternoon instead of the week it used to take a researcher. She built her own wireframes in Figma, good enough that engineering stopped asking for a formal design pass on anything that wasn’t customer facing in a new way. She wrote tickets detailed enough that the team rarely needed the clarifying conversation a delivery lead used to run. On a quarterly review, every one of her projects shipped roughly on time. The dashboard looked fine.</p><p>What took longer to surface, because it never shows up on a dashboard, was what she had stopped doing in order to do all of that. She had stopped running the kind of generative research that doesn’t start from a hypothesis, the open ended conversation that finds a problem nobody asked about yet, because there wasn’t time left after she’d synthesized the structured research she already had a backlog for. She had stopped questioning her own interface decisions, because a design partner’s job had always included telling her a pattern she’d built didn’t match the rest of the product, and nobody was doing that job anymore, including her, for her own work. And she had quietly stopped running the kind of planning conversation that surfaces risk early, the fifteen minute check in a delivery lead would have had with an engineer mid sprint, because she genuinely did not have a spare fifteen minutes most days.</p><p>None of this was her fault, and I want to be precise about that, because the easy read of a story like this blames the person absorbing the work for not saying no. She did say something, more than once, to her manager, in terms close to “I think I’m covering three roles here.” What she got back, both times, was some version of “you’re doing great, and we don’t have budget to backfill right now.” That is not a PM failing to advocate for herself. That is an organization treating one person’s capacity to cope as a substitute for a staffing decision it did not want to make.</p><p>Here is where the translate, don’t reinvent instinct actually earns its keep, and where most organizations get this wrong in the opposite direction. The mistake is not always refusing to acknowledge the expansion. Sometimes the mistake is swinging the other way, deciding that because AI changed what’s technically possible, the whole role needs to be reinvented from scratch: a new title, a new org chart, a wholesale rewrite of what a product manager is for. That overcorrection is its own kind of avoidance. It treats the problem as a structure question when the actual question is simpler and more uncomfortable: which specific function got absorbed, and does anyone with the authority to staff it actually know that it happened.</p><p>Translating the expansion, rather than reinventing around it, means doing something much less dramatic than redesigning a title. It means someone with budget authority sitting down and asking, function by function, what got absorbed, whether it was absorbed on purpose or by accretion, and whether the organization is comfortable with that as a permanent arrangement, or whether it was only ever meant to be a stopgap that quietly became load bearing. That conversation is less dramatic than a reorganization. It is also the one that actually protects the research rigor, the design consistency, or the delivery attention thinning out underneath a dashboard that still looks fine.</p><p><strong>Stride</strong></p><p>Before adding anything else to a product roadmap this quarter, take one function, research, design, or delivery coordination, and ask whether it is currently staffed on that team on purpose or quietly carried by the PM without anyone naming it. Pick whichever of the three functions is getting the least deliberate attention right now, not whichever feels most urgent to fix.</p><p>A rough way to spot which one that is: notice what a PM on the team has stopped bringing to a cross functional review in the last two quarters. If design reviews used to happen and now mostly don’t, design consistency is probably the function quietly absorbed and unattended. If research used to generate open ended findings and now mostly confirms what the roadmap already assumed, exploratory research is probably the one that has gone missing. If engineers used to ask clarifying questions mid sprint and now mostly don’t, delivery coordination has probably been folded into a role that no longer has time to run it.</p><p>Once that function is named, the move is not to immediately redesign the PM’s job or hire a replacement that budget may not support this quarter. The move is to have the conversation that has been skipped: the person doing the absorbed work tells their manager, specifically, what is being left out because there isn’t time for it, not in the vague terms of feeling stretched, but in the concrete terms of “exploratory research hasn’t happened since March” or “no one outside me has reviewed an interface pattern in two sprints.” That specificity is what turns a capacity complaint, which is easy for an organization to deprioritize, into a named gap in coverage, which is harder to wave off once it is written down.</p><p>This does not solve the staffing problem on its own. It does something narrower and more achievable: it makes the choice visible as a choice, rather than letting it continue to look like a role that simply expanded on its own.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=b288076d0c48" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[When the team shrinks and the title doesn’t]]></title>
            <link>https://stephanie-muxfeld.medium.com/when-the-team-shrinks-and-the-title-doesnt-2bddc349cc31?source=rss-697b881c9471------2</link>
            <guid isPermaLink="false">https://medium.com/p/2bddc349cc31</guid>
            <category><![CDATA[organizational-design]]></category>
            <category><![CDATA[product-leadership]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[future-of-work]]></category>
            <category><![CDATA[workplace]]></category>
            <dc:creator><![CDATA[Stephanie Muxfeld]]></dc:creator>
            <pubDate>Mon, 09 Nov 2026 12:31:01 GMT</pubDate>
            <atom:updated>2026-11-09T12:31:01.599Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*HOFhm2QkbZxK4YqpNDZNaw.png" /></figure><p><strong>Signal</strong></p><p>Ask an executive whether AI has reshaped their headcount this year and most will say no. Ask the people doing the work and a different picture shows up. A 2026 survey of 665 employed US adults and 354 executives, run by Omni Calculator on the Prolific research panel, found that about one in three workers say their team has gotten smaller over the past three years while their workload stayed the same or grew. Among mid-level employees the figure rises to a third. Among people working hybrid schedules it climbs to over a third. In tech and media specifically, it reaches 38 percent.</p><p>On the other side of the same question, only 10 percent of executives say they reduced headcount as a direct result of AI. Another 8 percent say they simply stopped backfilling open roles, a quieter version of the same outcome that doesn’t register as a cut in any announcement. Fifty-three percent say AI hasn’t affected their hiring decisions at all. Taken together, the worker data and the executive data are not two measurements of the same event. They read like descriptions of two different companies.</p><p>This gap is not new, and it did not start with the current wave of AI tools. The World Economic Forum’s Future of Jobs Report 2023 anticipated a version of it before generative AI reached most workplaces, projecting 23 percent structural churn across the labor market within five years and a contraction of 26 million jobs in clerical and administrative roles specifically, the layer of work that rarely gets its own headline when it disappears. That report also estimated that 44 percent of workers’ core skills would be disrupted over the same period, which is a reasonable description of what it feels like to watch a role’s actual content change out from under a title that stays exactly the same.</p><p>What ties the survey data and the labor market projections together is the same mechanism, work finding its way onto fewer people without the event ever getting named as a reduction. Nobody announces that three roles quietly became one and a half. A departure goes unfilled because the remaining team seems to be managing. A contractor’s engagement ends and nobody replaces it because the work seems to be getting absorbed. Each individual decision looks small and reasonable in isolation. The aggregate, measured a year or two later, looks like the 30 percent of workers in that survey who say their team shrank while their workload held steady or grew.</p><p>The reason this matters more now than it did five years ago is not that the mechanism is new. It is that AI tools give every one of those individual decisions a plausible justification in the moment it’s made. A role goes unfilled and the team “has AI now” to help absorb the gap, whether or not anyone has actually measured what the gap requires or what the tool actually covers. That justification doesn’t need to be tested before the next headcount conversation, because by the time anyone would test it, the team has already adjusted, quietly, the way teams do.</p><p>This is a structural story, not a performance story, and that distinction matters for how a leader responds to it. A team that has absorbed two departed roles into the people who remain is not a team performing worse. It is a team covering more ground with the same number of people, often successfully, right up until the ground gets uneven enough that something falls through. Treating the strain that shows up in that situation as an individual performance issue, someone seems stretched, someone seems slower than they used to be, misses what actually happened, which is that the structure around them changed and nobody updated the map.</p><p><strong>Steph’s perspective</strong></p><p>I watched this play out clearly inside an operations team I worked with earlier this year. Over about fourteen months, the team lost two people, one through attrition, one through a reorganization that moved a role to a different part of the company. Neither departure was framed as a reduction. Both were framed, accurately in the moment, as individual events with individual explanations. The team’s org chart, viewed at any single point during that stretch, looked stable.</p><p>Viewed across the full fourteen months, it was not stable at all. The work those two people had covered, a mix of vendor coordination, reporting, and a specific category of customer escalation, did not disappear when they left. It distributed itself across the four people who stayed, in pieces small enough that no single addition felt like the moment anything changed. A little more reporting here. A vendor call that used to route to someone else now routes to whoever is free. A category of escalation that used to have a dedicated owner now gets handled by whoever picks it up first.</p><p>By the time I was brought in, the team was functioning, in the narrow sense that the work was getting done and nobody had quit. But every person on it described some version of the same feeling, a sense that their job had changed without anyone deciding it should, and a related sense that it would be strange to bring this up, because nothing had technically happened to them. No one had been laid off. No new tasks had been formally assigned. The job just was not the job it used to be, and there was no event to point to as the reason.</p><p>What made this hard to name, for them and initially for me, is that every individual decision along the way had been reasonable. It made sense not to backfill the vendor coordination role immediately after that person left, because the team seemed to be handling it fine for the first few weeks. It made sense to route the escalation category to whoever was available rather than waiting to hire a dedicated owner, because escalations can’t wait on a hiring process. Each decision, examined on its own, was the obviously correct call given what was known at the time. The problem only existed in aggregate, and aggregates don’t show up in any single meeting where a single decision gets made.</p><p>The fix, once we named it, was not complicated, though it required something leadership had not done, which was to look at the team’s actual coverage the way you’d look at a staffing model rather than an org chart. We listed every function the team had covered at the start of the fourteen months and matched it against who currently owned each one. Two functions had no clear owner at all, they were being handled reactively by whoever had bandwidth that week. One function had been absorbed so completely into a single person’s role that losing that person would have meant losing the function entirely, with no documented process behind it. None of this was visible from the org chart. All of it was visible once someone asked the coverage question directly instead of the headcount question.</p><p>I think the executives in that 10 percent, who say they cut headcount specifically because of AI, are probably the most honest ones in the whole survey. The other 53 percent who say AI hasn’t affected their hiring at all may also be telling the truth about their own decisions, while missing the aggregate effect of a dozen small, individually reasonable choices not to backfill, not to replace, not to hire the contractor again this quarter. Structural change rarely announces itself as a single decision big enough to notice. It accumulates in the space between decisions that each looked fine on their own.</p><p><strong>Stride</strong></p><p>Pick one role on the team, including the reader’s own if it applies, and write down what it actually covered two years ago next to what it covers now. Not the job description. The real list: the vendor calls, the categories of work, the decisions that route through that role today that didn’t two years ago. Most people doing this exercise for the first time find at least one function that quietly migrated onto their plate with no formal moment attached to it.</p><p>Bring that comparison into the next planning or one-on-one conversation as a capacity question rather than a performance question. The framing matters. “Here’s what this role actually covers now compared to two years ago, and here’s where I think the coverage is thin” opens a structural conversation. “I’m struggling to keep up” invites a performance conversation about the wrong thing entirely.</p><p>If the exercise surfaces a function with no clear, documented owner, the kind that gets handled reactively by whoever has time, treat that as the finding worth escalating first, independent of whether the current arrangement happens to be working this month. An unowned function that is currently being absorbed successfully is still a structural risk. It just hasn’t been tested yet.</p><p><strong>Sources &amp; References</strong></p><p>Omni Calculator, “Ghost Downsizing” survey report, 2026, conducted on the Prolific research panel among 665 employed US adults and 354 executives. <a href="https://www.omnicalculator.com/reports/ghost-downsizing-ai-survey-2026">https://www.omnicalculator.com/reports/ghost-downsizing-ai-survey-2026</a></p><p>World Economic Forum, The Future of Jobs Report 2023. <a href="https://www.weforum.org/publications/the-future-of-jobs-report-2023/digest/">https://www.weforum.org/publications/the-future-of-jobs-report-2023/digest/</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=2bddc349cc31" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The judgment work AI cannot do]]></title>
            <link>https://stephanie-muxfeld.medium.com/the-judgment-work-ai-cannot-do-cf4ead25892b?source=rss-697b881c9471------2</link>
            <guid isPermaLink="false">https://medium.com/p/cf4ead25892b</guid>
            <category><![CDATA[leadership]]></category>
            <category><![CDATA[product-management]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[product-strategy]]></category>
            <category><![CDATA[critical-thinking]]></category>
            <dc:creator><![CDATA[Stephanie Muxfeld]]></dc:creator>
            <pubDate>Mon, 26 Oct 2026 11:31:04 GMT</pubDate>
            <atom:updated>2026-10-26T11:31:04.197Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*yLv-C8tBX8MenkWK_iIOnw.png" /></figure><h4>Signal</h4><p>Six weeks ago, this space opened with a distinction: confident output and correct output are not the same thing, and most AI tools are built to produce the first without any particular claim on the second. Since then, the Observatory has moved through structure, through ownership, through translation, through legibility, each week naming a different way that fluent output can crowd out a harder, slower kind of thinking. Read together, the five pillars that anchor this practice stop looking like five separate ideas and start looking like five faces of one underlying skill.</p><p>Call it judgment, and each earlier piece turns out to have been examining a different application of it. Calibration over confidence is judgment applied to certainty: knowing the difference between how sure a system sounds and how sure the evidence actually warrants. Structural not personal is judgment applied to cause: knowing whether a recurring problem sits in a person or in the system built around them. Ownership in the gaps AI creates is judgment applied to accountability: knowing that generating something and being answerable for it are two different acts, even as the gap between them keeps narrowing. Translate don’t reinvent is judgment applied to memory: knowing which part of an existing practice still earns its place before a new tool replaces it wholesale. Legible isn’t the same as true is judgment applied to evidence: knowing that a report reading clean is a claim about formatting, not a claim about accuracy.</p><p>Name the pattern and a second thing becomes visible. Every signal in this arc described a moment where well formatted, decisive sounding output let a team skip the step where someone actually decides whether to believe it. That step, the one where a person weighs evidence against context and consequence and arrives at a position they would defend out loud, is what judgment means here. And on the evidence this practice has gathered across dozens of product organizations, it is the one piece of the work that AI tools cannot do on a team’s behalf, however well they do everything that surrounds it.</p><p>This is not an argument that AI tools lack capability. It is a narrower claim, and a more useful one: a model can produce an answer, a summary, a recommendation, even a confident one, but it carries none of the consequence of being wrong the way a person who signs off does. It has no history with the team that will live with the decision, no memory of the last time a plausible sounding answer missed something the room would have caught, no stake in what happens next quarter if the call turns out wrong. Judgment is not a more advanced form of generation. It is what a person does with generated material before letting it become a decision.</p><p>That gap is where organizations are spending the least deliberate attention right now, even as they spend heavily on the tools that sit on either side of it. Teams have grown fluent at prompting. Fewer have built the habit of the moment after the answer returns, the moment someone has to decide what to do with it. That moment does not ship with any tool. It has to be built into how a team works, the same way the last five weeks have argued it has to be built into calibration, structure, ownership, translation, and legibility, one deliberate habit at a time rather than one sweeping policy.</p><p>This matters now, closing an arc rather than opening a new question, because the next stretch of this practice will inevitably return to specific tools and specific releases. Before that cycle starts, it is worth being precise about what does and does not change when the next tool arrives. The tools will keep improving at the comfortable half of the work. The expensive half, the half this piece is about, is not a technology problem, and no release notes will ever announce a fix for it.</p><h4>Steph’s perspective</h4><p>I have started introducing myself as an architect rather than a strategist, and the change is not cosmetic. A strategist is hired to have the answer. An architect is hired to build the structure that holds up whatever the answer turns out to be, including the version nobody has thought of yet. Across twenty years in product organizations, through three distinct waves of automation before this one, I have come to think that judgment is not a trait some people have and others do not. It is a capacity that structure either protects or wears down, and most organizations have never stopped to ask which one theirs is doing.</p><p>I want to work through one scene, because the pattern is easier to see than to argue in the abstract.</p><p>Last quarter I sat in on the closing session of a two day planning offsite for a product organization that had, by any fair measure, done the work this arc has described. Over the previous two quarters they had built a calibration check into their AI-assisted research process. They had mapped the structure around a recurring escalation instead of putting it on a single engineer’s performance review. They had named an accountable owner for a set of AI-assisted requirements documents. They had kept the parts of their retro format that worked instead of rebuilding it from scratch around a new tool. They had, at least once, pulled the raw numbers behind a dashboard that looked finished and found a quiet error in how a cohort had been defined.</p><p>Five practices, five different functions in the organization, each one built, each one real. At the closing session, when I asked the leadership team a simple question, which of these five did you actually do this quarter, not design, not announce, do, the room went quiet. Three hands went up for one practice. Two for another. For the ownership check, the one that required naming an individual accountable for a decision shaped by a tool, not a single hand.</p><p>I have seen this exact shape before, under different names, in every wave of this work. It was true when a research synthesis tool first got good enough to produce a plausible sounding literature review, and the teams that adopted it fastest were not the ones with the best research discipline. They were the ones willing to skip the discipline because the output already looked done. It was true when early automated test tools started generating coverage reports that looked exhaustive and were not, and the teams that caught the gap were the ones who had kept a habit of spot checking that predated the tool by a decade. The tool changes. The shape of what gets skipped does not.</p><p>What gets skipped is never the comfortable half of the work. The calibration check that got done consistently was the one that fit inside an existing meeting, five minutes added to a ritual that already existed. The ownership check that got skipped was the one that required a person to say, out loud, in front of colleagues, that a decision was theirs to defend even though a tool had generated the material underneath it. That sentence costs something to say. It exposes the speaker to being wrong in a way no one in the room can retroactively blame on the tool. Judgment, it turns out, is not unevenly distributed because some people think more clearly than others. It is unevenly exercised because some of its applications are cheap and some are expensive, and teams will reliably do the cheap ones first and call the work finished.</p><p>This is the part of the job I think about as architecture rather than strategy. A strategist would have told that leadership team which five practices to adopt, and they had already adopted them, correctly, on paper. What they needed instead was a structure that made the expensive application of judgment as unavoidable as the cheap one, something that did not depend on a person volunteering for discomfort in a room full of colleagues. We rebuilt the ownership check as a field in the planning template itself, a line that could not be left blank before a roadmap item moved from draft to committed. Not a new discipline asked of anyone. A structure that made the discipline the path of least resistance instead of the path of most courage.</p><p>I think about this every time a client asks me when the organization will be fully ready for the next version of whatever tool just shipped. There is no such state. The tools will keep producing more fluent material, faster, and the half of the work that depends on a person willing to be wrong in public will not get easier because the input got better. If anything it gets harder, because the output earns trust faster than the organization earns the habit of questioning it. The structure has to do more work precisely because the material looks more finished.</p><p>I do not think AI tools created this problem. I think they made it visible faster, because they are extraordinarily good at producing the comfortable half of any process (the synthesis, the summary, the draft, the dashboard) and entirely indifferent to the expensive half (the moment someone stakes their name on what to do with it). Twenty years of watching capable teams over-trust polished output has taught me that the teams who come through each wave of this intact are never the ones with the sharpest individual judgment. They are the ones who stopped treating judgment as a personal virtue and started treating it as a structural question, the same question I would ask of any system I was asked to build: what does this structure make easy, and is that the thing I actually want done.</p><h4>Stride</h4><p>Five pillars, five weeks, and one finding underneath all of them: judgment is the thing that does not arrive installed with a tool, and it does not improve on its own just because the material it is applied to gets more polished. The move this arc leaves with is not a new practice to add to the five already named. It is a single act of honest accounting.</p><p>Look back at the five pillars this arc has worked through: calibration over confidence, structural not personal, ownership in the gaps AI creates, translate don’t reinvent, legible isn’t the same as true. Name the one that is weakest on the team right now, not the one that sounds most urgent in the abstract, but the one that, examined honestly this week, would reveal the fewest hands going up when asked who actually did the work rather than who agreed it mattered.</p><p>Name it. Not “we should be better at ownership” but “no one on this team has been asked, out loud, to defend a decision shaped by an AI-assisted draft in the last quarter.” Not “our structure could be cleaner” but “the last three performance conversations addressed a problem that was actually a missing decision right.” Specificity is what turns a value into something that can be worked on.</p><p>Then make it the thing the team builds structure around next quarter, not a reminder, not a value statement, a structure. If ownership is the gap, that means a field in a planning document that cannot be left blank. If calibration is the gap, that means five minutes added to an existing ritual rather than a new meeting nobody will protect. If legibility is the gap, that means one standing question asked of any dashboard before it informs a decision: what would this look like if the underlying data were wrong.</p><p>The test of whether the structure is working will not be a feeling of confidence. It will be whether, next quarter, more hands go up when someone asks who actually did the work. That is the structure this practice exists to build toward, sturdy enough to hold judgment in place regardless of what tool arrives next.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=cf4ead25892b" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Stop Mistaking Clean for Correct]]></title>
            <link>https://stephanie-muxfeld.medium.com/stop-mistaking-clean-for-correct-f0e14e9fe81d?source=rss-697b881c9471------2</link>
            <guid isPermaLink="false">https://medium.com/p/f0e14e9fe81d</guid>
            <category><![CDATA[decision-making]]></category>
            <category><![CDATA[data]]></category>
            <category><![CDATA[product-leadership]]></category>
            <category><![CDATA[product-management]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <dc:creator><![CDATA[Stephanie Muxfeld]]></dc:creator>
            <pubDate>Mon, 19 Oct 2026 11:31:01 GMT</pubDate>
            <atom:updated>2026-10-19T11:31:01.564Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*XdOFZ6PEO0Ff6g0Amq669Q.png" /></figure><p><strong>Signal</strong></p><p>Every product team now works with documents it did not fully build. A roadmap deck returns from an AI tool with the empty cells filled and the formatting straightened. A research summary arrives with the quotes already selected and the themes already named. A metrics dashboard renders with a clean narrative sitting above the numbers, explaining what they mean before anyone in the room has asked a single question.</p><p>None of this is new in kind. Product organizations have always turned messy inputs into polished artifacts. What has changed is the ratio between the two. The distance between a rough draft and a presentable one used to take an analyst an afternoon, sometimes a week of revision, and that time asked something of the person doing it. Sitting with the data long enough to format it also meant sitting with it long enough to notice where it disagreed with itself, where a chart flattened a trend that the underlying rows did not support, where two metrics implied different stories and nobody had reconciled them yet. The friction was doing work nobody had named as work.</p><p>Now that distance closes in under a minute, and the friction closes with it. The dashboard reads as finished the moment it exists. The summary reads as authoritative because the sentence structure is confident and the formatting is clean, not because anyone has checked the sentence against the rows it claims to describe. Legibility and accuracy used to travel together often enough that teams stopped examining the difference between them. A report that was hard to read was usually also under-baked, the two problems shared a cause, which was that nobody had spent the time. That correlation is breaking. It is now entirely possible, routine even, to produce a document that reads as finished and is wrong, because the tool that built it was designed to produce something legible, not something examined.</p><p>Consider what a product review looks like in most organizations now. Someone opens a slide generated from a mix of real numbers and model-assisted narrative. The headline says retention held steady in the quarter. The chart beneath it is clean, labeled, color coded to house style. The room nods and moves to the next topic, because nothing about the slide signals that it needs more attention than any other slide. Nobody in that room asks whether the retention figure means what the label says it means, whether a change in how churn was counted upstream quietly moved the baseline, whether the clean line on the chart is smoothing over three weeks of volatility that the raw numbers would have shown on their own. The slide looked done. Looking done and being done are not the same claim, but they produce the same behavior in the room, and that is the actual mechanism worth naming here.</p><p>This shows up most sharply in three places. The first is roadmap communication, where a plan generated with AI assistance reads with more confidence than the thinking behind it has earned, and stakeholders calibrate their trust to the writing rather than to the plan itself. The second is research synthesis, where a tool trained to find patterns will find patterns whether or not the underlying sample supports them, and the resulting summary carries the same tidy tone whether it rests on forty interviews or four. The third, and the one worth naming directly, is the metrics dashboard, because a dashboard is the artifact leaders are most likely to treat as settled fact rather than as an argument that still needs examining.</p><p>None of this is an argument against using these tools. The tools are not the problem, and refusing to build with them does not restore the old correlation between legibility and accuracy. What has actually changed is the cost of looking finished, and any practice that assumed a document had earned its polish by virtue of being polished needs to be retired along with that assumption. A clean dashboard is now cheap to produce. Trust has to be earned somewhere else, and naming where it gets earned is the work ahead for most product organizations, whether or not they have noticed the correlation already broke.</p><p><strong>Steph’s perspective</strong></p><p>I have sat in more product reviews than I can count where a slide looked finished and was not, and the AI-assisted version of this problem is not new to me, it is a faster version of something I have watched for twenty years. The pattern is always the same. A report comes into the room looking authoritative, the room treats the look as the substance, and the gap between the two does not surface until something downstream breaks and someone finally asks where the number actually came from.</p><p>The example I keep returning to is a quarterly business review at a mid-size software company I worked with early in my consulting practice, well before any of the current tools existed. The product team had built a retention dashboard that the whole leadership group trusted by habit. It was clean. It was color coded. It had held the same shape for a year. In one review, a newer member of the team, someone without the history to know the dashboard was sacred, asked a simple question, what counted as retained in the denominator. Nobody in the room could answer with confidence. It took two weeks to trace the definition back through three tools and a spreadsheet formula nobody remembered writing, and what the team found was that a product change eight months earlier had silently shifted who counted as an active user, inflating the retention figure the whole time. The dashboard had looked done for a year. It had been wrong for most of that year. Nobody had examined it, because nobody had reason to, because it looked finished, and looking finished is exactly the quality that stops a room from asking the one question that would have found the problem.</p><p>What AI-generated reporting does to that same dynamic is remove the last remaining reason anyone had to look closely. Before, someone on the team had to build the dashboard by hand, and building it by hand meant touching the data enough to notice some of its seams. That person was an accidental check on the system, not because they were trying to be one, but because the work of assembly forced contact with the raw material. Now the assembly can happen without anyone touching the raw material at all. The tool pulls from the source, writes the narrative, formats the chart, and hands over something that looks exactly as finished as the version a careful analyst would have produced after a week of work, except nobody spent the week, and nobody had the accidental contact with the seams that the week used to guarantee.</p><p>I have started asking a specific question whenever I am handed a report, a roadmap summary, or a dashboard that looks finished, and I ask it before I ask whether I agree with the conclusion. The question is what this would look like if the underlying data were wrong. Not whether the data is wrong, I usually cannot know that on sight, but what the artifact itself would look like in that case. If a dashboard would look identical whether the retention number were accurate or silently broken, that dashboard is not actually evidence of anything, it is a rendering, and a rendering is not the same claim as a finding. A report that would look different if the data underneath it were wrong, one that shows its seams, that flags volatility rather than smoothing it, that names its own definitions in the header rather than assuming everyone already agrees on them, is doing something a purely legible report is not. It is giving the reader a way to check it rather than only a reason to believe it.</p><p>This is not a call to distrust every AI-generated artifact or to slow every review into a forensic audit. Most of what comes out of these tools is fine, and treating every dashboard as a crime scene would spend exactly the kind of attention this problem calls for directing somewhere else. The distinction I draw in practice is between a report built to be read and a report built to be checked, and the two are not the same design problem. A report built to be read optimizes for the room nodding. A report built to be checked carries its own evidence close to the claim, its sample size, its definitions, its date range, close enough that a reader does not have to go looking for them. The habit I am naming is not paranoia. It is the one question that restores the examination an AI tool quietly removed, ask what the artifact would look like if it were wrong, before deciding whether to trust what it says when it looks right.</p><p>The lesson from that early retention dashboard has stayed with me through every later version of this same mistake, because the mistake never actually changes shape. What changes is only how cheaply the next clean, confident, wrong artifact can be produced, and how little friction remains to slow anyone down long enough to notice.</p><p><strong>Stride</strong></p><p>Pick one dashboard or report the team already trusts without examining, the one nobody in the room would think to question because it has looked the same way for months. Before the next meeting where that artifact appears, run one exercise against it, ask what it would look like if the underlying data were wrong.</p><p>Start with the definition behind the headline number. Trace what the metric actually counts, not what the label implies it counts, and confirm when that definition was last checked against the system producing it. A number whose definition nobody can state with confidence is not evidence yet, however finished its chart looks.</p><p>Then examine what the report would hide if it were wrong. A clean trend line smooths volatility by design. Ask whether the raw numbers underneath it show the same steadiness the chart presents, or whether weeks of noise have been averaged into a single reassuring slope. A report that would look identical whether its inputs were sound or broken is not built to be checked, and that gap is worth naming directly to whoever owns the artifact, not as a criticism of their work but as a design note for the next version of it.</p><p>Finally, add one line to the artifact itself rather than only to the conversation about it, a visible note of the sample size, the date range, or the definition behind the headline number, placed close enough to the claim that nobody has to go looking for it. This single addition does more to restore the link between legibility and accuracy than any review step layered on top, because it gives the reader something to check without requiring a question the room has stopped thinking to ask.</p><p>The team that builds this habit into one dashboard this week will have a working template for every other report that currently earns trust by default. The team that does not will keep mistaking a document’s polish for its soundness, right up until a number everyone trusted turns out to have been wrong for longer than anyone wants to admit.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=f0e14e9fe81d" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Stop reinventing, start translating]]></title>
            <link>https://stephanie-muxfeld.medium.com/stop-reinventing-start-translating-9e346d63962b?source=rss-697b881c9471------2</link>
            <guid isPermaLink="false">https://medium.com/p/9e346d63962b</guid>
            <category><![CDATA[product-management]]></category>
            <category><![CDATA[leadership]]></category>
            <category><![CDATA[business-strategy]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[future-of-work]]></category>
            <dc:creator><![CDATA[Stephanie Muxfeld]]></dc:creator>
            <pubDate>Mon, 12 Oct 2026 11:31:01 GMT</pubDate>
            <atom:updated>2026-10-12T11:31:01.775Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*K2dJ08DTCEVAOuXf4AOZ4w.png" /></figure><p><em>Every new tool cycle tempts organizations to rebuild their processes from scratch. The smarter move is to translate what already worked.</em></p><h3>Signal</h3><p>A finance team automates its purchase approval chain. A customer success group rebuilds onboarding around an AI agent that sends the welcome sequence, schedules the kickoff, and answers setup questions. A sales organization replaces its weekly pipeline review with a dashboard and a generated summary. A recruiting team swaps the hiring debrief for structured scorecards that a tool rolls up into a recommendation. Each change is made for a good reason. The old approval chain was slow. Onboarding depended on a few overloaded people. The pipeline meeting ate an hour of every manager’s Monday. Six months later, the new versions are faster, cleaner, and easier to report on, and something has gone missing from each of them that nobody can put words to.</p><p>This pattern repeats with every tool cycle, and the current one is moving faster than any before it. When a capability arrives that seems to change how work gets done, the instinct is to treat the processes built around the old way of working as part of the old way of working. Approvals, reviews, handoffs, intake, and planning get filed under how we used to do it, and the team starts rebuilding from a blank page with the new tool at the center.</p><p>The trouble is that most business processes were never about the format they happened to take. A purchase approval is not a form. It is a moment where someone with context asks whether this spend makes sense right now. A pipeline review is not a meeting. It is the place where a manager hears the hesitation in a rep’s voice before the deal slips. A hiring debrief is not a scorecard. It is the room where one interviewer’s concern gets tested against everyone else’s read before an offer goes out. The format is what shows up on the calendar or in the workflow tool. The function is what the process actually does for the business, and a function does not expire because the tool changed.</p><p>When a process is rebuilt from scratch, the visible parts usually make it into the new version. Approvals still get routed. Onboarding emails still go out. The pipeline still gets reported. The invisible parts tend to disappear: the judgment call a person made at a handoff, the informal check that caught an exception before it became a problem, the moment where someone who disagreed got to say so in front of the people who could act on it. Those parts were rarely written down, because they were never anyone’s job description. They were simply the reason the process worked.</p><p>The current generation of AI tools makes this easier to miss, not harder. An agent can reproduce the artifact of a process with remarkable fidelity. It can route the approval, draft the summary, populate the scorecard, and send the follow-up. Because the output looks complete, the absence of the function underneath it does not announce itself. It shows up later, as an exception nobody caught, a risk nobody heard about, or a decision nobody can explain.</p><p>G.K. Chesterton described the general shape of this mistake nearly a century ago, in a <a href="https://fs.blog/chestertons-fence/">parable about a fence</a> standing across a road, from his 1929 book The Thing. A reformer sees no use for the fence and wants it cleared away. Chesterton’s answer: “If you don’t see the use of it, I certainly won’t let you clear it away. Go away and think. Then, when you can come back and tell me that you do see the use of it, I may allow you to destroy it.” The point is not that the fence must stay. The point is that the person removing it should be able to say what it was for.</p><p>Every organization redesigning its processes around AI right now is standing in front of a long row of fences. The speed the new tools bring is real and worth having. But a tool that reproduces the artifact of a process does not automatically carry its function, and the distance between those two things is where most of the value leaks out.</p><h3>Steph’s perspective</h3><p>I have led or supported more organizational transformations over twenty years than I can count cleanly, and the ones that went sideways almost never failed on the new method or the new tool. They failed on what got thrown away in the process.</p><p>One example involves a customer intake process. Requests came in through a shared queue, and a support lead read every request that did not fit an existing category before routing it. The team called that pile “other,” and it was treated as overhead, a sign the categories needed work. When intake was rebuilt around automated classification with tighter categories, “other” all but disappeared. Routing got faster. Reporting got cleaner. Everyone considered it a success.</p><p>Two quarters later, a long-standing customer escalated a problem that had been building for months. When we traced it back, the early signs had been there the whole time, spread across a handful of requests the new system had confidently routed into the nearest category, each one reasonable on its own. Under the old process, those requests would have landed in front of the support lead, who would have noticed three of them in a week and walked over to the product team. The “other” pile had been doing early warning work. It had just never been labeled that way.</p><p>The fix was not to bring back the manual queue, and it was not to add another tool to detect patterns. The fix was to translate. We named what the old step had actually been doing: putting the strange, hard to classify requests in front of a person with enough context to notice when strange was becoming a pattern. Then we built that function into the new process, with a weekly human review of every low confidence classification and a direct line from that review to product. The format changed. The function survived.</p><p>That experience changed how I approach every process redesign since. When I build a framework for a team now, I start by asking what the processes it replaces were for, not what they looked like. Most processes carry three things worth examining before anything gets retired.</p><p>The first is the function, the job the process did for the business, often different from its stated purpose. The second is the conditions that made it work, the person with context, the timing, the moment of judgment, the small agreement nobody wrote down. The third is the signal, how the organization knew the process was doing its job, and what would have told them it had stopped. When I can name all three, I know what I am translating. When I cannot, I know I am about to clear a fence I do not understand.</p><p>Translating is not the same as preserving. Some processes deserve to go. An approval step built to compensate for missing budget visibility has done its work once that visibility exists, and it can be retired with thanks. A status meeting that existed only because information was hard to find may have no function left once the information is easy to find. Part of translating well is noticing which parts of a process were ritual and letting those parts go without guilt. The discipline is in being able to say which is which, out loud, before the change ships rather than after something breaks.</p><p>The difference between the two approaches shows up in where the conversation starts. Reinventing starts from the new tool: here is what it can do, so here is the process we will build around it. Translating starts from the work: here is what our processes were doing for us, here is how the new tool can carry that forward, and here is what it cannot carry, which a person still has to hold. The first conversation is faster and feels more modern. The second one is slower by perhaps an afternoon, and it is the one that tends to hold up six months later.</p><p>I think of this as the architect’s part of the job. Frameworks last when they are built around function rather than format, because formats change with every tool cycle and functions mostly do not. Organizations still need to catch exceptions, make decisions with the right context, surface risk early, and remember what they chose and why. The tools that help them do it will keep changing, faster now than at any point in my career. The work of carrying a process forward, rather than starting over on it each time, is what keeps an organization from relearning the same lessons with better software.</p><h3>Stride</h3><p>Name one process your organization is rebuilding from scratch right now. It might be an approval chain, customer onboarding, intake, a hiring loop, vendor review, a planning cycle, or the weekly business review. Most organizations have at least one in motion, and if an AI tool arrived this year, there are probably several.</p><p>Before the replacement ships, write three lines about the version being retired. What was it for, meaning the job it did for the business, not the job printed in the workflow diagram. What made it work, meaning the conditions underneath it: who looked at it, what context they brought, what judgment they applied before it moved on. How did you know it was working, meaning the signal that told the organization the process was earning its time.</p><p>If you cannot answer the first line, do not retire the process yet. Find the person who has run it longest and ask them what would go wrong if it stopped tomorrow. Their answer is usually the function nobody wrote down.</p><p>With the three lines in hand, check the new version against each one. Does it carry the function, or only the artifact? Does it keep the conditions, or does it remove the person, the context, or the moment of judgment that made the old version work? Does it give you a signal, or will you only find out it stopped working when a problem surfaces two quarters from now? Anything the new version cannot carry needs a deliberate home: a standing human review, a named owner, or a short step the tool does not replace.</p><p>Then make one decision explicit. Either the old process’s function now lives somewhere specific in the new one, or the organization has agreed out loud that the function is no longer needed. Both are good outcomes. The one to avoid is the quiet third option, where the function disappears because nobody noticed it was there.</p><p>Keep the three lines somewhere the team can find them. The next tool cycle is already on its way, and the organization that wrote down what its processes were for will translate them forward in an afternoon instead of rebuilding them from a blank page.</p><h3>Sources and references</h3><ul><li>G.K. Chesterton, The Thing, 1929, chapter “The Drift from Domesticity.” <a href="https://fs.blog/chestertons-fence/">Overview of the passage</a></li></ul><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=9e346d63962b" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Who signed off on that?]]></title>
            <link>https://stephanie-muxfeld.medium.com/who-signed-off-on-that-12366bbb4249?source=rss-697b881c9471------2</link>
            <guid isPermaLink="false">https://medium.com/p/12366bbb4249</guid>
            <category><![CDATA[product-management]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[decision-making]]></category>
            <category><![CDATA[accountability]]></category>
            <category><![CDATA[leadership]]></category>
            <dc:creator><![CDATA[Stephanie Muxfeld]]></dc:creator>
            <pubDate>Mon, 05 Oct 2026 11:31:01 GMT</pubDate>
            <atom:updated>2026-10-05T11:31:01.462Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*cYBq4kZD7PpicV059jwGcw.png" /></figure><p><strong>Signal</strong></p><p>A product requirements document moves through the usual channel. It gets built from a mix of customer interview transcripts, a running document of scattered observations, and a tool that drafts a first cut of structure, flags open questions, and proposes acceptance criteria. The writer tightens the language, adds two examples from direct experience, and sends it for the engineering team’s review. The team has no follow-up questions and starts on design and development immediately. Maybe there is a formal “accepted” flag, but more likely the acceptance is tacit in that they start working off what’s included in the document.</p><p>Nothing about that approval is wrong on its surface. The document reads clearly. The acceptance criteria are specific. The open questions are the right ones. But ask the person who approved it why the third acceptance criteria exists, rather than a fourth one the tool also proposed and the writer cut, and the answer gets thin fast. Ask the writer the same question and the answer sometimes gets thin too, because the tool surfaced four plausible criteria, the writer kept the three that looked strongest, and the reasoning behind the cut lived in a few seconds of reading and contemplating rather than in a documented tradeoff anyone examined.</p><p>This is happening across the artifacts organizations produce every week, not only requirements documents. Research summaries get assembled from a stack of interview recordings. Code gets generated from a short natural language prompt and shaped into a pull request. Roadmap narratives get drafted from a quarter of scattered planning documents and status updates. Each of these moves through a review step that was built for the old version of an artifact, one where a single person generated the reasoning from the first word to the last and could, in principle, defend every line of it under direct questioning.</p><p>That assumption is no longer reliably true, and the review step has not caught up to the change. A team approving a requirements document today is often approving something assembled across two very different relationships to the reasoning inside it: a tool that generated structure and options at speed, and a person who selected, cut, and lightly edited what the tool produced. The review step treats this as equivalent to a document one person wrote end to end. It is not equivalent. The person who touched the document last can explain what the document says. Whether they can explain why it says that, rather than something adjacent the tool also proposed and they declined, is a separate question entirely, and it is the question most review steps were never built to ask.</p><p>The speed gain here is real and worth calling out rather than treating with suspicion. A requirements document that once took three days now takes an afternoon. A research summary that once took a week of transcription and theme coding now takes an hour. This is not a complaint about the tools themselves. The tools are doing what they were built to do, producing competent first drafts and structured options faster than a person working alone could ever manage. The problem sits one layer over, in the review and approval structure that still assumes the speed came from a person working harder or more efficiently, rather than from a different kind of authorship entirely, one with a different shape and a different set of blind spots.</p><p>Three places this shows up most clearly in product organizations right now are worth naming.</p><ol><li>The first is requirements and specification work, where a tool proposes structure, edge cases, and acceptance criteria, and a person edits and approves rather than originates from a blank page.</li><li>The second is research synthesis, where interview data gets summarized and theme coded by a tool, and a researcher checks the output for accuracy against the source material rather than building the themes directly from the transcripts.</li><li>The third is code review, where a reviewer evaluates a pull request that may be substantially tool generated, written to look idiomatic and internally consistent in a way that makes surface level review easier and the deeper question, does the reasoning behind this actually hold, harder to catch in the same pass.</li></ol><p>None of this means the resulting work is worse. Some of it is measurably better. More options get considered. More edge cases get named. Less time gets lost staring at a blank page waiting for the first sentence to appear. What it means is that the artifact and the accountability behind it have quietly come apart from each other, and most review processes in most organizations still operate as though they are the same thing, because until recently, they were.</p><p>The gap does not tell you it’s there. A document that reads well, moves through review quickly, and later turns out to rest on a choice nobody can explain looks, from the outside, exactly like a document that reads well and turns out fine. The difference only shows up when something goes wrong and the organization goes looking for the person who can explain the reasoning behind the original choice, and finds that the honest answer is that no one examined it closely enough to carry that explanation forward.</p><p><strong>Steph’s perspective</strong></p><p>I spent a meaningful part of my career building the review structures product organizations use to approve the things that matter: requirements documents, launch readiness checklists, go to market plans, the gates a decision has to clear before it becomes real and starts costing money to reverse. Every one of those structures carries a hidden assumption underneath its checklist items. It assumes that the person whose name sits on the document is the person who did the thinking behind it. The gate checks for completeness, for clarity, for alignment with strategy. It does not check who actually reasoned through the tradeoff at the center of the decision, because for most of the history of this work, those were the same person, and nobody had reason to separate the two questions.</p><p>I watched that assumption break in a launch review at a company I was consulting with about a year ago. A product manager I trust, someone with good judgment and a strong track record over several years, presented a go to market plan that was unusually tight. Clear segmentation, a sharp positioning statement, a sequencing plan that anticipated two objections I would have raised myself if given the chance. It was approved on the call where it was introduced, the way this group had approved dozens of similar plans before it. It was only in a follow up conversation a week later that I learned the segmentation had come from a tool that processed the account data and proposed three distinct segment definitions, and the product manager had picked the one that read best on the page and built the rest of the plan around that choice. The other two segmentations were never discussed with anyone on the team. Not because they were hidden or withheld, but because nobody thought to ask, myself included, since the plan in front of me looked complete.</p><p>That is the moment I started thinking about this differently than I had before. The plan itself was not wrong. The segmentation held up fine when we later tested it against actual pipeline data over the following two quarters. But a decision got approved without the approvers knowing that the decision point at its center, which segment to build the entire plan around, had been made by a selection from a tool’s output rather than by a reasoned argument a person in the room could walk us through. If the segmentation had turned out to be wrong, no one would have owned the why, because the honest answer would have been that it read better than the alternatives on the page, which is a legible sounding reason that is not actually a reason at all.</p><p>I want to be clear about what I am not arguing here, because it is easy to hear this as a case against the tools, and that is not where I land. I am not arguing that a person has to generate every option by hand for a decision to count as genuinely owned. That has never been how good decision making worked, even decades before any of these tools existed. Analysts have always produced options that executives select among, and nobody considered that a problem. The difference is that an analyst can be asked why they ranked one option first, and the ranking reflects a chain of reasoning a person built and can defend under real pressure in the room. A tool’s ranking reflects a different kind of process underneath it entirely, patterns learned across enormous amounts of text rather than a reasoned argument about this specific situation, and when a person adopts that ranking without doing the work of making it genuinely theirs, ownership has quietly moved from the person approving the document to no one in particular. Generated is not the same as owned. The ownership has not transferred to the tool. It has simply evaporated.</p><p>Ownership, in the sense I mean it here, has nothing to do with who gets credit when something goes well or who gets blamed when it does not. I think of it as a functional property a decision needs in order to be correctable later, when the world around it shifts. If I know that you own a decision, I know exactly who to go to when new information arrives, who can tell me what would change their mind and what would not, who understood the tradeoff well enough at the time to revisit it when circumstances change under it. A decision with no clear owner, even a good decision that happens to work out, is much harder to correct later, because nobody left behind can tell you what assumptions it actually rested on or which of those assumptions have since stopped being true.</p><p>This is the gap I think about now in every review I sit in, and I sit in a fair number of them. Not primarily whether the artifact in front of me is good, though that still matters. Whether there is a specific person in the room who can tell me why it says what it says, and who will still be answerable for that reasoning in six months, when the context around the decision has shifted in ways nobody can predict from here. That second question is the one most review processes stopped asking the moment the artifacts in front of them started arriving faster than anyone expected.</p><p>The uncomfortable part, and I think it is worth pointing out, is that naming an owner takes a few extra seconds in a process that has gotten dramatically faster everywhere else around it. It feels like friction precisely where everything else has sped up, and that is exactly the kind of step that gets skipped under deadline pressure, first occasionally and then as a matter of habit. I have started treating that friction as the signal that the step matters, rather than as evidence it should be cut to keep pace with everything else. The organizations I have watched handle this well over the past couple of years are not the ones using these tools less often or more cautiously. They are the ones who added one small, deliberate question back into a process that had sped up everywhere else, and who treat answering that question as part of the work itself rather than as overhead sitting on top of it.</p><p><strong>Stride</strong></p><p>The move here is small enough to make this week, and specific enough that it will not get skipped once it is actually in place. Add one line to the definition of done for any AI assisted deliverable in the organization, a requirements document, a research summary, a pull request, a roadmap narrative, naming the person accountable for the reasoning inside it. Not the person who generated the first draft, and not necessarily the person who gave it a final read before it shipped. Those can be the same person on any given document. The point of the line is that it forces the question to get asked out loud, every time, instead of assumed by default because the document looks finished.</p><p>Write the line so it asks for a name and a reason together, not a name on its own. Something close to this works well in practice: accountable owner, the person who can explain why this option was chosen over the alternatives the tool or the process surfaced along the way. Put that line in the same place the definition of done already lives, next to the checklist item for test coverage or the checklist item for legal review, so it reads as a normal part of shipping rather than a special step reserved only for decisions that already feel high stakes in the moment. The decisions that need this most are usually the ones that do not feel high stakes at all when they cross a desk: the segmentation that reads well, the acceptance criteria that looks obviously right on first read, the summary that happens to match what the reviewer already expected to see.</p><p>Test the question itself on the next AI assisted deliverable that crosses your desk for approval, before the line even exists in writing anywhere. Before approving anything, ask the person who touched it last one direct question: what was the alternative you did not choose, and why not. If the answer comes quickly and specifically, naming the alternative and the actual reason it lost, the ownership is already there, and writing the line into the definition of done will take seconds. If the answer takes a while to arrive, or arrives as a restatement of why the chosen option is good rather than a reason the others were worse, that gap is the thing worth noticing and naming before the document goes any further toward becoming real.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=12366bbb4249" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[When the org chart is the real bug]]></title>
            <link>https://stephanie-muxfeld.medium.com/when-the-org-chart-is-the-real-bug-ad272366016c?source=rss-697b881c9471------2</link>
            <guid isPermaLink="false">https://medium.com/p/ad272366016c</guid>
            <category><![CDATA[workplace-culture]]></category>
            <category><![CDATA[product-management]]></category>
            <category><![CDATA[management]]></category>
            <category><![CDATA[leadership]]></category>
            <category><![CDATA[organizational-design]]></category>
            <dc:creator><![CDATA[Stephanie Muxfeld]]></dc:creator>
            <pubDate>Mon, 28 Sep 2026 11:31:01 GMT</pubDate>
            <atom:updated>2026-09-28T11:31:01.899Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*IPhdLR5Rtaaong1Ro7EDwg.png" /></figure><h4>Handoffs, decision rights, and incentives explain more recurring problems than the person sitting across the table.</h4><h3>Signal</h3><p>A team misses the same deadline for the third release running, and the fix that gets proposed is always a conversation. Someone schedules time with the engineer who is not communicating enough, the product manager who needs to move faster, or the designer who keeps holding up review. The conversation happens. It goes well. Everyone nods, commits to doing better, and returns to their desks. Six weeks later the deadline slips again.</p><p>Watch enough of these cycles and a pattern surfaces. Leaders default to a performance conversation when the actual fault sits in the structure around the person, not in the person doing the work. A handoff has no clear owner, so whoever picks it up last absorbs the blame for a delay that started three steps earlier. A decision has no assigned decision maker, so the person closest to the deadline ends up making a call they were never given the authority to make, and gets reviewed on judgment they were forced to exercise without a mandate. A reporting line routes two functions through a single manager who has time to arbitrate one conflict a week, so the second and third conflicts get decided by whoever argues longest, and that person is then treated as difficult.</p><p>None of this shows up as a structural problem while it is happening. It shows up as a person problem, because a person is the visible surface of it. The engineer who missed the deadline is real, and so is the performance conversation where someone raises it with them. What is harder to see is the three steps upstream, where a handoff had no name attached to it, or where the org chart drew a reporting line that made this person the default arbiter of disputes that were never theirs to settle.</p><p>None of this makes the performance conversation wrong. Sometimes a person genuinely needs to communicate more, or move with more urgency, or let go of a review pattern that is slowing everyone else down. The point here is not that individual conversations never apply. It is that they get reached for first, by default, before anyone examines whether the structure would have produced the same outcome regardless of who was in the seat. Replace the person and leave the structure unfixed, and the same failure recurs with a new name attached. That recurrence is the tell we’re naming here. A genuine performance issue does not usually survive a change in personnel. A structural one does, because the structure was never the variable under discussion.</p><p>This is easy to miss because structural fixes are slower and less visible than individual ones. A conversation with one person can happen this week and feel like progress by Friday. Naming a decision owner, redrawing a handoff, or moving a reporting line touches more people, takes longer to land, and rarely produces the immediate sense of having done something that a one-on-one does. So the faster, more visible move gets taken, and the structure that actually produced the problem goes unexamined for another release cycle, waiting to hand the same failure to whoever stands in that spot next.</p><p>There is a third layer beneath the handoff and the decision, and it is the one leaders examine least often: the incentive sitting quietly beneath both. A team can have a clearly named handoff and a clearly assigned decision owner and still produce the same recurring failure, if the incentive structure rewards the wrong thing at the point where the failure occurs. A support function measured purely on speed of response will always look slower than a function measured on nothing at all, even when the first function is doing more work correctly. A manager rewarded for shipping on schedule has every reason to route disputes toward whichever direct report has the least standing to push back, regardless of who is actually right. None of this shows up on an org chart. It shows up in a scorecard, a bonus structure, or the metric a leader mentions in a review meeting, and it shapes behavior just as forcefully as a reporting line does.</p><h3>Steph’s perspective</h3><p>I have spent over twenty years watching what gets called a people problem resolve itself the moment the structure around the person is changed, without a single individual performance conversation taking place. Not most of the time. Nearly every time I have gone looking for the structural cause instead of settling for the obvious individual one.</p><p>The clearest version of this I remember involved a product organization where one particular team lead kept getting flagged as slow to ship. Three managers in a row had the same conversation with her: be more decisive, move faster, stop letting review cycles drag. Each manager left believing they had been direct and that she had agreed to change. Each release, the same pattern showed up in the retro anyway.</p><p>What I found when I traced the actual handoffs was not a decisiveness problem. It was a routing problem. Design signed off on a screen and handed it to her team. Her team built against that screen. Somewhere in the three weeks between design sign-off and ship, legal or compliance would flag a change, and because no one owned reconciling a late compliance flag against an already-approved design, the reconciliation fell to whoever was holding the work when the flag arrived. That was her, every time, because her team shipped last in the sequence. She was not slow. She was the only person in the chain with no standing to say no to a change that arrived after everyone else had already signed off and moved on. Once compliance review moved earlier, with an explicit deadline set before design sign-off rather than after it, the delays stopped. No one had a performance conversation with her about decisiveness again, because there was nothing left to be decisive about.</p><p>What stays with me from that case, and the many I’ve seen since, is how reasonable every manager’s diagnosis looked in the moment. She really was the person nearest the missed date. The evidence, read at face value, pointed straight at her. That is what makes the structural read harder to reach for than the individual one. The individual read is not lazy or careless. It is what the available evidence appears to support, right up until someone asks a different question.</p><p>The question I have learned to ask first, before scheduling any conversation about a recurring problem, is whether the person in front of me is the cause or the last stop. Those look identical from the outside. A cause originates the problem. A last stop simply inherits it, because the structure routes accountability to whoever happens to be holding the work when it becomes visible, regardless of where the fault actually sits. Much of what gets treated as a performance issue, in my experience, is a last stop being mistaken for a cause.</p><p>This is not an argument that individual accountability does not matter. It matters enormously, and a leader who cannot tell the difference between a structural failure and a genuine performance gap will avoid necessary conversations just as often as they conduct unnecessary ones. The discipline is not choosing structure over people. It is checking the structure before assuming the person, because the check is cheap and the misdiagnosis is expensive. A wrong individual conversation costs trust, and it costs that trust specifically with the people who are already absorbing structural failures that were never theirs to own. Repeat that mistake enough times with the same person and the lesson your best people take from it is that the fastest way to avoid blame is to stop being the one who catches problems late in the sequence, which is the opposite of what you actually want them learning.</p><p>I think of this as architecture, not diagnosis. An architect does not ask what is wrong with the people standing in a building that keeps flooding in the same corner. An architect asks what the building is doing that routes water there. The people are not the structure. They operate inside it, and a structure with a flaw will produce the same flaw regardless of who is standing where, until someone examines and changes the structure itself.</p><h3>Stride</h3><p>Before the next one-on-one scheduled to address a recurring problem, map the structure around the person first.</p><p>Start with the handoff. Trace the work backward from the point where it became visible as a problem to the point where it entered this person’s hands. Name who owned each step before them and what was true about the work when they received it. If the work arrived already late, already ambiguous, or already carrying a decision no one upstream wanted to make, the structure is the more likely cause, and the conversation belongs with whoever owns the step where the ambiguity first appeared, not with the person who happened to be holding the work when it surfaced.</p><p>Then examine decision rights. Ask directly whether this person had the authority to make the call they are being evaluated on. A missing decision owner does not announce itself. It shows up as someone making a judgment call under pressure, without a clear mandate, and then absorbing the consequences of a decision the organization never formally assigned to anyone. If the answer is unclear, that ambiguity is the finding. Name a decision owner before the next version of this problem reaches someone else.</p><p>Then look at the reporting line itself. Notice whether this person sits at a structural chokepoint, the place where two functions, two priorities, or two timelines collide and someone has to arbitrate. If they do, the frequency of conflict reaching them is not a reflection of their judgment. It is a reflection of how much unresolved tension the structure routes through their desk. A person who looks like they cannot manage competing priorities may simply be managing more of them than the role was ever designed to hold.</p><p>Then check the incentive sitting beneath all of it. Ask what this person, and whoever they answer to, is actually rewarded for at the moment the recurring problem shows up. If the scorecard, the bonus structure, or the metric mentioned points one direction while the behavior you want points another, the structure is working exactly as designed, just not for the outcome anyone intended. An incentive misaligned with the outcome will beat a good intention every time, and it will keep beating it until someone changes what gets measured or rewarded.</p><p>Only after this mapping is complete does the individual conversation become worth having, and it will be a different conversation than the one that would have happened without it. Sometimes the mapping confirms that the person genuinely needs to change something, and that conversation is now sharper, because it is not carrying the weight of a structural problem the person could never have fixed alone. More often, the mapping surfaces a handoff with no owner, a decision with no assigned authority, or a reporting line asking one person to arbitrate what should have been decided somewhere else. Fix that first, and notice whether the pattern that looked like a person problem simply stops recurring.</p><p>The habit worth building is not complicated. It is the order of operations: structure before person, every time a problem repeats with a different name attached to it. The org chart is not neutral scaffolding sitting quietly beneath the work. It is often the actual source of the bug that everyone keeps trying to fix by talking to whoever it lands on last.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=ad272366016c" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The confidence problem]]></title>
            <link>https://stephanie-muxfeld.medium.com/the-confidence-problem-8d510373bbe9?source=rss-697b881c9471------2</link>
            <guid isPermaLink="false">https://medium.com/p/8d510373bbe9</guid>
            <category><![CDATA[technology]]></category>
            <category><![CDATA[product-management]]></category>
            <category><![CDATA[product-leadership]]></category>
            <category><![CDATA[decision-making]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <dc:creator><![CDATA[Stephanie Muxfeld]]></dc:creator>
            <pubDate>Mon, 21 Sep 2026 11:31:01 GMT</pubDate>
            <atom:updated>2026-09-21T11:31:01.910Z</atom:updated>
            <content:encoded><![CDATA[<p><em>Why fluent AI output keeps getting mistaken for correct output</em></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*iuu7bvhUUB1yRvg_TY0yUg.png" /></figure><h3><strong>Signal</strong></h3><p>AI tools now produce polished, decisive sounding output on nearly everything: a market analysis, a competitive teardown, a roadmap recommendation, a customer segmentation, a one-page brief distilling a hundred pages of research into three clean bullets. The output arrives complete. It reads like it already survived scrutiny, formatted the way a finished deliverable is formatted, worded with a level of certainty that used to be earned only after several rounds of review. Ask a tool for the top three risks in a launch plan and it hands you three, ranked, each one stated flatly, with no hedge anywhere in the sentence structure. Ask it to size a market and it returns a number, sometimes to two decimal places, with no visible seam where an assumption was made.</p><p>This is a genuinely new problem, not a repackaged old one. Poor recommendations have always existed. What is new is how they are dressed. A weak analysis used to come with tells: a shaky verbal delivery, a document that looked rushed, a source list that was thin or missing, a person who could not answer a follow-up question without stumbling. Those tells did real work. They told a reviewer where to press. AI-generated output removes almost all of them. The prose is even. The structure is clean. The confidence is uniform whether the underlying reasoning is airtight or invented outright, because the model is optimizing for fluent, plausible language, not for calibrated uncertainty. It has no reason to sound less sure about a guess than about a fact, because from the model’s side, both are simply the next most probable sequence of words.</p><p>Teams are starting to treat that fluency as a proxy for correctness, and the mistake is understandable. For twenty years, confidence in a business setting correlated, imperfectly but usefully, with competence. The person who spoke plainly and directly usually had done the work. The recommendation that arrived clean and well-argued usually reflected a clean and well-argued process behind it. That correlation is what let organizations move fast: a leader could read the tone of a memo, or hear the tone of a person, and use it as a shortcut for how much scrutiny the underlying claim needed. AI severs that correlation almost completely. Fluency is now cheap to produce and available to anyone, on any claim, true or false, well-supported or entirely fabricated. The tone of the output has stopped being informative about its accuracy, and most teams have not yet adjusted the way they read it.</p><p>For a product organization specifically, the pattern shows up in the artifacts a CPO already reviews every week: the one-pager framing a build versus buy decision, the competitive teardown feeding a positioning call, the PRD that translates a customer request into a set of requirements, the roadmap slide that ranks initiatives by expected impact. Each of these documents has always carried some risk of being wrong. What has changed is how much more of the drafting work behind them can now be delegated to a tool, and how much less of the visible effort in the final document reflects the effort that actually went into getting it right. A PRD that took a team three weeks to reason through and a PRD that a model drafted in an afternoon from a rough prompt can look identical on the page. Both will be well organized. Both will read decisively. The organization’s existing review habits, built for a world where visible polish correlated with invested effort, do not distinguish between them, and that is exactly the gap this problem exploits.</p><p>The effect shows up first in places where speed already outpaces verification: financial forecasts assembled from a partial data pull, competitive analyses built on the model’s general knowledge rather than current pricing, technical estimates that treat every unknown as a known. In each case, the deliverable looks the same whether it is right or wrong. A reviewer scanning for red flags in the old sense, in awkward phrasing, in hedging language, in visible gaps, will find none, because there are none to find. The gaps have not gone away. They have simply stopped being visible in the text.</p><p>None of this is an argument against using the tools. It is an argument for building a new habit around them, because the old habit, reading tone as a signal of substance, no longer works, and most organizations have not replaced it with anything yet.</p><h3>Steph’s Perspective</h3><p>I noticed this first not in a dramatic failure but in a small, almost boring one, which is usually how the real problems show up. A product manager on my team brought a demand forecast into a planning review, built with a fair amount of AI assistance: a model had been fed historical order data and asked to project the next two quarters, broken out by region. The document that came out of it was genuinely well put together. Three tidy paragraphs of narrative context, a table with clean columns, a confident verb in every sentence, a short list of caveats at the bottom that read like due diligence had been done. Nothing in it looked unfinished. Nothing in it invited a question.</p><p>The room accepted it in about four minutes, which in my experience is close to the fastest a demand forecast has ever moved through a planning meeting in twenty years of doing this work. Nobody asked what the model had assumed about seasonality. Nobody asked whether the historical window it was trained on included the supply shortage from the year before, or whether that anomaly had been excluded, or whether it mattered either way. Why would they, when nothing in the tone of the document invited the question. It read the way a finished, reviewed piece of analysis reads, and the room responded to the reading, not to the reasoning underneath it.</p><p>The forecast was wrong. Not wildly wrong, which might have caught someone’s attention sooner, but wrong in the quiet, expensive way that shows up two months later as slipped inventory and a scramble to reallocate. When we went back to find the error, it turned out the model had simply carried the prior year’s seasonal pattern forward without adjusting for a known one-time event, a detail any of the three people in that room could have caught in under a minute if the document had given them a reason to look. It did not give them a reason to look. That is the entire problem in one sentence.</p><p>What struck me afterward was not that the forecast was wrong. Forecasts are wrong constantly, and that has never been the standard we hold them to. What struck me was how completely the fluency of the language had substituted for the verification the room would normally have supplied on its own. For most of my career, a hedge in someone’s voice, a pause before answering a hard question, a raised hand in the back of the room, told you exactly where to push. Those signals were not perfect, but they were honest, in the sense that they correlated with the presenter’s own actual uncertainty. A model has no equivalent tell, because it is not tracking its own uncertainty in a way that leaks into its tone. It is producing the most fluent available sentence, and fluent is not a synonym for sure.</p><p>I have also seen the opposite version of this story, on a different team, working through a competitive positioning brief. The person presenting it had used AI heavily to draft it, and said so upfront, then added a specific line I still bring up in reviews: I am confident in the second and fourth sections, because I checked those claims against source pricing pages myself, and I am not confident in the third section, because the tool inferred a competitor’s roadmap from public statements that could be marketing rather than intent. That one sentence changed how the room used the document. Nobody threw it out. They used it exactly as much as it deserved to be used, section by section, because someone had done the work of separating the parts that were verified from the parts that were fluent guesses wearing the same formatting. The document was not better written than the demand forecast. The difference was entirely in what the presenter volunteered about it before anyone had to ask.</p><blockquote>I have come to think of this as a redistribution of work rather than a new kind of work.</blockquote><p>The calibration that used to arrive for free, carried in a person’s voice or a document’s rough edges, now has to be supplied deliberately, by a human, because the tool will not supply it and was never built to. That is not a criticism of the tool. It is a description of what changed. The teams handling this well are not the ones avoiding AI-assisted analysis, and they are not the ones blindly trusting it either. They are the ones who have built a specific, named step back into the process where a human states, out loud, how confident they actually are in a claim, independent of how confident the claim sounds. That step used to be implicit. It has to be explicit now, or it disappears entirely, and the room moves as fast on a guess as it does on a fact.</p><h3>Stride</h3><p>Build one deliberate calibration check into an AI-assisted workflow this week, and make it a named step rather than a hoped-for habit. Before any AI-assisted recommendation, forecast, or analysis moves from draft to decision, have the person who generated it state their actual confidence level out loud or in writing, in plain language or a rough number, before the document’s own fluency gets to state it for them. Put it on the agenda the same way a budget line is put on an agenda: not implied, written down, sequenced ahead of the discussion of the recommendation itself.</p><p>Keep the check specific enough to be usable. A general instruction to be more skeptical of AI output tends to fade after a week, because it asks people to sustain an abstract vigilance with nothing to anchor it to. A specific question does not fade the same way: what is the one assumption this analysis would fall apart without, and how confident is the person presenting it that the assumption holds. That question takes under a minute to ask and under a minute to answer, and it does the same work the old tells used to do, surfacing the gap between how a recommendation sounds and how solid it actually is, before the room commits to acting on it.</p><p>This does not require a new process on top of an already full calendar. It fits inside a step most product organizations already have and mostly skip: the moment right before a document leaves draft status and enters review. Add the confidence line to whatever template already exists there, a PRD header, a one-pager’s opening block, a slide’s speaker notes, so stating it becomes part of finishing the document rather than an extra conversation someone has to remember to start. Within a few cycles, the absence of that line will start to read the way an unfinished section used to read: a visible gap, not a small omission. That is the point at which the habit has actually taken hold, when the room notices its absence faster than it noticed its arrival.</p><p>This will feel like friction at first, mostly because it is friction, a deliberate slowing of a process that AI tools have made faster in every other respect. Add it anyway. The forecasts, the analyses, and the recommendations that AI tools now help produce are not going to slow down, and neither is the pace at which teams are asked to act on them. The only lever available to a team that wants the speed without the blind spot is to name where the confidence actually lives, on purpose, every time, until it becomes as automatic as the fluency it is meant to check.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=8d510373bbe9" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Generated Doesn’t Mean Owned.]]></title>
            <link>https://stephanie-muxfeld.medium.com/generated-doesnt-mean-owned-4b7b544c619b?source=rss-697b881c9471------2</link>
            <guid isPermaLink="false">https://medium.com/p/4b7b544c619b</guid>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[career-advice]]></category>
            <category><![CDATA[product-management]]></category>
            <category><![CDATA[future-of-work]]></category>
            <category><![CDATA[leadership]]></category>
            <dc:creator><![CDATA[Stephanie Muxfeld]]></dc:creator>
            <pubDate>Mon, 31 Aug 2026 12:49:12 GMT</pubDate>
            <atom:updated>2026-08-31T12:49:12.970Z</atom:updated>
            <content:encoded><![CDATA[<p>A model can draft the sentence. It can’t decide to send it.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*B5oG9m3MifvtivokD3S3hw.png" /></figure><p>Only a few people are on the call. It’s a launch review. The kind that always runs long. Someone asks who approved the line in the customer email that went out an hour early, to the wrong distribution list. A pause. Then the same four words, offered by three different people, as if agreed on in advance. <em>The tool did it.</em></p><p>Nobody disagrees. Nobody has to. The answer closes the loop and the meeting moves on to the next item, and a real mistake becomes a small footnote instead of an open question. This is not the first time this exact answer has ended this exact kind of conversation. It will not be the last.</p><blockquote><em>Generation is not ownership. It’s the draft before the signature, not the signature itself.</em></blockquote><p>A model can produce the sentence, the summary, the recommendation, the whole first pass. What it cannot do is decide to send it. Somewhere between the draft and the outbound email, a person looked at the words on the screen and pressed send, or nodded, or simply let the automation run without stopping it. That decision doesn’t disappear because the sentence it acted on came from somewhere else.</p><p>This isn’t a hypothetical the law is still catching up to. In August 2026, the European Union’s Artificial Intelligence Act reached full enforcement, and its requirements are specific about exactly this gap. High-risk systems must log their actions for traceability, and they must remain under meaningful human oversight, not nominal oversight, actual review by an actual person.¹ Regulators didn’t write that requirement because oversight was already happening reliably. They wrote it because it wasn’t.</p><p>The instinct to treat a capable model as something closer to a colleague makes this worse, not better. Research published in Harvard Business Review this year, led by Boston Consulting Group’s Matthew Kropp, found that giving artificial intelligence (AI) agents employee-like framing inside an organization reduced individual accountability and lowered the quality of human review, without improving how readily people actually used the tool.² A separate part of the same research found the framing increased unnecessary escalation instead. The framing meant to make responsibility clearer made it quieter instead. Call something a teammate, and people start treating its output the way they’d treat a teammate’s: trusted by default, and rarely questioned until something has already gone wrong.</p><p>Courts are already drawing the line the other way. A German court ruled in June 2026 that a major technology company was liable for the content of its own AI-generated search summaries, rejecting the argument that users should simply know not to trust them.³ The court treated the summaries as belonging to the company itself, not as an independent, disclaimable voice standing apart from it. Gartner projects that by mid-2026, new categories of unlawful AI-informed decisions will generate more than ten billion dollars in remediation costs across enterprises and vendors combined.⁴ That isn’t the cost of AI being wrong. That’s the cost of nobody having been willing, in advance, to say whose decision it actually was.</p><p>Naming who owns a result, especially when a model did the drafting, is not a compliance exercise. It’s the difference between a team that can explain its decisions and one that can only point at a tool. A compliance exercise produces a document nobody reads until something breaks. Real ownership looks closer to what the Singaporean telecommunications company Singtel has started doing this year: restructuring parts of the organization so specific leaders are explicitly responsible for blended teams of people and AI agents, not the software vendor, not “the platform,” a named person.⁵ Singapore’s Infocomm Media Development Authority built the same principle into its own governance framework for agentic AI, naming humans staying “meaningfully accountable” as one of its four core pillars, precisely because autonomy has a way of blurring who’s actually answerable once it gets folded into a routine workflow.⁶</p><p>None of this means slowing down, or treating every drafted paragraph like a legal filing. It means one habit, practiced early enough that it never has to be reconstructed after the fact: before something goes out, someone specific decides to stand behind it, and that decision gets said out loud, not assumed. Not the department. Not “the team.” A name.</p><blockquote><em>A signature was never really about the pen. It was about the hand that stayed willing to be found later, when someone finally asked who wrote this.</em></blockquote><p>Whose hand is on the last thing your team shipped?</p><p><strong>Sources &amp; References</strong></p><ol><li>European Union Artificial Intelligence Act, high-risk system obligations under Articles 12 and 14 (logging and human oversight), enforcement activated August 2, 2026.</li><li>Matthew Kropp et al., “Research: Why You Shouldn’t Treat AI Agents Like Employees,” <em>Harvard Business Review</em>, May 6, 2026.</li><li>German court ruling on liability for AI-generated search summary content, reported June 2026.</li><li>Gartner, projection on remediation costs from unlawful AI-informed enterprise decision-making, cited 2026.</li><li>“Your newest employee is an agent. Who is accountable?” <em>TNGlobal</em>, August 28, 2026 (Singtel organizational restructuring for blended human-AI teams).</li><li>Singapore Infocomm Media Development Authority, Model AI Governance Framework for Agentic AI, 2026.</li></ol><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=4b7b544c619b" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The Room You Give a Decision]]></title>
            <link>https://stephanie-muxfeld.medium.com/the-room-you-give-a-decision-3571915b1765?source=rss-697b881c9471------2</link>
            <guid isPermaLink="false">https://medium.com/p/3571915b1765</guid>
            <category><![CDATA[deadlines]]></category>
            <category><![CDATA[organizational-drift]]></category>
            <category><![CDATA[parkinsons-law]]></category>
            <category><![CDATA[decision-making]]></category>
            <category><![CDATA[leadership-judgment]]></category>
            <dc:creator><![CDATA[Stephanie Muxfeld]]></dc:creator>
            <pubDate>Mon, 24 Aug 2026 11:31:01 GMT</pubDate>
            <atom:updated>2026-08-24T11:31:01.338Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Ww1IAdFnJXvdST_ByavoWA.png" /></figure><h4>Why generosity with time doesn’t always produce better work</h4><p>Someone asks when the roadmap decision needs to be locked. The answer comes back easy and well-meaning: “No rush, take the time you need to get it right.” The meeting ends. Everyone feels good. Nothing gets decided that week. Or the next.</p><p>Six weeks later the same question surfaces in a different meeting, still open, now wrapped in more context, more opinions, more reasons it’s complicated. Nobody lied. Nobody dropped the ball. Nobody was lazy. The time just did what time does when nothing is pushing against it.</p><p>By the time it finally gets resolved, the decision looks almost nothing like what it would have looked like in week one. Not because six weeks of research changed the fundamentals. Usually the fundamentals were visible early. What changed is the number of people who weighed in, the number of alternatives that got floated and half-explored, the number of times someone said “let’s just make sure we’re not missing anything” before circling back to roughly the original instinct, three months later, with a much longer document attached to it.</p><p><strong>A joke that held up</strong></p><p>Cyril Northcote Parkinson wrote, mostly as a joke, that work expands to fill the time available for its completion.¹ He was a British naval historian writing for a general audience in 1955, poking fun at the way a retired person can spend an entire morning writing a single postcard: finding the card, hunting for glasses, checking the address, drafting and redrafting a sentence or two, deciding whether to bring a coat to the mailbox. An hour of actual effort, stretched across a day because a day was available.</p><p>The joke has held up better than most serious management theory. A task with two weeks of real content takes two weeks if given two weeks, and four if given four, not because the work grew, but because the time did.</p><p><strong>Generosity that isn’t generous</strong></p><p>What makes this worth taking seriously isn’t the observation itself. It’s what it implies about a common instinct in how leaders try to help their teams: giving more time, more scope, more room, as an act of generosity.</p><p>The instinct comes from a reasonable place. Constraints feel like pressure. Pressure feels like something to remove wherever possible, especially for a leader who remembers what it felt like to be squeezed by an unreasonable deadline earlier in their own career and swore they’d do it differently. Generosity with time can feel like the corrective. It can feel like trust: <em>I’m not going to breathe down your neck about this, take what you need.</em></p><p>But removing a constraint doesn’t create more capacity. It creates more room for a task to expand into, and expansion isn’t the same thing as improvement.</p><p>The clearest version of this shows up in product timelines. A team given a tight, honest deadline tends to make faster decisions, because indecision has a visible cost: less time to build whatever gets decided. Every day spent debating is a day not spent building, and that tradeoff is felt immediately. A team given an open-ended timeline tends to defer decisions longer, because indecision has no visible cost yet.² Nothing is lost today by waiting one more week. Nothing is lost next week either. The cost only becomes visible in aggregate, much later, by which point it’s hard to trace back to any single choice to wait.</p><p>The work doesn’t get better in the extra time. It gets debated longer, revised more, and finished later, often to a similar or lesser standard than it would have reached under real pressure. This is a strange thing to sit with, because it runs against the instinct that more deliberation should produce better outcomes. Sometimes it does. But past a certain point, additional time mostly produces additional versions of a decision that was already essentially available, dressed in more caveats.</p><p><strong>Why self-imposed deadlines work anyway</strong></p><p>There’s a name for the shape this takes. Research on pre-commitment shows that people who set their own deadlines, even self-imposed ones with no external enforcement, perform measurably better than people given unlimited time, and better than people who set deadlines but space them too generously.³ The finding is counterintuitive on its face. Why would a deadline you invented yourself, one you could ignore without consequence, actually change your output? The answer seems to be that a deadline forces a choice that would otherwise keep sliding. It’s not really about the clock. It’s about converting an open question into a closed one, at a specific point in time, by design rather than by accident.</p><p>This isn’t an argument for arbitrary deadlines or constant urgency, which produce their own failure mode: burnout, corner-cutting, work rushed past the point where more time would have genuinely improved it. Anyone who has shipped something six weeks early only to spend the next six months fixing what got skipped knows that failure mode intimately. Compression has real costs, and pretending otherwise is its own kind of dishonesty.</p><p>The point is narrower than “always move faster.” Time and scope aren’t neutral resources that simply get consumed by however much work exists. They actively shape how much work gets created, debated, and revisited. An open calendar doesn’t produce clarity. It produces more surface area for uncertainty to occupy. Give a question six weeks instead of six days and you haven’t given it more time to be answered correctly. You’ve given it more time to accumulate stakeholders, alternatives, and reasons to keep it open.</p><p><strong>Diligence and avoidance look the same from the inside</strong></p><p>The practical implication isn’t to constrain everything ruthlessly. It’s to notice when generosity with time is actually functioning as an escape from making a decision, for you or for the people you lead.</p><p>That distinction matters, because from the inside, the two things feel identical. Genuine thoroughness and comfortable avoidance both look like “we’re still working through it.” Both produce activity. Both can be narrated, honestly, as diligence. The difference usually only becomes visible in hindsight, when you can see that the extra five weeks produced a slightly longer document but not a materially different decision.</p><p>A deadline isn’t always pressure imposed from outside. Sometimes it’s the thing that forces a choice that would otherwise keep getting deferred indefinitely, dressed up as diligence. Deferred decisions don’t feel like avoidance while they’re happening.⁴ They feel like thoroughness. That’s what makes them so easy to keep extending, and so hard to notice you’re extending, because every individual extension has a reasonable justification attached to it. One more round of feedback. One more stakeholder to loop in. One more data point that might change things. Each step is defensible. The pattern, seen from above, is something else.</p><p><strong>The permission underneath this</strong></p><p>If you’ve been the leader handing out extra time as a kindness, this isn’t a case against kindness. It’s permission to ask a sharper question before you offer more room: does this task actually need more time, or does it need a decision, and is “more time” the thing that’s letting everyone, including you, avoid making one.</p><p>That question can feel uncomfortable to ask, because it risks looking like you’re taking something away. It can help to notice that the opposite is usually true. A team drifting inside an open-ended timeline isn’t experiencing freedom. It’s experiencing an absence of the one thing that would let them stop debating and start building. Naming a real constraint, an honest one, tied to something that actually matters, isn’t pressure for its own sake. It’s often the most direct form of relief available. It tells everyone involved: this is where the debate ends, and now we build something instead of continuing to discuss what we might build.</p><blockquote>The room you give a piece of work isn’t a passive container. It’s an active ingredient in what that work becomes.</blockquote><p><strong>Sources &amp; References</strong></p><ol><li>Parkinson, C. N. (1955). Parkinson’s Law. The Economist.</li><li>Goldratt, E. M. (1997). Critical Chain. North River Press.</li><li>Ariely, D., &amp; Wertenbroch, K. (2002). Procrastination, Deadlines, and Performance: Self-Control by Precommitment. Psychological Science, 13(3), 219–224.</li><li>Tversky, A., &amp; Shafir, E. (1992). Choice Under Conflict: The Dynamics of Deferred Decision-Making. Psychological Science, 3(6), 358–361.</li></ol><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=3571915b1765" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>