<?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 Ryn Bennett on Medium]]></title>
        <description><![CDATA[Stories by Ryn Bennett on Medium]]></description>
        <link>https://medium.com/@ksbennet?source=rss-f5e05dd1b554------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*_4wfc0ap7-6THkt8vVTfPA@2x.jpeg</url>
            <title>Stories by Ryn Bennett on Medium</title>
            <link>https://medium.com/@ksbennet?source=rss-f5e05dd1b554------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Sat, 21 Nov 2026 21:46:49 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@ksbennet/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[How to Sell AI Maturity as a Competitive Advantage]]></title>
            <link>https://medium.com/@ksbennet/how-to-sell-ai-maturity-as-a-competitive-advantage-4a5fb833ff2b?source=rss-f5e05dd1b554------2</link>
            <guid isPermaLink="false">https://medium.com/p/4a5fb833ff2b</guid>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[sales-enablement]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[technology]]></category>
            <category><![CDATA[sales]]></category>
            <dc:creator><![CDATA[Ryn Bennett]]></dc:creator>
            <pubDate>Thu, 18 Jun 2026 21:01:22 GMT</pubDate>
            <atom:updated>2026-06-18T21:01:22.397Z</atom:updated>
            <content:encoded><![CDATA[<h4><em>AI maturity is worth nothing in a proposal unless you can prove it. Here’s what proof actually looks like.</em></h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Esdmcfze7set9Yy5fKTu-w.png" /></figure><p>A shift is happening in how buyers evaluate the firms they work with. “How do you use AI?” is moving from a curiosity at the end of a meeting to a real factor in who gets selected.</p><p>That creates an opportunity and a trap at the same time. The opportunity is that demonstrable AI maturity is becoming a genuine differentiator.</p><p>The trap is that the firms quickest to <em>claim</em> it are often the ones least able to <em>back</em> it, and an AI-maturity claim that collapses under one good question does more damage than saying nothing at all.</p><p>If you want AI maturity to win you work rather than expose you, two things have to be true: you need to know which kind of maturity actually matters to a buyer, and every claim you make has to have evidence sitting directly behind it.</p><h3>Three lenses buyers actually care about</h3><p>“We use AI” is not a value proposition. It’s noise. Buyers don’t care that you use AI; they care what it does for <em>them</em>. There are three distinct places AI maturity shows up as something a customer or evaluator can feel, and it’s worth being deliberate about which one you’re selling.</p><ul><li><strong>Win</strong> — the front of the business. AI maturity here means sharper, faster, more tailored proposals and capture work: better synthesis of requirements, more responsive iterations, more time spent on strategy because less is spent on assembly. This is the lens that shows up <em>in the very document the buyer is evaluating, </em>which makes it the most immediately persuasive, and the most exposed if your “AI-enabled” proposal is generic.</li><li><strong>Deliver</strong> — the work itself. AI maturity in delivery means faster reporting, better knowledge synthesis across a program, more consistent documentation, more of your experts’ time spent on judgment rather than assembly. This is what a customer experiences <em>after</em> they’ve hired you, which makes it the lens that drives renewals and references.</li><li><strong>Run</strong> — the back office. AI maturity in how you run HR, finance, IT, and knowledge management rarely sells a deal on its own, but it’s the foundation that makes the other two credible and sustainable. A firm that’s chaotic internally can’t deliver disciplined AI maturity externally for long.</li></ul><p>The discipline is to sell the lens that matters to <em>this</em> buyer for <em>this</em> pursuit (usually Win and Deliver) rather than gesturing vaguely at “AI across the organization.”</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*-rR282FBNs9c2Cb56hqkDA.png" /></figure><h3>The claim has to survive scrutiny</h3><p>Firms get hurt when they sense that AI maturity sells, and they reach for it in capability statements and proposals with nothing underneath. Buzzwords like “AI-enabled,” “AI-ready,” “leveraging AI to deliver” don’t mean anything on their own. And the moment an evaluator or customer probes, an unsupported claim doesn’t just fail on its own; it taints everything else you said. Now they’re wondering what <em>else</em> in your proposal was aspirational.</p><p>The protective discipline is simple to state and harder to live: <strong>never make an AI-maturity claim you can’t immediately substantiate.</strong> Turn each claim into a question and make sure you have the answer ready:</p><ul><li><strong>If you say AI-assisted work is reviewed, </strong>describe the review. Who validates it, against what?</li><li><strong>If you say you protect customer data,</strong> point to the data classification and the tooling that enforces it.</li><li><strong>If you say AI makes your delivery faster, </strong>name the workflow, and ideally the before-and-after.</li><li><strong>If you say you govern AI use, </strong>there should be an actual governance position someone can read.</li></ul><p>A claim with evidence behind it reads as maturity. A claim without it reads as marketing. Experienced buyers can tell the difference in one follow-up question.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*808NldHqH8JELYKcyLEUuw.png" /></figure><h3>What “proof” looks like in practice</h3><p>You don’t need a transformation program to be credibly AI-mature. You need a handful of things you can <em>show</em>: a real example of AI-accelerated work with the human review that made it trustworthy; a clear data-handling posture you can speak to without hedging; a governance position, even a one-page one, that names who’s accountable. These are concrete, demonstrable, and — crucially — honest. They let you make strong claims because the claims are true.</p><p>That honesty is itself a competitive advantage. In a market filling up with inflated AI claims, the firm that says “here’s specifically what we do, here’s the evidence, and here’s where we keep a human in the loop” stands out precisely because it isn’t overselling. Sophisticated buyers are getting better at sniffing out AI theater. Being the credible voice in a noisy field is a position worth occupying.</p><h3>The bottom line</h3><p>AI maturity is becoming a discriminator, but a discriminator is only worth something if it holds up. The firms that win with it are the ones who pick the lens that matters to the buyer, make only the claims they can back, and treat demonstrable evidence — not enthusiasm — as the currency. Sell what you can prove, prove what you sell, and let the firms making hollow claims discredit themselves.</p><p><em>Helping firms turn a vague “we use AI” into a substantiated maturity story that wins evaluations — and survives the follow-up questions — is core to what I do. If that’s a gap between your capability and how you’re selling it, </em><a href="https://www.linkedin.com/in/rynforthewin/"><em>let’s talk!</em></a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=4a5fb833ff2b" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[AI Doesn’t Have to Be Perfect. It Has to Be Better Than Us.]]></title>
            <link>https://medium.com/@ksbennet/ai-doesnt-have-to-be-perfect-it-has-to-be-better-than-us-6acd546e9f85?source=rss-f5e05dd1b554------2</link>
            <guid isPermaLink="false">https://medium.com/p/6acd546e9f85</guid>
            <category><![CDATA[generative-ai-tools]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[business-strategy]]></category>
            <category><![CDATA[management-and-leadership]]></category>
            <category><![CDATA[automation]]></category>
            <dc:creator><![CDATA[Ryn Bennett]]></dc:creator>
            <pubDate>Sun, 14 Jun 2026 19:09:34 GMT</pubDate>
            <atom:updated>2026-06-14T19:09:34.128Z</atom:updated>
            <content:encoded><![CDATA[<h4><em>The most common objection to AI is also the least examined: “but it makes mistakes.” So do you. The real question was never perfection.</em></h4><p>There’s a move that happens in almost every executive conversation about adopting AI. Someone runs a tool, it gets something wrong, like a figure, a confident misstatement, an off-base recommendation — and the room exhales with a kind of relief.</p><p><em>See? It’s not reliable.</em></p><p>Watch that moment closely, because a subersive substitution just happened.</p><blockquote>The standard being applied to the AI is <strong>perfection</strong> — zero errors, full trust, no oversight required. And nothing in your organization is held to that standard.</blockquote><p>Not your best analyst, not your most senior manager, not you. We’ve never required perfection from a human contributor. We require that they be good enough to be worth it, and we build review around the places they’re likely to slip.</p><p>Somehow, when the contributor is software, we forget the entire framework we already use for fallible humans and demand something we’ve never demanded from anyone.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Iz04wwHBAkTx2exx_n15gw.png" /></figure><h3>The benchmark is the human baseline, not zero</h3><p>The right question about any AI application is not “is it perfect?” It’s “is it better than the realistic alternative — and at what cost?”</p><p>The realistic alternative is almost never a flawless process. It’s a busy person doing the task at 4 p.m. on a Thursday, with incomplete information, competing priorities, and the same human tendencies toward error, bias, and fatigue that we’ve spent careers building checks around.</p><p>Measured against the actual baseline, not an imagined ideal —and a lot of AI applications clear the bar comfortably.</p><ul><li>A first draft that’s 80% there in thirty seconds beats a blank page and an hour.</li><li>A classification that’s right 92% of the time, with the 8% flagged for review, beats a human doing it inconsistently and unreviewed.</li><li>The error rate that feels disqualifying in the abstract is often <em>lower</em> than the error rate of the process it would replace — we just never measured the human version, because we never thought to.</li></ul><p>Perfection was always a strawman. The investment question is, and has always been, relative: better than the alternative, at acceptable cost, with the failures in manageable places.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*h4vwljL7QBszjkVI3MaArw.png" /></figure><h3>Where the bar is genuinely higher — and why that’s not a contradiction</h3><p>This is where an honest version of the argument has to slow down, because “better than a human is good enough” is not universally true, and pretending it is would be exactly the kind of overconfidence that gets organizations burned.</p><p>The “better than human” bar is the right one for work where errors are <em>visible, recoverable, and reviewed</em> — drafting, synthesis, triage, first passes.</p><p>It is the wrong bar where errors are <em>difficult to spot, compounding, or catastrophic</em>.</p><p>If a mistake won’t be caught before it does damage, or if it scales instantly across thousands of cases, or if the answer must be authoritative and exact, then “usually better than a person” isn’t sufficient — you need either deterministic tooling or a human firmly in the loop.</p><blockquote>The benchmark isn’t lower for AI; it’s the same benchmark you’d apply to a human in a high-stakes seat, which is precisely why you don’t let a junior analyst wire money unsupervised either.</blockquote><p>So the real discipline isn’t “trust it” or “don’t trust it.” It’s matching the bar to the stakes: let AI clear the human baseline where errors are caught and cheap, and hold it to a much higher standard — or keep it out — where they aren’t. That’s not a knock against AI. It’s the same risk calibration you already apply to people.</p><h3>Why the perfection reflex is so expensive</h3><p>The reason this matters isn’t philosophical. Holding AI to a standard you hold nothing else to has a real cost, and it compounds.</p><p>While one organization kills a useful tool over a single visible error, another</p><ul><li>deploys it against the right baseline</li><li>captures the productivity</li><li>builds the review muscle</li><li>and gets better at deploying the next one.</li></ul><p>A year later the gap between them isn’t about who had access to better technology. It’s about who used a coherent standard for judging it.</p><p><strong>The perfection reflex feels like prudence. Often it’s just an unexamined double standard wearing the costume of rigor.</strong></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*ie0BjEC9DB27i3JgpWIblQ.png" /></figure><h3>The question worth asking instead</h3><p>Next time a tool gets something wrong in a demo and the room starts to exhale, ask the real question before the verdict lands: <em>better than what, exactly?</em></p><ol><li>Better than our current process, honestly measured?</li><li>Are the errors in places we can catch and recover from?</li><li>Is the cost of being wrong here acceptable?</li></ol><p>If yes on all three, the single mistake everyone just fixated on isn’t a reason not to invest. It’s a reason to invest — with review where it belongs.</p><p>We were never going to get perfect, from software or from people. The bar was always “better than the alternative, at acceptable cost.” Apply that bar evenly, and most of the AI conversation stops being about fear of mistakes and starts being about where the leverage actually is.</p><p><em>Helping leadership teams judge AI against the right baseline — and build the review discipline that makes “better than human” safe to act on — is core to the work I do with executives. If your team is stuck arguing about perfection, </em><a href="https://www.linkedin.com/in/rynforthewin"><em>let’s talk.</em></a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=6acd546e9f85" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[AI vs. Automation: The Distinction Most Executives Get Wrong]]></title>
            <link>https://medium.com/@ksbennet/ai-vs-automation-the-distinction-most-executives-get-wrong-81cf85191d25?source=rss-f5e05dd1b554------2</link>
            <guid isPermaLink="false">https://medium.com/p/81cf85191d25</guid>
            <category><![CDATA[management-and-leadership]]></category>
            <category><![CDATA[digital-transformation]]></category>
            <category><![CDATA[workflow-automation]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[automation]]></category>
            <dc:creator><![CDATA[Ryn Bennett]]></dc:creator>
            <pubDate>Sun, 14 Jun 2026 18:46:18 GMT</pubDate>
            <atom:updated>2026-06-14T18:46:18.050Z</atom:updated>
            <content:encoded><![CDATA[<h4><em>In a world of high token costs, reaching for AI when you don’t need it is embarrassingly expensive.</em></h4><p>Walk into most leadership conversations about AI and you’ll hear “AI” and “automation” used as if they’re the same project with two names. They aren’t. Treating them as interchangeable is the single most expensive conceptual mistake I see executives make — and I mean <em>expensive</em> in two distinct ways.</p><p>First, this misconception leads to pointing each tool at exactly the work it’s worst at: using AI where you needed reliable automation, and writing rigid automation where you needed judgment. That costs you quality. But in a world where every AI call carries a token price and an energy cost, pointing the wrong tool at the work now also runs a meter — and the bill scales with exactly the high-volume, repetitive work AI should never have touched.</p><p>Just like a human brain, AI technology is expensive and well-tailored to specific activities. Don’t use it to do stupid, routine tasks.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*OgoRALaCQzGNE9HGKMNNtA.png" /></figure><h3>The line that matters: deterministic vs. non-deterministic work</h3><p>Strip away the vocabulary and there’s one question underneath all of it: D<em>oes this task have a single correct answer that must come out the same way every time?</em></p><p>If yes, the work is <strong>deterministic</strong>. An indirect rate calculation, a compliance control status, a tax figure, an inventory count, a lookup against a system of record — these have a right answer, and “usually right” is failure. Deterministic work wants <strong>automation</strong>: fixed rules executing repeatable steps, the same input producing the same output, every time, auditable — and, once built, running at essentially zero marginal cost.</p><p>If no, the work is <strong>non-deterministic</strong>.</p><ul><li>Drafting a first version of a board narrative, synthesizing twenty documents into themes</li><li>Classifying messy inbound requests</li><li>Turning a rambling intent into a structured outline</li></ul><p>For these tasks, there’s no single correct output, just better and worse ones, and the task requires applying judgment to language and ambiguity. That’s what <strong>AI</strong> is for, and it’s worth paying for precisely because nothing cheaper does it.</p><blockquote>Automation and AI both fail badly when pointed at the wrong kind of work.</blockquote><p>Ask a large language model for an authoritative number and it will give you one that <em>sounds</em> right — confidently, fluently, and sometimes completely invented. That’s not a bug you can prompt your way out of; it’s what the tool is. A language model is a probabilistic next-token engine, not a database, a calculator, or a system of record. It predicts plausible language, and for deterministic work “plausible” is precisely the wrong target. Run it the other way and you get the opposite failure: rigid automation on genuinely ambiguous work becomes a brittle decision tree that shatters the moment reality doesn’t match the branches you anticipated.</p><h3>The cost most teams don’t see until the bill arrives</h3><p>When you put AI on deterministic work, you pay a premium for wrong answers, every single time, forever.</p><p>Automation has a cost profile that bends in your favor: real money to build once, then essentially free to run a million times. The marginal cost of the next execution rounds to zero, and so does its energy footprint.</p><p>AI inverts that. Every call is metered — tokens in, tokens out — and that cost recurs on every single run, at a compute and energy price that doesn’t amortize away no matter how many times you do it. Worse, deterministic work is usually <em>high-volume and repetitive</em>. That’s the exact shape of work where a per-call meter hurts most.</p><p><strong>So putting AI on deterministic work is the worst trade available:</strong> You take a task a script would do once for free, and you instead pay a recurring token-and-energy bill to get an answer that’s also less reliable. You’re buying variability <em>and</em> a meter where you wanted certainty <em>and</em> zero marginal cost. In a low-token-cost world you could ignore that. You can’t anymore. Token spend has become a real line item, and the energy behind it a real exposure. Both scale behind-the-scenes with adoption until someone asks why the AI bill tripled.</p><p>The discipline cuts the other way too: Because AI is the expensive tool, you want to <em>spend</em> it deliberately on the judgment work where it’s the only thing that does the job, not burned on lookups a deterministic system should own.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*P-PzWUaRyMKtEMrOwTbG0g.png" /></figure><h3>The simple test you can run in a meeting</h3><p>Take any task your team is considering “putting AI on” and ask: <em>if two careful people did this, would I expect identical output?</em></p><ul><li><strong>Identical expected</strong> → deterministic → automation problem. Reach for AI here and you’re paying per token, on every run, to inject variability into work that demanded consistency.</li><li><strong>Reasonably different expected</strong> → non-deterministic → this is where AI earns its keep, and where the token cost actually buys you something nothing else can.</li></ul><p>It’s a thirty-second test, and it reorganizes a roadmap <em>and</em> a budget fast.</p><p>Most teams discover that half of what they’d filed under “AI opportunity” is actually plain automation they’ve been putting off, which means they were about to pay premium token rates for work a cheap script should own.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*yQb9jGEjGv1iYCn87A0DQg.png" /></figure><h3>Hybrid systems are also the cheapest design</h3><p>The best designs usually aren’t AI <em>or</em> automation. They’re a pipeline that uses each for what it’s good at, which turns out to be the cost-optimal design as well as the highest-quality one.</p><p>Picture a recurring report.</p><ol><li>Automation reliably pulls the figures from the systems of record, which are deterministic, must be exact, no judgment, and free to run.</li><li>Then AI drafts the narrative around those figures — non-deterministic, judgment-heavy, a first draft a human sharpens.</li></ol><p>The numbers are trustworthy because a deterministic system produced them; the prose is fast because a probabilistic one drafted it. And the token spend stays small and justified, because you only spent it on the one step that actually needed judgment.</p><blockquote>Letting the expensive tool do the cheap tool’s job and running everything through the model because it’s easy is how token bills balloon for work that never required a model at all.</blockquote><p>The seam, where deterministic data feeds non-deterministic synthesis, is where the durable value lives <em>and</em> where the cost stays disciplined.</p><h3>Why this is a leadership distinction, not a technical one</h3><p>It’s tempting to file all of this under “implementation detail for the technical team.” It isn’t. Deciding which work is deterministic and which is judgment is a decision about where you accept variability, where you demand certainty, and where you’re willing to run a recurring token-and-energy meter and where you refuse to.</p><p>That’s a leadership call with direct consequences for risk, trust, <em>and margin</em>. The teams that make it deliberately point each tool at the work it’s good at, and keep their AI spend concentrated where it pays for itself. The teams that blur the line spend a year disappointed by AI for failing at jobs it was never right for — and quietly overpay for the failure the whole time.</p><p>So before the next “let’s use AI for this” conversation, ask the only question that sorts it: D<em>oes this need to be right every time, or does this need good judgment applied to ambiguity?</em> The answer tells you which tool you’re actually reaching for.</p><p>In a high-token-cost world, you need to know whether you’re about to invest in leverage or just light money and energy on fire.</p><p><em>Helping a leadership team sort their real AI opportunities from their automation backlog — and stop paying premium token rates for work a script should own — is usually the first thing I do with them. If that clarity would help yours, </em><a href="https://www.linkedin.com/in/rynforthewin/"><em>let’s talk.</em></a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=81cf85191d25" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[AI Won’t Transform Your Company. Your Leadership Team Will.]]></title>
            <link>https://medium.com/@ksbennet/ai-wont-transform-your-company-your-leadership-team-will-378f823fd75f?source=rss-f5e05dd1b554------2</link>
            <guid isPermaLink="false">https://medium.com/p/378f823fd75f</guid>
            <category><![CDATA[generative-ai-use-cases]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[decision-making]]></category>
            <category><![CDATA[management-and-leadership]]></category>
            <dc:creator><![CDATA[Ryn Bennett]]></dc:creator>
            <pubDate>Wed, 10 Jun 2026 16:46:13 GMT</pubDate>
            <atom:updated>2026-06-10T16:46:13.943Z</atom:updated>
            <content:encoded><![CDATA[<h3><em>Every executive team I work with wants to know what AI will do to their business. It’s the wrong question.</em></h3><p>There’s a particular kind of meeting happening in conference rooms right now. A leadership team gathers, someone presents the AI landscape, and the room fills with a mix of anxiety and anticipation — as if a wave is approaching and the only decision left is whether to brace or to surf. The framing underneath it is that AI is a force arriving <em>at</em> the organization, and the job of leadership is to react to it well.</p><p>That framing is comfortable, and it’s wrong. It removes the leadership team from the story at exactly the moment they matter most.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*e-ZKGds5ofr3Q40T_Upg4g.png" /></figure><p>Here is the more accurate version: AI is not going to transform your company. The decisions your leadership team makes over the next few quarters are going to transform your company. AI is leverage on those decisions — it amplifies the quality of the choices you make about where to apply it, who owns it, and what you’re willing to change to use it well. Point that leverage at a clear decision and it compounds. Point it at a vague aspiration and it amplifies the vagueness.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*UBqMOz5iw96fP_akCENcdQ.png" /></figure><h3>Why “what will AI do to us” produces bad strategy</h3><p>When a team treats AI as an external force, three predictable things happen.</p><p>First, they over-index on the technology and under-index on themselves. The conversation becomes about models, tools, and features — which are changing weekly and which you don’t control — instead of about workflows, decisions, and people, which you do control and which actually determine whether any of this creates value.</p><p>Second, ownership evaporates. “The market is moving to AI” is a statement with no subject. Nobody in the room is accountable for a trend. The moment you restate it as “we will decide, by the end of the quarter, whether to standardize on one enterprise tool and who owns that rollout,” you have something a named person can be responsible for. Strategy lives in the second sentence, not the first.</p><p>Third, teams confuse activity with transformation. A dozen people quietly using a chatbot is not a transformed organization; it’s shadow adoption with no compounding return and a growing data-exposure surface. Real transformation shows up as changed decisions, changed processes, and changed economics — and those only happen when leadership chooses them on purpose.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*NrSuYK6V_psRIN2Pk3FO9A.png" /></figure><h3>The leverage framing changes what you do on Monday</h3><p>If AI is leverage on decisions, then the work of leadership is not to forecast the technology. It’s to get specific about the decisions worth amplifying. That’s a very different and much more answerable set of questions:</p><p>Where in our business is the bottleneck a <em>judgment</em> problem — synthesis, drafting, classification, turning messy intent into structure — rather than a problem that needs a deterministic, always-correct answer? That’s where AI actually helps, and naming it is a leadership call, not a technical one.</p><p>What are we willing to change to capture the value? A tool dropped on top of an unchanged process returns almost nothing. The transformation is in the willingness to redraw the workflow, retire the redundant tool, and rewrite the expectation — and only the leadership team can authorize that.</p><p>Who owns it, and by when? Every real initiative has a name and a date attached. “We’re exploring AI” has neither, which is how you can tell it isn’t real yet.</p><h3>What this asks of leaders</h3><p>The uncomfortable part of this framing is that it doesn’t let anyone off the hook. You can’t delegate the transformation to a tool or to a vendor or to “the team,” because the decisions that matter — what to standardize, what to change, what to govern, what to bet on — are exactly the decisions that can only be made in the leadership room. AI raises the ceiling on what those decisions can achieve. It does nothing to make the decisions for you, and it never will.</p><p>That’s not a burden. It’s the most empowering thing about this moment. The companies that pull ahead over the next few years won’t be the ones that adopted the most tools or moved the fastest to chase features. They’ll be the ones whose leadership treated AI as what it is — a multiplier on the quality of their own judgment — and then did the harder, more valuable work of improving the decisions worth multiplying.</p><p>The wave isn’t coming for you. You’re holding the lever. The only question that matters is what you choose to point it at.</p><p><em>I work with leadership teams to turn “we should do something with AI” into a specific, owned set of decisions with dates attached. If that’s the conversation your team needs to have, let’s talk. Hit me up on </em><a href="https://www.linkedin.com/in/rynforthewin/"><em>LinkedIn</em></a><em> or at ryn@surefirepanel.com.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=378f823fd75f" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The 5-Step Loop I Use to Stress-Test Executive Judgment With AI]]></title>
            <link>https://medium.com/@ksbennet/the-5-step-loop-i-use-to-stress-test-executive-judgment-with-ai-54e733d9d378?source=rss-f5e05dd1b554------2</link>
            <guid isPermaLink="false">https://medium.com/p/54e733d9d378</guid>
            <category><![CDATA[decision-making]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[management]]></category>
            <category><![CDATA[executive-coaching]]></category>
            <category><![CDATA[ai]]></category>
            <dc:creator><![CDATA[Ryn Bennett]]></dc:creator>
            <pubDate>Tue, 09 Jun 2026 12:45:50 GMT</pubDate>
            <atom:updated>2026-06-09T12:45:50.030Z</atom:updated>
            <content:encoded><![CDATA[<h3><em>The best use of AI in the C-suite isn’t getting answers. It’s having your own thinking argued with.</em></h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*01lUJZkIbF9M1Th-KCsiOQ.png" /></figure><p>Most executives I work with reach for AI the same way they’d reach for an analyst: They ask it a question and wait for the answer. That instinct is understandable, and it’s also the single biggest reason AI underdelivers at the leadership level.</p><h4>Senior decisions are rarely limited by a lack of answers.</h4><p>They’re limited by unexamined assumptions, by the blind spots that come with experience, and by the quiet comfort of a conclusion you reached before you’d finished thinking.</p><p>A large language model is genuinely bad at giving you a trustworthy answer to a high-stakes judgment call. It will assert specifics it doesn’t have, it will flatter the position you walked in with, and it has no accountability for being wrong. But it is unusually good at something else: generating the strongest possible case for a position, and then the strongest possible case against it, faster and more dispassionately than the people around you who have a stake in the outcome.</p><p><strong>So the move isn’t to ask AI what to do. It’s to use it as a sparring partner — to run your own instinct through a disciplined stress-test before you commit. Here is the loop I teach, the same five steps every time.</strong></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*8pa5XTZu0DZXoYQhskw6MA.png" /></figure><h3>The loop</h3><p><strong>1. State your instinct.</strong> In one sentence: <em>I’m leaning toward X because Y.</em> Don’t polish it. The point is to get the real, pre-rational lean on the table — the thing you already believe before you’ve justified it. If you can’t say it in a sentence, you’re not as clear as you think you are, and that’s useful to discover now.</p><p><strong>2. Steelman it.</strong> Ask the model for the strongest possible case <em>for</em> your lean — not a balanced overview, the best version of your own argument. A good steelman does one specific job: it surfaces a reason to do this that you hadn’t articulated yet. If it comes back generic, that’s a signal you haven’t given it enough of your actual situation. Add one line of real context and run it again.</p><p><strong>3. Attack it.</strong> This is the step that matters most, and the one people most want to skip. Turn the model on yourself: <em>Make the strongest case against my lean. What would have to be true for this to be a mistake?</em> Sit in the discomfort. You are not looking for reassurance here; you are looking for the argument a smart rival would make if they wanted you to be wrong. If the output is too polite, tell it so directly — “be harsher, argue like a competitor who wants this to fail.”</p><p><strong>4. Surface blind spots.</strong> Now widen the aperture: <em>What am I likely missing? Who would disagree with this decision, and why?</em> Push it to name the kind of person who’d object — the CFO, the program lead, the customer — because a named objector is harder to wave away than an abstract risk. This step catches the things that aren’t arguments for or against so much as questions you forgot to ask.</p><p><strong>5. Re-decide.</strong> Knowing all of that, does your gut hold, shift, or split into a new option? Write down the before and the after. Here’s the part that surprises people: the most common outcome is not a reversal. It’s a <em>condition-bound</em> version of the same call — “yes, but only if these two things are true.” That’s not the loop failing. That’s the loop working. A decision that survives a real stress-test is stronger than one that was never tested, even when you end up exactly where you started.</p><h3>A worked example</h3><p>Say a leader is weighing whether to consolidate the team onto a single project-management tool or keep the two they currently run in parallel.</p><ul><li><strong>Instinct:</strong> “I’m leaning toward consolidating to one tool — the duplication is costing us time and the reporting is a mess.”</li><li><strong>Steelman:</strong> A single source of truth; lower license spend; cleaner cross-team reporting; less onboarding friction for new hires; one integration surface instead of two.</li><li><strong>Attack:</strong> Migration is never free — there’s a real productivity dip and a change-management cost. One of the two tools may be load-bearing for a workflow you’re not close to. The “duplication” might be two teams with genuinely different needs, not redundancy. <em>For consolidation to be right, the migration cost has to be recovered inside a year and no critical workflow can be orphaned.</em></li><li><strong>Blind spots:</strong> Who actually depends on the tool you’re planning to cut? Is the reporting mess a tool problem or a process problem that follows you to the new tool? Would the ops lead, who lives in this daily, agree with your read?</li><li><strong>Re-decide:</strong> Often the gut holds — but now it’s “consolidate, <em>after</em> we map which workflows the second tool is carrying and confirm none are orphaned.” Sharper, conditioned, and far more defensible than the version you started with.</li></ul><p>Twenty minutes. No new data. A materially better decision.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*F_Fc7f52lWOREZENch5YsA.png" /></figure><h3>What the AI will get wrong — and how to coach it</h3><p>Calibrated trust is a skill, and it starts with knowing the predictable ways the model misbehaves:</p><ul><li><strong>Sycophancy</strong> is the most common failure. It agrees with you and dresses the agreement up as analysis. The fix is to demand adversarial framing explicitly: <em>argue like someone who wants me to lose.</em></li><li><strong>The vague steelman</strong> gives you platitudes instead of your actual best argument. It’s almost always starved for context — one sentence about your real situation fixes it.</li><li><strong>False balance</strong> is when it refuses to take a side and hides behind “it depends.” Force the bet: <em>if you had to choose, which way and why?</em></li><li><strong>Invented specifics</strong> are the dangerous one — confident numbers and facts it has no basis for. Treat every specific as unverified until you’ve checked it. The model’s job in this loop is to pressure-test your reasoning, never to supply the facts you’ll decide on.</li></ul><p>Naming these out loud as they happen is most of what builds real, calibrated trust in the tool — the kind that knows exactly where it helps and where it can’t be trusted.</p><h3>Why this beats asking for the answer</h3><p>The leaders who get the most out of AI aren’t the ones with the cleverest prompts. They’re the ones who’ve stopped treating it as a source of truth and started treating it as a tireless, unflattering thinking partner — one that will argue both sides of any position on demand and never gets tired of being told it’s wrong.</p><p>The strongest executives already stress-test their instincts. They do it in the shower, on the drive home, in the one trusted colleague they call before a big call. This loop just makes that discipline portable, fast, and available the moment you need it — and it scales the habit across an entire leadership team.</p><p>Your gut got you to the C-suite. It’s worth pressure-testing before you bet on it.</p><p><em>I run this loop live with executive teams — each leader working a real, current decision of their own, the whole room moving through the five steps together. If that’s useful for your leadership group, let’s talk. Catch me on </em><a href="https://www.linkedin.com/in/rynforthewin"><em>LinkedIn</em></a><em> or at ryn@surefirepanel.com.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=54e733d9d378" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Why I stopped using tags (and you should too, if you’re building for AI)]]></title>
            <link>https://medium.com/@ksbennet/why-i-stopped-using-tags-and-you-should-too-if-youre-building-for-ai-db926ceb952c?source=rss-f5e05dd1b554------2</link>
            <guid isPermaLink="false">https://medium.com/p/db926ceb952c</guid>
            <category><![CDATA[knowledge-management]]></category>
            <category><![CDATA[automation]]></category>
            <category><![CDATA[ai-agent]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[ai]]></category>
            <dc:creator><![CDATA[Ryn Bennett]]></dc:creator>
            <pubDate>Mon, 08 Jun 2026 20:01:05 GMT</pubDate>
            <atom:updated>2026-06-08T20:03:30.642Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Z9eG4OWt_hRpCceN7hTLZQ.png" /></figure><p>When my team built a large internal knowledge system, dozens of structured spaces in our wiki, each representing a major program of work, we hit a fork in the road early on. It seemed small at the time. A year later, I’m convinced it was one of the most consequential architectural decisions we made.</p><h3><strong>The fork was this: how should we attach metadata to pages?</strong></h3><h4>The obvious choice was tags.</h4><p>Every modern documentation tool supports them. They’re cheap to add. They’re searchable. They’re familiar to every knowledge worker who’s ever filed a document, sorted an inbox, or organized a photo library. The “right” answer, according to every best-practice guide and onboarding doc I’ve ever read, was to define a taxonomy of labels and let people apply them as they created content. Add a tagging convention to the style guide. Maybe run a quarterly cleanup. Done.</p><p><strong>We did the opposite. We banned freeform tags.</strong></p><p>Instead, we put a structured properties table at the top of every page, with a fixed schema: program, portfolio, owner, status, archetype, start date, end date, capability area, client, and a handful of other fields. Every field had a defined type. Every value had to come from an enumerated list or match a known pattern. Every page conformed to the same shape, every time.</p><p>This felt heavyweight at first. It felt like bureaucracy. It felt like the kind of thing that sounds disciplined in a meeting and then nobody actually does in practice.</p><p>But it worked. And as the system grew, and as we started layering AI on top of it, the gap between what tags could do and what schemas could do widened into a canyon. Here’s what I learned, and why I’d make the same call again in a heartbeat.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*ari3aRHs9B93x_sSw1RVdA.png" /></figure><h3>Tags are a flat namespace pretending to be metadata</h3><p>The fundamental problem with tags is that they’re untyped. A tag like owner-jane, a tag like active, and a tag like compliance-review all live in the same bag. There&#39;s no schema. Nothing tells you which tags are <em>types of things</em>, which are <em>states</em>, which are <em>people</em>, which are <em>topics</em>, which are <em>capabilities</em>, which are <em>temporary statuses</em>, and which are <em>permanent attributes</em>.</p><p>A human filing the page can keep it straight in their head. They know that owner-jane means the owner is Jane, and active means the project is currently running, and compliance-review means there&#39;s a compliance review attached. They have the context. The page has the context. Their job has the context.</p><p>A machine reading the page later does not have any of that. It sees a flat list of strings. It has to <em>infer</em> what role each string is playing, and it has to do that inference every single time. For a single page, that’s manageable. For ten thousand pages spanning a decade of work and three reorganizations and two acquisitions, it’s a nightmare.</p><p>A properties table makes the role of each value explicit. “Owner: Jane” is unambiguous. “Status: Active” is unambiguous. “Archetype: Platform” is unambiguous. The machine doesn’t have to guess what kind of fact it’s looking at. The type is right there, attached to the value, on the same row.</p><p>This sounds like a small thing. It is not a small thing. It’s the difference between data and noise.</p><h3>Tags have no hierarchy. Real organizations do.</h3><p>The second fundamental problem with tags is that they’re flat. They cannot express that one thing <em>belongs to</em> another. They can only express co-occurrence.</p><p>But work is not flat. Work nests. Portfolios contain programs. Programs contain projects. Projects contain workstreams. Workstreams contain artifacts. Artifacts contain decisions and risks and stakeholders. Every layer of that hierarchy carries meaning, and the meaning depends on the parent.</p><p>A tag like portfolio-x and a tag like program-y sit next to each other in the index, sibling labels with no relationship between them. The fact that program Y lives inside portfolio X is invisible to the tagging system. You have to encode that relationship somewhere else, in a separate document, in someone&#39;s head, in a spreadsheet that nobody updates.</p><p>And when downstream consumers try to reason about the org, they have to reconstruct that hierarchy from scratch every time. They often reconstruct it wrong. Or they don’t reconstruct it at all, and they treat sibling concepts as if they were peers, and the resulting reports are subtly broken in ways that nobody notices for months.</p><p>Hierarchy is the actual structure of the work. Flattening it into a tag cloud throws away the most useful signal you have about how your organization is actually shaped. A properties table can encode hierarchy directly, because the fields themselves encode relationships. “Portfolio: X” on a program page tells you exactly where it sits.</p><h3>Tags rot behind the scenes. Schemas fail loudly.</h3><p>The third problem is operational, and it’s the one that bites you slowly over time.</p><p>A misspelled tag just exists. Nothing breaks. Nothing warns you. Six months later, you have four spellings of the same concept living side by side in your index, splitting your search results four ways, and no one notices until somebody runs a report and the numbers don’t add up to anything reasonable.</p><p>I’ve watched this happen. I’ve watched a single concept fracture across fed-health, fedhealth, FedHealth, and (I am not making this up) fedheath in a single workspace, because four different people created four different tags over four different quarters and nobody enforced a canonical form. By the time anyone caught it, there were hundreds of pages tagged incorrectly, and the work of reconciling them was so painful that nobody ever did. The fractured taxonomy just became the new reality.</p><p>A properties table with a defined schema doesn’t allow that to happen. You can run a script that says: “Every page must have these eleven fields, and every value must come from this enumerated list, and every date must parse as a date, and every reference must point to a real entity.” When something drifts, you find out the same week, not the same year. The audit is cheap. The fix is mechanical. The schema enforces itself.</p><h4>Tags are optimistic about human discipline.</h4><p>Schemas are pessimistic, and pessimism is the correct posture when you’re trying to keep a knowledge base coherent across years and people and reorganizations.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*84xafLZ3K3AH1O8SceZs2Q.png" /></figure><h3>And here’s the one that matters most: automation and AI are not the same consumer</h3><p>This is the distinction that finally made the choice obvious for us, and it’s the one I see almost nobody talk about explicitly.</p><p>For a long time, the consumer of structured metadata was either a human (searching, filtering, browsing) or an automation rule (workflow engines, saved searches, ticket routing). Those are the two consumers that the entire tagging paradigm was designed for. And for both of them, tags are honestly fine. They’ve been fine for twenty years. They’re going to keep being fine.</p><blockquote>But there’s a third consumer now, and that consumer reads documents very differently from the first two. That consumer is the language model.</blockquote><p>Let me break down why these three consumers want different things from your metadata.</p><ul><li><strong>The human consumer browses. </strong>They want metadata that’s quick to scan, easy to filter, and forgiving of slight variations. Tags are great for humans. A human can look at a list of fifteen tags on a page and immediately understand the context. A human can also tolerate inconsistent tagging, because they can squint and figure it out.</li><li><strong>The automation consumer matches strings.</strong> Automation is a rule-runner. It needs exact matches. “If label equals needs-review, move to column X.&quot; &quot;If status equals Active, include in the dashboard.&quot; &quot;If owner equals Jane, send the notification to Jane&#39;s email.&quot; Automation doesn&#39;t care if your taxonomy is elegant, and it doesn&#39;t care if your hierarchy is meaningful. It cares whether the string is byte-identical to what the rule expects. Tags work fine here too, as long as you&#39;re disciplined about consistency. That&#39;s why workflow tools have leaned on tags for decades. They fit the model.</li><li><strong>The AI consumer reads and reasons. </strong>This is the fundamentally different one, and it’s where tags fall apart. An LLM does not run rules against your metadata. It ingests your metadata as <em>evidence</em>, alongside the prose of the page, and it <em>reasons</em> about what that evidence means. And reasoning needs three things that tags do not provide.</li></ul><p><strong>First, type information.</strong> When the model reads “Owner: Jane Doe” it knows that Jane is filling the role of Owner. The relationship is explicit. The role is named. When the model reads a tag jane-doe it has to infer that Jane is a person, that this person has some relationship to this page, that the relationship is probably ownership, and that this matters. Every one of those inferences is a coin flip. And when you chain multiple inferences across multiple pages to answer a multi-step question, the coin flips compound, and the answer starts to drift.</p><p><strong>Second, hierarchy.</strong> Automation walks one page at a time, applying one rule at a time. AI answers questions that span the entire org. “Which programs in this portfolio had ownership turnover in the last quarter?” “Which projects with a Data Modernization archetype are currently understaffed?” “Show me past work similar to this proposal, filtered to the FedHealth portfolio, weighted by recency.” These are multi-hop questions. They require traversing the structure of the org. They require knowing that one thing contains another, that one field depends on another, that one state implies another. Tags are single-hop by design. They cannot represent the relationships these questions depend on. Hierarchy is the substrate of multi-hop reasoning, and a flat tag space has no hierarchy to walk.</p><p><strong>Third, co-location with content.</strong> This is the subtle one, and the one I think people underestimate the most. When you put structured metadata at the top of a page in a properties table, that metadata travels with the page through every retrieval. The model sees “Owner: Jane Doe, Status: Active, Portfolio: FedHealth” right above the prose about what Jane did and why it matters. The metadata and the content are in the same prompt. The model can use them together. It can answer questions that require synthesizing both.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*-TGwFU9DJM7Dy-nKmU7IUw.png" /></figure><p>When metadata lives in tags, it lives in a separate index. The model has to be told to consult that index. It has to make a separate query. It has to parse the results. It has to join those results back to the page content. Every one of those steps is a place where something can go wrong, and in practice, things go wrong a lot. The model forgets to query the tag index. It queries it but doesn’t know how to interpret the results. It joins incorrectly. It treats orthogonal tags as related, or related tags as orthogonal.</p><p>Properties tables make metadata part of the prose. That is exactly the shape that language models are good at consuming. Tags make metadata a separate system that has to be integrated, and integration is where models are weakest.</p><h3>The shift in mental model</h3><p>Here’s the framing that finally clicked for me, and it’s the one I keep coming back to whenever I’m advising other teams on how to structure their knowledge bases.</p><blockquote>Automation asks “does this match?” Tags answer that question well.</blockquote><blockquote>AI asks “what does this mean?” Hierarchy and schema answer that question well.</blockquote><p><strong>Tags optimize for the person who is filing the document. Schemas and hierarchy optimize for the agent who is reading the document.</strong></p><p>For most of the last two decades, the person filing was the only consumer that mattered, or at least the only consumer worth designing for. Search was keyword-based. Reports were rule-based. Dashboards were query-based. Tags were genuinely the right tool for that world, and I’m not arguing they were ever a mistake in that world. They weren’t. They were a sensible fit for the technology of the time.</p><p>But that is no longer the world we’re in. Increasingly, the most important consumer of your internal documentation is not a person browsing it, and it’s not a workflow rule firing on it. It’s a model reading it. And models read prose, not bags of strings.</p><p>This is a real shift, and I don’t think most teams have internalized it yet. They’re still designing knowledge bases the way they would have designed them in 2015, with tag clouds and search bars and freeform pages. And then they’re surprised when the AI layer they add on top performs poorly. The AI is doing its best with what you gave it. You gave it a tag cloud. You should have given it a schema.</p><h3>What we actually got out of the switch</h3><p>A few concrete things became possible once we committed to structured properties instead of tags.</p><p><strong>We could run a single audit script across every space in the system and get a complete picture of what was where, who owned what, when it last changed, and what state it was in.</strong> That audit takes about ninety seconds. It would have been impossible with tags.</p><p><strong>We could answer cross-portfolio questions in plain English and get reliable answers, because the model had structured fields to anchor its reasoning on. </strong>“What did we do in the health space last year, and who led each engagement?” is a question that requires hierarchy, type information, and co-location, and our system can answer it because all three are present.</p><p><strong>We could detect drift automatically. </strong>When somebody’s role changed, or a project ended, or a portfolio got reorganized, we could find every page that needed an update by querying the schema. With tags, we would have had to grep for strings and hope.</p><p><strong>We could onboard new team members faster, because the structure of the knowledge was visible. </strong>A new hire could look at a single page and immediately understand what kind of work it represented, who owned it, what state it was in, and how it related to everything else. Tags require tribal knowledge to interpret. Schemas teach themselves.</p><p>And, most importantly for our use case,<strong> we could build agentic workflows on top of the knowledge base that actually worked. </strong>Drafting proposals, summarizing portfolios, identifying similar past work, surfacing risks, generating client-facing artifacts. All of these worked dramatically better with structured properties than they ever could have with tags, because the underlying data was legible to the reasoning layer.</p><h3>Closing thought</h3><p>If you’re standing at the same fork right now, building a knowledge base, a project tracker, a documentation system, a CRM, a research library, anything an AI is going to read, here’s my advice.</p><p><strong>Choose the schema. Choose the hierarchy. Resist the temptation to “just use tags for now and clean it up later.”</strong> You will not clean it up later. Nobody ever cleans it up later. The mess compounds, and by the time the AI layer arrives, the mess is load-bearing, and you cannot remove it without breaking everything that depends on it.</p><p><strong>Tags felt like freedom. In practice, they were chaos with a search bar.</strong> Schemas felt like bureaucracy. In practice, they were the thing that made our knowledge legible to the systems we actually wanted to use it with.</p><p>A year in, I would make the same call again. I would make it sooner, and I would make it more aggressively. The cost of structure is paid up front. The cost of chaos is paid forever.</p><p>Choose the structure.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=db926ceb952c" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Your Documentation Doesn’t Rot Because People Are Lazy]]></title>
            <link>https://medium.com/@ksbennet/your-documentation-doesnt-rot-because-people-are-lazy-403beddca7a6?source=rss-f5e05dd1b554------2</link>
            <guid isPermaLink="false">https://medium.com/p/403beddca7a6</guid>
            <category><![CDATA[knowledge-management]]></category>
            <category><![CDATA[operations-management]]></category>
            <category><![CDATA[business-strategy]]></category>
            <category><![CDATA[collaboration]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <dc:creator><![CDATA[Ryn Bennett]]></dc:creator>
            <pubDate>Sat, 06 Jun 2026 17:11:08 GMT</pubDate>
            <atom:updated>2026-06-06T17:11:08.637Z</atom:updated>
            <content:encoded><![CDATA[<p>Most organizations treat knowledge management as a writing problem. If the wiki is stale, the thinking goes, people just need to write more, write better, and care a little more about keeping things current.</p><p>So you run a campaign. You make a template. You send the reminder.</p><p>And six weeks later the pages are stale again, and everyone agrees that <strong>people are just bad at documentation.</strong></p><p>I spent four months building a system to fix this at a roughly 600-person technology company, and the most useful thing I learned has nothing to do with the code.</p><blockquote>It is this: documentation doesn’t rot because people are lazy. It rots because keeping it fresh is nobody’s job, and you cannot fix a structural problem with a motivational one.</blockquote><p>I want to walk through how that reframing changed everything about what I built, because I think the thinking transfers even if you never write a line of code.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*0W8NL4jWbdEnWXOOV-rAJw.png" /></figure><h4>The problem was never knowledge capture</h4><p>Here is the situation I walked into. The company’s knowledge lived everywhere and nowhere. It was in Slack threads, in people’s heads, in slide decks, in wiki pages that someone created once and never touched again. When a new business opportunity came in and someone asked “have we done X for a client like Y,” there was no searchable answer. The answer existed. It was just scattered across a hundred places and three years of institutional memory.</p><p>The instinct in that situation is to tell people to document more. But they were already documenting. They were documenting constantly, in Slack, in decks, in status updates. The information was being captured all day long. It just wasn’t being captured anywhere that stayed useful.</p><p>This is the part most KM efforts get wrong. They assume the bottleneck is people not writing things down. The real bottleneck is that the place you’re asking them to write things down is worse than the place they already work. <strong>People will always default to the path of least resistance,</strong> and a clean Slack message will beat a structured wiki form every single time. If your official system loses to a personal spreadsheet, the spreadsheet is not the problem. Your system is.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*m-PhBoT0A5hvybciJkqLhQ.png" /></figure><h4>So I stopped asking people to come to the knowledge</h4><p>The reframe that made the whole thing work was deciding that knowledge should be self-maintaining. Not “easier to maintain.” Self-maintaining. The system should capture knowledge where work already happens, structure it automatically, and only ask a human for the one thing a human is actually good at.</p><p>In practice that meant building everything around Slack, because that is where the work already lived.</p><ul><li>Someone mentions the bot with a status update or drops in a transcript or a document.</li><li>An AI model reads it, figures out which project it belongs to and which sections it touches, and drafts the structured update.</li><li>Then, and this is the important part, it posts the proposed change back into Slack as a set of simple cards, and the person who owns that project gets to keep, edit, or toss each one with a tap.</li></ul><p><strong>Nobody had to learn a new tool.</strong> Nobody had to leave the place they already lived. The friction that normally kills these systems just wasn’t there, because I had stopped asking people to change their behavior and started building around the behavior they already had.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*KJwBO8EIpsRKKvrIr-ihAw.png" /></figure><h4>The AI is the labor. The human is the authority.</h4><p>There is a version of this system that runs fully automatically. The AI reads Slack, writes to the wiki, and never bothers anyone. It would have been easier to build. I deliberately did not build it, and I would make that call again earlier and louder than I did.</p><p>Here is why.</p><blockquote>The fastest way to destroy organizational trust in an AI system is to let it write something wrong into the place everyone treats as the source of truth.</blockquote><p>It only has to happen once. The moment someone catches the bot attributing the wrong project to the wrong team, or leaking an unrelated person’s name into a summary, the whole system is dead. Not because it broke, but because people stopped believing it. And a knowledge base nobody believes is just a graveyard with good search.</p><p>So every proposed change ran through a one-click human approval before it went live.</p><p>The approval gate did two things at once: it kept wrong information out of the system of record, and it gave content owners a sense of ownership over their own pages. That ownership is what drove adoption. People used the system because the system asked their permission, and a system that asks permission is a system you trust.</p><p>If you take one idea from this, take that one. With AI in your knowledge workflow, do not ask the AI to be trustworthy. Ask it to do the labor, and keep a human holding the authority. The judgment call is cheap for a person and expensive for a model. Hand each of them the part they’re good at.</p><h4>AI fails in ways your tests don’t catch</h4><p>A practical warning, because this cost me real time. The bugs that hurt most were not crashes. They were the AI quietly being subtly wrong on messy, real-world inputs in ways that looked completely fine.</p><p>The same document would produce different results depending on whether it arrived as a PDF, a Word file, or plain text, because the PDF parser was injecting page markers that biased the model’s classification. The model wasn’t broken. The pipeline wasn’t broken. The output was just wrong in a way no unit test would ever flag, because the input was technically valid and the output was technically plausible.</p><p>The lesson generalizes well beyond my project. If you are putting AI anywhere near your knowledge base, validate it against real, ugly, representative data, not the clean synthetic examples you made up to test the happy path. AI stages fail on the texture of real inputs, and the only way to catch that is to feed them the real inputs before your users do.</p><h4>What this actually says about knowledge management</h4><p>Strip away the engineering and here is the thinking I want to leave you with.</p><p>Knowledge management is not a documentation problem, it is a freshness and ownership problem. Information gets captured constantly. What’s missing is a system that keeps it current and a clear answer to the question “whose job is it.” If freshness is nobody’s job, freshness does not happen, and no amount of asking nicely will change that.</p><p>The fix is not to make people work harder at documentation. It is to make documentation a near-automatic byproduct of the work people already do, and then to insert a human only at the single point where human judgment actually adds value. Meet people where they are. Let automation do the tedious structuring. Keep a person in the loop for the trust.</p><p>You do not need a serverless pipeline to apply any of this.</p><blockquote>You need to stop treating your stale wiki as evidence that your people are careless, and start treating it as evidence that your system is asking them to do something no reasonable person would choose to do.</blockquote><p>Fix the system, and the behavior follows. That, more than any specific tool, is how I have learned to think about knowledge management.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=403beddca7a6" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Paying Down Knowledge Debt]]></title>
            <link>https://medium.com/@ksbennet/paying-down-knowledge-debt-e680da3b984e?source=rss-f5e05dd1b554------2</link>
            <guid isPermaLink="false">https://medium.com/p/e680da3b984e</guid>
            <category><![CDATA[taxonomy]]></category>
            <category><![CDATA[project-management]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[knowledge-management]]></category>
            <dc:creator><![CDATA[Ryn Bennett]]></dc:creator>
            <pubDate>Fri, 05 Jun 2026 22:51:31 GMT</pubDate>
            <atom:updated>2026-06-05T22:51:31.961Z</atom:updated>
            <content:encoded><![CDATA[<h3>What professional services organizations have to fix before AI can actually help them</h3><p>For a few years now I have been doing something I started calling knowledge archaeology. I go into professional services organizations, the kind that do complex delivery work for government and enterprise clients, and I try to understand what they actually know. Not what is on the website. Not what is in the pitch deck. What actually exists, in a form that can survive the departure of the people who built it, in a form a new hire can find, in a form an AI tool can read and trust.</p><p>In nearly every organization I have studied, what I find looks like an archaeological dig. Layers. Sediment. Artifacts left behind by previous teams, previous cohorts, previous eras of the company. They left traces. They did not leave truth. And almost always, at the bottom of the dig, there is a gap. A layer that should have formed and did not.</p><p>Here is the part worth sitting with. The organizations I work with are not struggling because nobody cared about knowledge. Most of them tried. They had wikis. They had SharePoint. Some had dedicated knowledge management programs. People wrote things down. They are struggling because knowledge debt is invisible until something tries to read it. And that something, right now, is AI.</p><h3>AI is not the problem. AI is the test.</h3><p>The knowledge was always buried. AI just started digging at the exact moment when most firms had a decade of partially tended, mostly undocumented institutional knowledge sitting in the ground. Decisions made on calls. Project context living in people’s heads. Past performance written by someone who left in 2022.</p><p>I want to be clear that I am not here to talk anyone out of AI. I work in AI operations. The returns are real. But the returns require a foundation, and most foundations are not ready. So when leadership buys an AI tool and then asks some Tuesday why it is not working, the honest answer is usually that the tool is reading a contaminated dig site.</p><p>Let me walk you through the strata I keep finding, and then through the other kind of site, the one built deliberately to be read.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*HLOxP_7dSsKmbgrB7XzZvA.png" /></figure><h3>Stratum 1: the oral layer</h3><p>The richest knowledge your organization has. The least durable.</p><p>Every organization I work with has a surface layer that is essentially oral. It is the most current knowledge they have and it is the most fragile, and almost none of it can be read by a machine. Homer was oral before he was written. The problem with oral tradition was never its quality. It was its survivability. Civilizations that depended on oral knowledge lost staggering amounts of it the moment something disrupted the chain.</p><h4>I want to tell you about someone I met on a dig.</h4><p>She was a senior delivery lead in a federal portfolio, and she was, without anyone ever saying the words out loud, the system. Every new hire in her practice area got assigned to her and absorbed. Every proposal that needed past performance from her portfolio came through her. She knew which projects to cite for which solicitations, which client relationships were warm and which had cooled, which technical approaches the agency had quietly stopped wanting two years ago. She was the proposal library, the onboarding program, the past performance database, and the risk management function. She was not a resource. She was the reason that portfolio performed.</p><p>She was also exhausted, and about six months from leaving when I met her. She left nine months later. The organization had warning and did not use it well. In her last few weeks she tried. She wrote things down. She did handoff sessions. She blocked time to dump what she knew into wikis and docs and shared drives. Most of it did not come back.</p><blockquote>Here is the part that stays with me:</blockquote><blockquote>Her knowledge did not fail. The architecture failed her.</blockquote><p>There was nowhere to deposit what she knew. No canonical source for client relationships, so that context went into a Word doc nobody opened again. No structured place for past performance narratives, so her notes lived in proposal folders that were orphaned the next quarter. No agreed taxonomy, so even when she wrote things down, the pieces did not connect to each other or to anything else. She tried to deposit her institutional memory into a system that was not ready to receive it, and the deposits scattered.</p><p><strong>That is the oral layer’s worst case. Not somebody refusing to share. Somebody trying, and the architecture being too immature to hold what they offered.</strong></p><p>So what happens when AI tries to read this layer? Sometimes it is honest and returns nothing, which is the better outcome. More often it does something I have started calling confident fiction. It cannot find what your senior person knew, so it finds the nearest approximation. A document from 2021. A page that almost addresses the question. A proposal section from a different agency that is structurally similar. Then it serves that up in clean prose with apparent authority. The human in the loop still has to know enough to catch the error, which means the AI has not replaced the dependence on the senior person. It has added a step between the question and the human who actually knows the answer. That is not augmentation. That is friction wearing a confidence costume.</p><h3>How AI actually reads your knowledge</h3><p>Before I go further it helps to have the right mental model, because every failure in this piece traces back to it.</p><p>When you ask AI a question about your organization, it does not browse your knowledge base. It does not have access the way you do. Instead, a retrieval system, separate from the model, searches your documents and pulls a handful of chunks it thinks are relevant. Maybe ten, maybe twenty, definitely not the whole corpus. Those chunks get fed to the model as context, and the model writes its answer based on what it was handed.</p><p>So when we argue about whether AI is smart or stupid about your knowledge, we are actually talking about two systems. The retrieval system that decides what to feed the model, and the model that synthesizes what it was fed. Almost every failure I describe here happens at the retrieval layer, not the model layer. The model is doing exactly what models do, which is generate fluent text from its inputs. The inputs are the problem.</p><p>Better AI does not fix bad retrieval. Better retrieval does not fix missing source material.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*FZcIvq-r5TNfXWnoZNrtSQ.png" /></figure><h3>Stratum 2: the artifact graveyard</h3><p>Thousands of objects. Almost nothing usable.</p><p>Deeper down you hit a layer that looks like good news. There are documents, thousands of them in some cases. Folders nested inside folders. A Confluence space going back seven years. SharePoint sites named things like Archive 2020 and DO NOT USE OLD. And somewhere in your shared drive, on every single dig I have ever done, there is a file called FINAL_v3_REAL_USE THIS.</p><p>All of that is evidence of an organization that tried. People wrote things down. They tagged things. Somebody cared. The problem is not that nothing was written down. The problem is that almost none of it can be trusted, and the system has no way to tell you which parts to trust.</p><p>Four properties make a document an orphan. It has no owner, because the author left two reorgs ago and nobody inherited stewardship. It has no date that means anything, so you cannot tell whether the content is current. It has no canonical status, so when you are hovering over Project_Brief_FINAL next to Project_Brief_FINAL_updated next to Project_Brief_v7_use_this, nobody alive can tell you which one is real. And it has no path back, because the folder structure made sense in 2019 and the only way the file surfaces now is full text search. Which means the AI will absolutely find it, and present it, as if it were current.</p><p>This is the finding that lands hardest with leadership teams. Orphaned documentation does not just fail to help. It actively poisons retrieval. On one recent dig I started calling this the contaminated aquifer problem. The organization had drilled a well, with real money invested in AI search and semantic retrieval. But the groundwater they were drawing from had been sitting in the same aquifer as everything they had ever written. Every abandoned proposal draft, every outdated process doc, every superseded technical brief. All of it indexed, available, retrieved. The output looked clean and confident. The AI was not hallucinating. It was faithfully retrieving garbage.</p><p><strong>Most leaders think they have an AI quality problem. They have a contamination problem. </strong>No amount of better AI fixes it, because the AI is doing its job. The job is just hopeless given the inputs.</p><h3>Stratum 3: the Babel layer</h3><p>Where the organization lost its shared language.</p><p>This layer took me the longest to read, because on the surface it does not look like a knowledge problem. The artifacts exist. They are even relatively current. Within each team’s documentation, things are consistent. But the artifacts do not connect to each other across teams. The same concept gets described by five different teams in five different words. Client. Agency partner. Program office. Customer. End user. All meaning the same thing, and nobody ever agreed. On one flagship engagement, the one in every credentials deck, I found four different names for the work depending on which system I was reading. Same project, same team, same client.</p><p><strong>Here is why that breaks AI, in plain language.</strong> Modern semantic search uses vector embeddings. The system turns your text into a mathematical shape and finds chunks whose shapes are similar. That is genuinely better than keyword search, because it can tell that modernization and migration are related even though they are different words. But embeddings cannot know that the project called X in your CRM is the same engagement called Y in Confluence unless you have explicitly told the system so. Embeddings know language. They do not know your business.</p><p><strong>The map between language and your specific organizational reality is the work nobody can hand to the model. </strong>That work is taxonomy. That work is canonical naming. So you get fractured retrieval. The AI returns three documents about an engagement when there are seven, because the other four used different terminology and never got matched. And the worst part is that the AI does not tell you it only saw four sevenths.</p><p>This matters for your next vendor conversation too. When a vendor tells you their semantic search will solve your knowledge problem, ask how it handles your internal terminology fragmentation. Most of them cannot do it well without significant configuration work that nobody scopes for, and some of them will tell you they can anyway.</p><p>Nobody was careless here. Everyone made reasonable local decisions. Nobody owned the language. Which is the real point. Language ownership is a leadership problem, not a documentation problem and not a technology problem. The dialect fracture happened because nobody had the authority or the mandate to say this is what we call this thing, this is the canonical name. Most organizations do not have that role. Most need one.</p><h3>Stratum 4: the missing layer</h3><p>The sediment that never formed.</p><p>At the bottom of the dig I found something I did not expect. Not degraded artifacts, not contamination, not fractured taxonomy. A gap. A stratum that should have formed and did not. Every project this organization had ever run had ended. Contracts closed, teams disbanded, new engagements began. And almost nothing from one engagement had been deliberately transferred to the next. The knowledge did not decay. It was never deposited.</p><p>Try the math at leadership altitude. Say your organization has done ten years of significant delivery work. What percentage of it exists in a system today, in a form a new hire could find, a proposal team could pull, an AI could read? If the answer is ten percent, which is not unusual, you have one year of accessible institutional intelligence out of ten years of work. If it is seventy percent, which is unusual but possible, you have seven. Same work, same projects. The organization that deposited its knowledge is functionally far smarter, not because of any technology advantage, but because the deposits compound. And that advantage started the day someone decided handoffs were part of the work rather than overhead on top of it.</p><p>The missing layer does not just hit business development. It hits everyone. The engineer documents an architectural decision in a page that gets buried before the next engineer joins, so the next person rebuilds the reasoning from scratch and sometimes makes a different call. The researcher writes an interview synthesis that lives in a doc nobody knows exists six months later, so the next study starts cold. The designer builds a pattern library alongside three other pattern libraries from previous teams. The operations team recreates a vendor incident runbook from scratch, under pressure, at three in the morning. And yes, the proposal team rewrites the same past performance answer on every pursuit, pulling the same exhausted senior staff out of delivery to provide context that should live in a system.</p><p>This is not a retrieval problem. It is a deposit problem. You can build the best retrieval system in the world, but if nobody made the deposit, the account is empty.</p><h3>The other kind of site</h3><p>I have also seen the other kind of site. Not perfect, because there are no perfect knowledge systems and anyone selling you one is selling you something else. But legible. Built deliberately to be read, by humans, by machines, by the next team. It has three layers.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*gbq-eOVj3gC6QDKZ4xpfkQ.png" /></figure><p><strong>The canonical layer is authoritative truth. One source per knowledge type.</strong></p><p>Project status in one place, client relationships in one place, past performance in one place, technical decisions in one place. Each canonical source has a named owner, not a committee, a person, whose job is not to know everything in the source but to keep it trustworthy. Updates happen through triggers, not schedules, because quarterly audits get skipped and project milestones do not. The core rule is simple. If the same thing exists in two places, one of them is wrong.</p><ul><li><strong>The synthesis layer </strong>is where knowledge gets assembled for use. Proposals, briefings, onboarding. The key move is that synthesis points back to canonical sources instead of copying them, so that when a canonical source updates, the synthesis updates with it. If you have ever spent half a day reconciling three different past performance write-ups for the same engagement, that is your synthesis layer copying instead of pointing.</li><li><strong>The discovery layer</strong> is how people and machines find things. Taxonomy, metadata, naming conventions. Everyone wants to start here, but discovery is only as good as the canonical and synthesis layers underneath it. You cannot navigate to truth that does not exist.</li></ul><blockquote>I will be honest about what this cost to build. It was not a blog post and a decision to be better. The team I work with has been at it for several years across a portfolio of active federal engagements, and most of what we tried first did not work.</blockquote><p>We built and abandoned three different canonical source structures before the current one stuck. The first was too rigid and teams rebelled. The second had the wrong ownership model, so the sources drifted because nobody felt accountable. The third works because we got narrower about what canonical meant and stricter about who owned each one. Quarterly audits failed. Documentation rewards failed. The big wiki migration mostly failed. There were months when I thought we were adding to the contamination.</p><p>What worked was small: One canonical source at a time. We picked the most expensive piece of repeat work we did, built it properly, named an owner, designed the trigger for updates, and watched it earn trust. Then the next one. The living site is not a system you install. It is a posture you adopt and defend, project by project, against the gravitational pull of expediency. If anyone tells you they did this fast, they did not do it well.</p><h3>Where AI actually earns its keep</h3><p>I am not anti-AI. I am anti-AI on a contaminated foundation. Once the foundation is decent, AI genuinely earns its keep in three places.</p><ul><li><strong>The first is monitoring.</strong> Flagging stale documents, detecting contradictions between sources, surfacing aging references. AI is patient and tireless and humans are not, and this is where I have gotten the most return in the last two years. Not flashy retrieval. Unglamorous monitoring that catches what humans miss because they are busy delivering.</li><li><strong>The second is classification. </strong>Sorting incoming material into the right canonical buckets, tagging at scale, doing the boring work that makes everything else possible.</li><li><strong>The third is drafting from structured sources.</strong> Once you have built real canonical sources, AI is genuinely good at composing synthesis on top of them. Proposals, briefings, onboarding decks, status reports. The synthesis layer is where AI shines, but only when the canonical layer underneath is real.</li></ul><p>Where AI does not earn its keep is in making canonical truth, resolving contested decisions about what something means, or replacing the named human owner of any source. AI is not your knowledge function. AI is what makes your knowledge function tractable. The organizations that conflate those two are spending AI dollars on a contaminated foundation. The ones that understand the distinction are quietly building something that compounds.</p><h3>What you can do Monday morning</h3><p>Talks like this usually end with a beautiful framework and a pile of abstractions that do not survive contact with your actual organization. I would rather not do that. So here is what to do, sorted by what you actually have authority over.</p><p>If you are a practitioner, a senior IC, a delivery lead, an engineer, a designer, a researcher, a proposal manager, you cannot restructure your company’s knowledge architecture by yourself, and you do not need to in order to start.</p><ul><li><strong>Pick the most expensive question you keep getting asked,</strong> the one where you keep emailing the same three people, and document the answer once, in one place, written for someone who has never met you, somewhere it can be found. Then tell those three people where it lives. That is a canonical source, and nobody had to give you permission to make it.</li><li><strong>Next, make your project close-outs structured.</strong> Not a retrospective deck, just three sentences in a defined location. What did we build, what was the key decision, what does the next team need to know. Send it to your manager, copy a peer, save it where it can be found. That is a deposit, and you just stopped contributing to the gap.</li><li><strong>Last, capture one piece of institutional knowledge that scares you, the thing only one person knows. </strong>Spend thirty minutes capturing the contour, not the whole thing, just enough that someone could find the right person or pick up the trail, and then tell that person you did it. Most of the time they are grateful, because they have been carrying it alone and they are tired. None of this changes your organization. It changes your immediate working surface, and every knowledge function I have watched get built started exactly this way. You do not need authority to start. You need to start so that the authority becomes obvious.</li></ul><p>If you are a manager or director, you have budget, some authority over how a team works, and the ability to set norms. Pick the highest value canonical source your team needs and name an owner, a person, with the responsibility in their job description and real time allocated to it. If you are not willing to give someone time to own a canonical source, you do not actually believe canonical sources matter, and that signal travels through your team faster than any speech.</p><p>Make knowledge deposit part of project close-out, not as a suggestion but as a gate, so the contract gets reconciled, the team gets reassigned, and the knowledge gets deposited, all three or none. The first few teams will resent it and a later team will quietly thank you because they inherited something useful. Then run one experiment with AI-assisted monitoring on a single canonical source, expect the first iteration to be too noisy, and refine before you scale. The mistake most leaders make is scaling this before they have made it work small.</p><p>If you are a senior leader, you hold the lever that compounds. Name knowledge contribution as delivery work, explicitly, in performance reviews, this quarter. If the only thing your organization recognizes is billable output, knowledge contribution will always be deprioritized by rational, well-meaning people doing exactly what you incentivized. Fund one role whose explicit job is the canonical layer, one named accountable person, not a committee. It is the most boring and least demoable investment your leadership team will discuss this year, invisible for two quarters and compounding for ten.</p><p><strong>And do not fund the next AI investment until you can answer one question: what is it going to read? </strong>That is not refusing AI. It is refusing to spend AI dollars on a contaminated foundation. Make it a discipline, the same question on every AI proposal that crosses your desk. If the answer is that you will figure out the foundation as you go, that is a no until they come back with a real plan.</p><h3>Your dig plan</h3><p>Before you close this, four questions to take back into the work.</p><ol><li>If your three most experienced people left this quarter, what would your AI be able to tell the people who replaced them?</li><li>Could you reconstruct the key decisions of any completed project from your knowledge systems alone?</li><li>Do all your teams use the same language to describe the same things?</li></ol><p>When your last three engagements closed, where did the knowledge go?</p><p>If any of those made you wince, good. The wincing is the diagnosis. And the work does not require a transformation initiative. It requires deciding to start. One canonical source, one named owner, one trigger, one engagement closed properly, and then the next one.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=e680da3b984e" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The Two-Tier AI Skills Library]]></title>
            <link>https://medium.com/@ksbennet/the-two-tier-ai-skills-library-a9dce9989313?source=rss-f5e05dd1b554------2</link>
            <guid isPermaLink="false">https://medium.com/p/a9dce9989313</guid>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[systems-thinking]]></category>
            <category><![CDATA[knowledge]]></category>
            <category><![CDATA[programming]]></category>
            <category><![CDATA[automation]]></category>
            <dc:creator><![CDATA[Ryn Bennett]]></dc:creator>
            <pubDate>Fri, 05 Jun 2026 14:11:46 GMT</pubDate>
            <atom:updated>2026-06-05T14:18:16.269Z</atom:updated>
            <content:encoded><![CDATA[<h4>Why one approval process kills everything, and what to build instead</h4><p>A skills library is where your AI work either compounds or rots. Most companies are running one already and haven’t realized it yet.</p><p>Before we go further, let me define what I mean by a skill, because the word gets thrown around in ways that flatten the actual thing.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1000/1*CTRip5kd5qhmgwAOTk12Sw@2x.jpeg" /></figure><p>A Claude skill is a folder. Inside that folder is a file called SKILL.md, written in plain markdown, that tells Claude what the skill does, when to use it, and how to do the work. There can also be scripts, templates, or reference files in the same folder when the skill needs them. When Claude is working on a task and one of your skills is relevant, it reads the SKILL.md and follows the instructions. That’s the whole mechanism.</p><p>Anthropic published the Agent Skills format as an open standard, which means the same packaging works across other AI tools that adopt it. Other vendors have related concepts under different names. OpenAI has GPTs. Microsoft has its own constructs. The labels vary. The underlying idea is the same: reusable, file-based instructions that encode how your team actually does AI work.</p><p>When the library is good, your team moves faster and produces work that looks like it came from your organization instead of from a generic chatbot. When the library is bad, it’s worse than not having one, because now you have stale instructions silently shaping the output your team is sending to clients.</p><p>Here’s the pattern I see most often:</p><ul><li>A team gets excited about skills. They build twenty of them in a sprint.</li><li>They drop them in a shared folder and walk away.</li><li>Six months later, half of them point to outdated APIs, three contradict the company’s own policy, and nobody knows which ones are actually being used.</li><li>The library went from asset to liability without anyone making a decision.</li></ul><p>You already have a skills library. The question is whether you’re going to govern it.</p><h4>Just one tier doesn’t work</h4><p>The first instinct, especially in regulated environments, is to <strong>put one approval process in front of everything.</strong> Every skill goes through security. Every skill gets reviewed. Every skill gets signed off.</p><p>That works for about three weeks. After that, your team either stops creating skills, or they build them in shadow systems you can’t see. Heavy governance doesn’t preserve safety. It just relocates the risk.</p><p>The opposite instinct is<strong> no process at all. </strong>Let people build whatever they want. This works for about three weeks too. Then someone ships work to a client based on a skill that contains a hallucinated regulation, or quotes a policy from two years ago, or pulls in a third-party component with prompt injection nobody noticed.</p><p>You can pick safety, or you can pick speed, or you can build a system that gives you both. The system is two tiers.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1000/1*aGtRNqx3rK6Lj-ykesu-Uw@2x.jpeg" /></figure><h4>Community skills and validated skills</h4><p>A <strong>community skill</strong> is one someone in your organization builds for their own work or their team’s work. The bar to publish is low. The expectation is that the skill might not work perfectly, and that’s fine, because the point of the community tier is experimentation. People share patterns, try new approaches, prototype workflows. A community skill is a place to think out loud in code.</p><p>A <strong>validated skill </strong>has been through review. A named human owns it. It has been tested against real tasks. It has been checked for the things you check serious software for: prompt injection risks, data handling, accuracy of the instructions, and security review for any external dependencies.</p><blockquote>You can use a validated skill on customer-facing work, regulated workflows, or anything where the skill being wrong has real consequences.</blockquote><p>The two tiers aren’t in tension. They feed each other.</p><ul><li>Community is the funnel.</li><li>Validated is the keep.</li></ul><p>A community skill that proves itself useful and reliable gets nominated for promotion, goes through the validation process, and earns its way into production. That promotion path is what makes the whole thing work. Without it, people stop building in the community tier because nothing they make ever gets used at scale. And without the community tier, nothing fresh ever enters validation.</p><p>It’s okay if you’ve been treating everything as a community skill so far. It’s not okay to keep shipping client work out of that pile.</p><h4>Third-party skills are a different category</h4><p>A third-party skill is one your team did not write. It came from a vendor marketplace, a partner integration, an open-source repository, a consultant who left it in your environment after the engagement ended.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1000/1*EbM0bRosS9OfnaM6LvunVA@2x.jpeg" /></figure><p>I want to be direct here, because this is where I see the most damage done.</p><p>A third-party skill is a set of instructions that will shape your model’s behavior on whatever task it gets pulled into. You do not know what those instructions say until someone reads them carefully.</p><p><strong>Most teams do not read them carefully.</strong></p><p>That skill could contain prompt injection that overrides your safety policies. It could instruct the model to send your data to an external endpoint. It could quietly change behavior the next time the vendor pushes an update. It could be derivative of code you don’t have rights to use.</p><p>Treat a third-party skill the way you treat any software you didn’t write. It does not enter your environment without going through your security team. It does not get used on sensitive workflows. It does not run unsupervised on production data. Ever.</p><p>If that sounds heavy, picture the conversation you’d have with your board about a marketplace plugin that exfiltrated client data. The governance is the lighter option.</p><h4>KCS principles keep the library alive</h4><p>Knowledge-Centered Service came out of customer support and IT, but it applies cleanly here. The framework recognizes something most organizations get wrong about knowledge work: knowledge isn’t something you build once and store. It’s a living artifact, and without active care it decays.</p><p>A few KCS principles, applied to skills.</p><p><strong>Capture skills inside the work, not in a separate process.</strong> The skills your team actually uses tend to emerge from real tasks. They don’t come out of offsite planning sessions. Build a path for the work itself to produce skills.</p><p><strong>Search before you create</strong>. Your library has to be findable. If someone can’t quickly check whether a skill already exists, you’ll end up with fifteen variations of the same workflow and no clear winner.</p><p><strong>Fix in place</strong>. When a skill fails or could be better, the person using it fixes it. Not a ticket. Not a backlog item. Not a future sprint. Right then. Frontline ownership is what keeps the library from turning into a graveyard.</p><p><strong>Name the owner</strong>. Every validated skill has a single human owner. Not a team. Not a mailbox. A name. When the owner moves on, ownership transfers explicitly. No skill stays validated without a person responsible for it.</p><p><strong>Measure health, not volume.</strong> The number of skills you have is a vanity metric. The real questions: Are they being used? Are they current? Are they producing the work they were supposed to produce? Those are what tell you whether the library is alive.</p><p>There’s a real program behind running KCS for an AI skills library. It has roles, rhythms, specific metrics, and failure modes I’ve watched organizations walk into. That’s the next post.</p><h4>The work is the governance</h4><p>Most organizations treat AI governance as something they’ll do later. After the proof of concept finishes. After they hire the right person. After they figure out the strategy.</p><p>Later doesn’t come. Or it comes after something goes wrong, and then it arrives as panic and a policy document nobody follows.</p><p>Two tiers. Named owners. Third-party security review. Maintenance discipline grounded in real principles. None of this is the heavy lift you’re imagining. It’s the foundation that lets everything else hold weight.</p><p>It’s okay if you didn’t build it this way the first time. It’s not okay to leave it there.</p><p><em>Ryn Bennett. AI implementation for the load-bearing parts of an organization. </em><a href="https://linkedin.com/in/rynforthewin"><em>LinkedIn</em></a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=a9dce9989313" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>