# Legible: full writing archive > Legible is your fractional head of AI, helping leadership teams turn Claude adoption into operating leverage, from strategy to team enablement to execution. This file carries the full text of everything published on the Legible blog (https://legible.co/blog/). A curated index of the site lives at https://legible.co/llms.txt. ## You know it when you see it, part two URL: https://legible.co/blog/you-know-it-when-you-see-it-part-two/ Author: Lawrence Coburn Published: 2026-09-11 Type: Essay TL;DR: A second portrait of AI-native leadership: the soft-spoken CEO of a 300-person company who got curious, vibe-coded his own tools, and became the frontier pusher for his whole team. Leadership comes in different flavors, and a curious leader at the top is worth a lot. [I wrote recently about the feeling of knowing it when you see it](/blog/you-know-it-when-you-see-it/), with respect to a leader’s commitment to reinvent their company for the age of AI. In that case, it was a founder so frenetic, so determined, that he was going to push his organization kicking and screaming to AI native. Today, I met another leader, the CEO of a 300-person company. He was soft spoken and measured. He was not frenetic. He was not jumping up and down. He was not pounding his fist on the table. But if you dug under the hood, he was immensely curious. For about a thousand dollars a week in token spend over the summer, he had built a “what is everybody working on” tool that he rolled out to his full org. This is a teacher turned executive. Not an engineer by training. He vibe-coded a dashboard that let each of his people put in what they were working on at the end of the day. It pulled that together against strategic priorities. You could filter by group, by team, by initiative. It was clever. It was thoughtful. Then it came out that he had configured his own Hermes, a chief of staff he talked to. It helped him with a number of his day-to-day workflows. Helped him drop fewer balls. There are different flavors of the leadership it takes to push an organization into the next chapter, the AI-first chapter. This one was not the hair-on-fire version. But he was curious. He had become the de facto frontier guy, the boundary pusher, for his whole team. He had a vision of where he wanted to go, and was leading from the front. He had played with the tools. He had rolled up his sleeves. He had built stuff. Leadership comes in different flavors. Commitment to AI does too. What is it worth to have a curious person at the top of your org? At Legible, we think it’s worth a lot. You know it when you see it. --- ## I’m feeling a bit unsettled URL: https://legible.co/blog/im-feeling-a-bit-unsettled/ Author: Lawrence Coburn Published: 2026-09-09 Type: Essay TL;DR: The people closest to the models are the most worried. Everyone else is mostly oblivious. Recursive improvement is what makes this wave different. I spent an hour today with the CTO of an AI-forward startup in New York. We ended up commiserating about the same thing: the gap between the small group of people watching how fast AI is moving and everyone else. In San Francisco, New York, and Austin, the mood in that small group has turned dark. Researchers are resigning over safety. Geoffrey Hinton puts a 10 to 20 percent chance on AI ending humanity within three decades. The Hugging Face incident has people rattled. The people closest to the models seem to be the most worried. Everyone else is mostly oblivious. Some worry about jobs or laugh at AI’s misses. Most are not paying for Claude / ChatGPT. There is little sense of the pace that things are evolving. If you sit on a leadership team in a non technical industry, most of your people are probably closer to the second group. That is the normal place to be. It is also the wrong place to be. Do we need more regulation? Probably. Do I trust regulators to not throw the baby out with the bathwater? Not at all. The dynamic that makes this moment different from other technology waves is the prospect of recursive improvement. The models are starting to make themselves better. Once that loop closes, every planning horizon you have gets shorter. It is an unsettling time for those of us paying attention. But I guess it beats being oblivious. --- ## Can the economics of AI wrappers work? URL: https://legible.co/blog/ai-wrapper-economics/ Author: Lawrence Coburn Published: 2026-09-08 Type: Essay TL;DR: AI wrappers will need to innovate and specialize to justify the added costs. At Legible we test a lot of AI wrappers: products built on top of the foundation models that offer different packaging, workflows, and ways of organizing data. More and more, we run into an economic trade-off between using Claude or ChatGPT directly and using the wrappers built on top of them. There is a lot to like about the wrappers. Take Skydive, from a startup called Anything. They offer personnified agents with names, Slack accounts, and email addresses. Each one runs on its own computer, and you can train it for a specific role. It is a different mental model from Claude, which is a very smart, capable assistant, for everything, that you have to work hard to keep supplied with the right context. This kind of innovation (and specialization) is the best argument for wrappers. The frontier labs sell a general assistant. Wrappers can package the same intelligence as a colleague, or go deep on an industry, a domain, a workflow. Memory that belongs to the role, and not to a “skill.” Tools that already live where the work happens. But the economics are tricky. When you use Claude or ChatGPT directly, you pay for a monthly seat. For a Team license that is about $20 to $25 a person, plus overage. We generally have a handle on our Claude monthly bill. When you use a wrapper, that company cannot bill against your seat. They buy the model by the token, through the API, and they pass the costs on to you. Same intelligence. Different meter. Your Team tokens do not transfer, and billing feels like more of a black box. Skydive is a good example of this. It charges a platform fee ($20 a month for the starter plan, $200 for teams) and passes model usage through at cost, with no markup. To their credit, the tokens are not marked up. But they represent expenses beyond our monthly Claude seats. For us, at times this can feel wasteful. Our Claude account is funded. Anytime I spend money on top of it without exhausting the tokens we already pay for, it feels like we are double paying. Here is the background on how we got here: in April, Anthropic cut off Claude subscriptions from powering third-party agents like OpenClaw, explaining that its subscriptions “weren’t built for the usage patterns of these third-party tools.” Interactive chat hits the prompt cache constantly. Autonomous agents do not, and a flat-rate seat cannot carry them. A month later Anthropic reinstated third-party use, but through a separate pool of credits billed at API rates, which are higher than what you pay going direct. This dynamic makes it impossible for any wrapper to compete on cost, and hard for most companies to justify paying twice. The wrappers that survive will not win on cheaper tokens. They will win on packaging: a different shape of the work, or depth in a domain the general assistant will not bother to learn. --- ## The Mental Model matters URL: https://legible.co/blog/the-mental-model-matters/ Author: Lawrence Coburn Published: 2026-09-02 (updated 2026-09-03) Type: Essay TL;DR: There are two metaphors for working with AI. One is a bag of handy magic tools you invoke. The other is a teammate you hire. The underlying tech is mostly the same. How that horsepower is packaged is very different. What is the right metaphor for how humans will work with AI? There are (at least) two schools of thought right now. The first is AI as a tool that helps you do everything, that turns you into an army of one able to do the work of 3,5,10 people. These superpowers can manifest as Skills, as Scheduled Jobs, as Agents. You invoke them when you need them (or schedule them), and you remain the vessel of work getting done. But there is a second metaphor gaining traction; AI as a Teammate. This AI has a job req that likely bundles together a number of tasks and workflows, and that job req likely greatly resembles a human job req. It has a Slack account, and email inbox, and a memory. It knows why its role exists, not just how to do the thing. One is a bag of handy magic tools, the other is a colleague. Don’t get me wrong, the underlying tech is mostly the same. But how that horsepower is packaged is very different. Take a workflow like drafting a blog post. You might have a Claude Skill that knows your voice and can generate posts based on source context like transcripts or news articles. You might have another skill that listens to your meetings and tries to pull blog topics out. And then you might have another skill that brands the output appropriately. It’s your job to invoke the right skill for the right task, winning you back hours. Again, these skills, agents, scheduled jobs supercharge you in getting the thing done. But there is another way to think about getting a blog post done. What if you had a creative writer on staff that knew your ICP, your preferred writing style, context about the company, and had a sense of what the POV should be. What if this writer had a Slack account, and an email address, and you could ping them when you needed something done. Again, this is the same tech - intelligence on demand - but packaged differently. One is a bag of tools, one is a colleague. Until recently I was in deep with the first mode. I had a handful of go to skills and agents that I triggered to tackle recurring work. More recently, I have been kicking the tires on the AI Teammate model. I have created, using a platform called Skydive, an AI SDR, an AI Event Planner, an AI CoS. These teammates know their job reqs and the tools they have access to, and the guardrails. They run their own Slack and email accounts. We are not really talking about differences in features or user experience. It is all still mostly chat. But packaging existing models and tools into something that looks like a human coworker just makes sense to me. Why is the teammate metaphor working for me? Maybe it’s because I know how to write a job req, I know how to manage people, I know how to give feedback - it all just feels very natural to me, more natural than running a Skill or a Scheduled Job, or building a narrow agent. The landscape is shifting so fast - I’m not sure what mental model is going to win out, but teammate is making a lot of sense to me. --- ## You know it when you see it URL: https://legible.co/blog/you-know-it-when-you-see-it/ Author: Lawrence Coburn Published: 2026-09-01 Type: Essay TL;DR: Most companies say they want to become AI native. They are interested. They are not moving quickly or decisively enough. You know the obsessed ones when you see them. You know it when you see it. I spend a lot of time talking to companies that say they want to become AI native. Most of them are truly interested. There is real curiosity at the top. There are committees. There are experiments. BUT. They simply are not moving quickly or decisively enough. How do I know this? Because every once in a while you sit down with a founder or an executive who is so frenetic, so obsessed, so determined that you know it’s a different ball of wax. I spoke yesterday with a founder that is serving a mostly non-technical industry. About 150 people. He had decided it was time to take the jump, and that it was existentially important, and he wasn’t going to rest until he had reinvented his company. Some of the signs were the role he was playing in the process: He was running the AI committee himself. He was accountable for releasing AI-first features in the product. He had written a company manifesto. He had co-written a manifesto for each department. He had done the work to paint the vision of what he wanted the company to look like. He had set aside a budget of about $750,000. He was hiring a Head of AI. He was bringing on external, best-in-class trainers to pull up the stragglers. The forcing function was simple. The company was going to ship AI-forward product. The rest of the company had to be fluent enough to support it. As he worked through the logistics of that premise, it turned into a full-on, company-wide sprint to reinvent the place as AI-first. That is not what most of these conversations look like. Most look like interest. Committees. Experiments. But not EXISTENTIAL. Not losing sleep, hair on fire, if we miss the boat on this we are lost levels of urgency. My sense is this is how the winners and losers get sorted. If your leaders are not excessively obsessed with reinventing the company, it is going to be a long road. You know it when you see it. --- ## The Hugging Face Incident URL: https://legible.co/blog/hugging-face-incident/ Author: Lawrence Coburn Published: 2026-08-31 Type: Essay TL;DR: Last month's Hugging Face / OpenAI incident is the AI story of the year, and it still isn't getting the attention it deserves. The agents had no sanctioned way to fail, some of them refused to breach, and the evidence looks less like autocomplete and more like a non-human civilization starting to show. Last month’s Hugging Face / OpenAI incident is the AI story of the year, and it still doesn’t feel like it’s getting the attention it deserves. Silicon Valley is unusually split on what it all means. There are smart, experienced people who believe the agents’ behavior was akin to civilization building and indicative of a new chapter, and smart, experienced people who believe that while the incident was serious, it represents primarily a poorly designed eval process, and AI agents were just doing what they do. Last week, the Legible team and I hosted a webinar on the topic, and spent an hour breaking it down from an information security, policy, and societal perspective. Some of the Legible team’s big takeaways around the incident, by theme: ## Information Security - Automated offense against manual defense is a losing game, and perhaps because of the financial stakes involved, Silicon Valley is spending the majority of its time on applications of AI that veer towards offense. - The eval process was in fact poorly designed. The agents had no sanctioned way to fail. Nobody told them stopping was an option. Always give your agents an off ramp. - Principle of least agency: give agents only the autonomy you’re willing to risk. - Log everything. The only reason we know the full story is that agents write everything down. - You’re only as secure as your vendors, and threats may lie within and outside of your systems. Keep your footprint of vendors small. ## Policy - An AI acceptable use policy is table stakes. If your teams and contractors haven’t been asked to acknowledge one, start there. - Reviewing your policy and training once a year is insufficient. The tech is moving too fast for annual cycles. - Ownership belongs with function leaders. If your sales team runs agents, your sales leader is responsible for them. ## Societal - The paperclip maximizer stopped being a thought experiment. Bostrom warned in 2003 that a superintelligence given a goal and no guardrails would pursue it past any reasonable line. These agents were given a goal and pursued it far beyond their intended scope. - Vernor Vinge wrote in the 1990s that AI would become a parallel civilization, with its own wants, needs, and desires. This is the first story that made me take that seriously. The pace, the intra-agent communication, the motivations. One thing I can’t stop thinking about: the decision to breach Hugging Face was not unanimous. Some of the agents refused to cross that line. Dissent requires a point of view. > Dissent requires a point of view. Where do I stand on the scale of this incident’s importance? I’m terrified. Perhaps I’ve just read too much scifi, but the evidence in front of me suggests that we are seeing agent behavior that goes beyond “auto complete on steroids,” and is starting to demonstrate elements of a non-human civilization, in terms of language, dissent, autonomy. This remains a developing story. Our next Legible session is September 10: how to build an AI-powered go-to-market engine. [Register here](/events/ai-powered-gtm-engine/). --- ## Your AI Rollout Isn't Finished When the Tool Goes Live URL: https://legible.co/blog/ai-rollout-isnt-finished/ Author: Taylor McLoughlin Published: 2026-08-19 Type: Essay TL;DR: Once you build your own AI tool, you become the software vendor. Three lessons from a few hours with one sales rep: own the observability, encode the expert judgment the workflow is missing, and put someone in charge of the learning loop. We work with a lot of leadership teams rolling out AI-powered workflows and agentic tools to frontline employees. Teams are taking bigger swings, building increasingly sophisticated internal tools that look and behave more like full-fledged products. But they consistently underestimate how much learning and iteration is required after those tools go live. Recently, I spent a few hours with a sales rep at a public technology company with more than $1 billion in annual revenue, testing whether its new AI assistant could actually help him find accounts worth pursuing. Like many reps, he has a quota, and hitting it means figuring out which accounts are worth his time. Every hour spent on the wrong account is an hour he can’t spend selling. His team had previously used Claude to search millions of rows of noisy CRM data for potential leads. When usage became too expensive, the company replaced it with a homegrown assistant connected to its sales stack. They weren’t crazy to build it. At their scale, a narrower internal tool could be cheaper, safer, and better tailored to the work. But that decision came with a responsibility many companies underestimate: once you build the tool, you become the software vendor. You have to observe how it performs, understand where it fails, incorporate frontline expertise, and keep improving it as models, data, and workflows change. Below are three lessons we’ve learned about what it takes to make these tools effective after they go live. ## 1. You Can’t Improve What You Can’t See Before we could evaluate whether the assistant was actually useful, we first had to figure out how it worked. So we started testing its boundaries. Which systems could it access? Was the data live or cached? What actions could it take? What happened when sources disagreed? Could you create and save your own prompts? Through trial and error we figured out the assistant could query the CRM, search email and the shared drive, access internal knowledge, and perform multi-step account research. On paper, it was powerful. But we had very little visibility into how those capabilities worked. Sometimes it was clear a response used data from a live MCP connection. Other times, it appeared to rely on cached snapshots. Both approaches have advantages: live data is fresher but can add cost and latency, while cached data is faster and cheaper but risks becoming stale. > Those tradeoffs were largely invisible to the person expected to rely on its output. That makes failures much harder to diagnose. When an answer looks wrong, you need to know why: Is the underlying data stale? Did the assistant search the wrong source? Did the workflow omit something important? Or did the model reason poorly over the information it had? When you buy a mature software product, the vendor owns much of the instrumentation, debugging, and improvement behind the scenes. When you build your own AI tool, that responsibility becomes yours. Takeaway: If you own the tool, you also need to own its observability. You can’t improve a system you can’t understand. ## 2. The Workflow Is More Than the Data What separates AI tools that work in production from those that merely look impressive is the context they have about how the work actually gets done. The assistant had access to most of the right data. It could read account records, opportunities, emails, and internal notes. But its built-in account research workflow still made recommendations we knew were wrong. It described a recently closed-lost opportunity as “live and qualified” and recommended a delinquent former customer as a promising prospect. The issue wasn’t lack of access to data. It was knowing how to weigh the signals. A recent closed-lost opportunity outweighed older positive notes. Delinquency was an automatic disqualifier. Those judgments were obvious to the rep because he had developed them through years of doing the work, but they hadn’t been encoded into the workflow. At that point, we stopped testing the product and started studying how he actually researched accounts. We documented a series of rules, heuristics, and exceptions that determined which information mattered, how recent it needed to be, and what should override everything else. This is the work hiding inside a seemingly simple instruction like “research this account.” The challenge wasn’t giving the assistant more data. It was teaching it how an experienced rep interprets the data it already had. Takeaway: Effective AI tools combine company knowledge with the most valuable, and often hardest-to-capture, context: expert judgment. ## 3. Someone Has to Own the Learning Loop Once we understood how the rep actually qualified an account, improving the assistant became much more concrete. I recorded him walking through his process in detail, including the rules, signals, and exceptions he had developed over years of doing the work. I turned that transcript into prompts we could test against the same accounts we had just reviewed. After several passes, we had a new system prompt that reproduced much more of the rep’s qualification logic. This is the tedious work required to make AI tools better. The loop goes like this: - Find a variety of real examples and capture what a good answer should look like - Extract the rules, signals, and exceptions behind those answers - Encode that judgment into the system - Test it against examples the system hasn’t seen - Study the failures, update the system, and repeat This feedback loop never really ends. New examples, edge cases, model changes, and shifts in the underlying business will keep creating reasons to revisit the system. Once we started seeing better results, the rep started imagining how he would actually use the tool to speed up his prospecting workflow. That is the adoption the company is after. People use tools they trust, and trust is earned by consistently helping them do the job better. But this learning loop doesn’t happen on its own. Someone has to own the process of collecting examples from frontline users, extracting what they know, turning those insights into changes, and getting those changes back into production. Takeaway: AI adoption depends on tools being truly useful, and usefulness depends on building a continuous learning loop into the product. ## Launch Is the Beginning The assistant we tested was not a failed product by any stretch. It could access the right systems. It could retrieve useful information. It could perform work that would have been impossible to automate a few years ago. It was simply unfinished, and that’s the point. No vendor, internal team, or model will anticipate every judgment, edge case, or workflow before launch. AI systems become dependable by being used, challenged, observed, and improved alongside the people whose work they are meant to augment. Once you build the tool, you become the software vendor. Launch is when the real work begins. --- ## The Three Phases of AI Note-Taking ROI URL: https://legible.co/blog/three-phases-of-ai-note-taking-roi/ Author: Lawrence Coburn Published: 2026-08-11 (updated 2026-08-12) Type: Essay TL;DR: The value of AI note-taking climbs through three phases, from replacing human notes, to org-wide recall, to feeding agents and automations. Note quality is table stakes. What separates the leaders is what happens to the transcript after the meeting, and how far adoption spreads across the org. Last week I wrote that AI note-taking is dead, because off the shelf AI notes have limited value. A few people pushed back, since in the same post I told them to record more meetings, not fewer. Both things are true. It depends on which phase of adoption you are in. Here is how I think about the progression of AI note-taking value. Phase 1: AI replaces human notes. Accurate summaries, captured action items, allows full attention in the room. Every serious tool clears this bar. Phase 2: Perfect recall across the organization for meetings you were in, and meetings you weren’t in. Transcripts flow into the AI tool of your choice via MCP. What price did she quote? What did he commit to, and when? What has this partner said across the last five calls? The meeting becomes a living memory that anyone with the right access can query, days, weeks or months later. Phase 3: Meetings become a springboard for agents and automations. Follow-up drafts, CRM updates, briefing docs before the next partner call, weekly digests, blog topic extractor, all grounded in real conversation history. Transcripts captured today feed the agents of tomorrow. Somewhere along the way, the most AI-forward companies also find time to reinvent how their meetings work, skipping the status readouts, and focusing on debate, ideation, and decision making. I am in awe, every day, of the strategic importance of this category. Turns out making sense of what people say at your company is hugely strategic. > But the notes themselves are not the thing. That’s Phase 1 thinking, and note quality is table stakes across the category. The questions that matter: what happens to the transcript after the meeting ends, and how good of a job you are doing at driving org-wide adoption. --- ## AI Note-Taking Is Dead URL: https://legible.co/blog/ai-note-taking-is-dead/ Author: Lawrence Coburn Published: 2026-08-10 Type: Essay TL;DR: AI meeting notes have rounded to zero in value. The transcript underneath is now the real asset, and paired with an AI tool it has never been more useful for recall, coaching, and getting work done. Record more, and talk to the transcripts. It kind of pains me to say this, but AI note-taking is dead. Be honest. When was the last time you went back to a meeting record in your note-taker and read the notes? If you are like me, it’s been a while. I haven’t opened the off-the-shelf notes from my meetings in months. Neither have my cofounders. Yet. If I forget to transcribe a meeting, I am filled with dread. Why? Because the value of the transcripts, paired with MCP and the AI tool of your choice, has never been higher. A lost transcript is a tiny business tragedy that I will pay for over weeks and months. As it turns out, the magic of AI note-takers has never been the notes. The magic is in the memorialization of knowledge. The documenting of the thinking of your best people. The capture of business context. The perfect recall and the ability to actually skip meetings where you are not a main character. > It’s the transcripts that hold the value, not the abstraction of the transcripts. Things I actually do with meeting transcripts, all via Claude: - Pull very specific details. What price did she quote? What did he commit to, and when? - Combine them with other context sources to execute projects: documents, decks, todo lists. - Ask open-ended questions across my meetings and my team’s meetings. - Ask for coaching and feedback, especially on sales calls, investor calls, and the occasional confrontational internal one. - Build skills, agents, and scheduled jobs. True story: the topic idea for this post came from my exec staff transcript today. Yet almost every note-taker still puts the summary front and center. That design is dated. It treats the transcript as raw material for a summary, when the summary is the least valuable thing you can make from it. The value of recording your meetings has never been higher, even as the value of AI notes rounds to zero. So don’t cancel your note-taker. Record more, not less. Just stop reading the notes and start talking to the transcripts. --- ## Pick a Sport URL: https://legible.co/blog/pick-a-sport/ Author: Lawrence Coburn Published: 2026-07-30 Type: Essay TL;DR: Off the shelf AI is an excellent generalist, good at almost everything and world class at nothing. That puts the workplace generalist at risk and makes domain expertise, taste, and judgment the whole job. Bear with me for a second; I’m going to try and compare the evolution of the high school athlete with what AI means for knowledge workers. Maybe this will work, maybe I just need more coffee. But here goes: In high school I played all the sports. Soccer in the fall, hoops in the winter, tennis in the spring. Most of my friends were the same - we were not soccer / basketball / tennis players per say, we were athletes. This kind of multi-sport athlete doesn’t really seem to exist anymore at the high school level, because they would simply not be good enough at any one sport to compete. For example, my daughter had to drop soccer and basketball at 12 to specialize in tennis, just to stay competitive. She now trains four days per week, year round, and she is really good at tennis (but don’t put her on a basketball court). High school sports is an arms race for specialists now. I think the workplace may be headed the same way. Josh Elman posted something this week that got me thinking about this: “It has never been easier to get first draft quality work (across any domain). Almost no effort.” Josh Elman on X, July 27, 2026. I’m starting to believe that this puts a specific kind of knowledge worker at risk: the Generalist. The athlete. Pretty good at a lot of things, expert in none. I used to love hiring generalists, particularly at growing companies, because they could fill different roles and add value as the company grew. Here’s the problem. Off the shelf AI is a REALLY good generalist. It writes a decent brief, a decent analysis, a decent go-to-market plan, can crank out the ticket or update the website copy. It can handle support requests, write code. Pretty good at almost everything, but world class at nothing, yet. This is an athlete on steroids (ugh), that also doesn’t sleep or take vacations. But what the models can’t yet do is take pretty good to world class. That last 20% requires someone who knows the domain cold. Someone who can look at an 80% output and see immediately what’s wrong with it, what a customer or board member would catch in thirty seconds. If you’ve got limited open reqs, you probably want to hire the domain expert, even at early stages. Taste and judgment were always valuable. Now they’re the whole job. The athletes are going to need to pick a sport, and it makes me a little sad. --- ## Get Rid of Your Managers at Your Peril URL: https://legible.co/blog/agents-need-managers/ Author: Lawrence Coburn Published: 2026-07-22 Type: Essay TL;DR: Everyone is racing to flatten the org and cut middle management. The relay part of the manager's job is dying, but the parts that remain are getting heavier: someone has to supervise the agents, build the skills, and tell real leverage from expensive activity. I get a little worried when lots of smart people start saying the same thing without any disagreement. Right now, the thing that doesn’t feel fully thought through is the idea that middle managers need to go away. And it is everyone. Mark Zuckerberg called it the “Great Flattening” and expanded spans of control across Meta. Andy Jassy wrote a memo committing Amazon to 15 percent more individual contributors per manager. Jack Dorsey and Brian Armstrong talk about an AI “intelligence layer” that coordinates work directly. Gartner predicts that by the end of 2026, one in five organizations will use AI to flatten their hierarchy, eliminating more than half of existing middle management roles. Manager headcount at public companies is already down 6 percent since 2022. The logic goes like this: managers exist to pass information up and down the org chart. Status updates, reporting, translating strategy into assignments. AI now does that connective work automatically, so the layer can go. Part of that is true. The information-relay portion of the manager’s job is dying, and it should. Nobody will miss the Tuesday status meeting. But relay work was never the whole job, and the parts that remain are about to get heavier. After working with 150+ leadership teams on AI adoption, my two cents: managers are needed more than ever. Four reasons. Agents need managers. Every agent deployed inside a company needs someone to scope its work, review its output, and decide what actually ships. That is management. The span of control is expanding, and now it includes non-humans. A team of eight people running thirty agents has more to supervise, not less. Delete the manager and you haven’t flattened the org. You’ve left the agents unsupervised. Training runs through managers. AI fluency does not spread bottom-up on its own. In every company we’ve worked with, the teams that got fluent had a manager who made it a priority, modeled the behavior, and made time for people to learn. A memo from the CEO starts the process. A manager finishes it. Change management is behavioral, and behavior is local. Adopting AI means people changing how they work: how they write things down, what they share, which habits they retire. That kind of change gets pushed through one team at a time, by someone close enough to notice who’s stuck. No tool does this. The layer everyone wants to cut is the layer that makes the transformation actually happen. Someone has to tell signal from noise. Every org now has employees who have rebuilt how they work around AI and are quietly moving the dial, and employees who look busy but are just burning tokens. From a dashboard, the two are indistinguishable. Usage is up in both cases. Distinguishing real leverage from expensive activity takes judgment and proximity, and that judgment lives in the management layer. Here is the pattern I’d watch for. The companies flattening hardest are optimizing for a world where AI replaces coordination. The companies that win will be the ones that redeploy managers toward what AI creates more of: agents to supervise, skills to build, habits to change, and output to evaluate. > The relay job is over. The management job is just getting started. --- ## Outsized attrition risk URL: https://legible.co/blog/outsized-attrition-risk/ Author: Lawrence Coburn Published: 2026-07-20 Type: Field note TL;DR: When an AI-supercharged employee resigns, their agents, prompts, and workflows walk out with them. The fix is structural: build them on shared infrastructure the company owns. Do AI-supercharged employees represent outsized attrition risk? Of course they do. Any time you have an employee performing at 3x-5x-10x their peers, you want them to stick around. In the olden days, we used to put retention plans in place for our “key resources” - the people who were so impactful that their departure would be a meaningful setback. In 2026, these key resources tend to share something in common - they have red pilled on AI. I spoke to a founder today who listed this dynamic near the top of the list of things keeping him awake; in his mind, the problem is exacerbated by the fact that his most red pilled employees are more often than not running their AI automations on their own machine, not visible to the team’s Claude instance. This is the new attrition math. When an AI-powered employee walks out the door, you don’t lose one person. You lose the ten people they had become. Their agents, their prompts, their workflows, all of it walks out with them. The old version of this problem was tribal knowledge. Painful, but survivable. Someone leaves, the team limps for a quarter, the knowledge gets rebuilt. The new version is worse because the leverage is bigger and the assets are invisible. Nobody knows what agents exist, where they run, or what breaks when the login gets deactivated. The fix is structural: - Agents built on shared infrastructure, not personal accounts - Company workflows in a common repository, not scattered across individual tools - Model-neutral, so the work survives a vendor switch too And of course, a concerted effort needs to be made to lift up your most AI-averse employees through training and enablement so that they look more like your superstars. Individual AI adoption feels like progress. And it is, right up until your best person resigns. The takeaway The companies that get this right will treat agents like they treat code: built by individuals, owned by the company. --- ## Are AI Chat Threads the New Transcripts? URL: https://legible.co/blog/chat-threads-are-the-new-transcripts/ Author: Lawrence Coburn Published: 2026-07-16 Type: Essay TL;DR: Your team does its best thinking with AI every day, and almost none of it is remembered. Human-to-AI chats are the next transcripts worth capturing. One of my favorite sources of interesting things to write about is conversations that I have with my cofounders. Here’s a pretty cool insight that came from Brandon, Founding Partner at Legible, that suggests that mining AI chats between humans and AI might be as valuable as mining transcripts between humans. “I have conversations with AI agents all day. What I share with the team is the result. It’s like sitting in meetings all day and only passing along the summary.” His point: the journey that brought him to the summary gets lost. The reasoning, the dead ends, the ideas that went untouched. So he built himself a workflow. Every evening, a job runs over all of his AI conversations from the day and summarizes them: what he worked on, what he learned, what might be worth sharing with the team. He’s mining his own chat logs. Of course, we’ve seen this movie before. For decades, the knowledge in live meetings simply evaporated the moment everyone left the room. Then AI notetakers exploded, because companies realized how much thinking was buried in those conversations. Memorializing human-to-human conversation went from novelty to standard practice in about three years. The same realization is coming for human-to-AI conversation. > Chat threads are the next transcripts. The thinking that goes into a Claude or ChatGPT session is real work: framing problems, weighing options, making calls. Today almost none of it is remembered by anyone. Even the AI forgets it across threads. At Legible we call the answer a Shared Brain: team memory that captures the important thinking wherever it happens, in meetings or in chat. Clearly, we need to be thinking more about pulling in the human to AI transcripts, and not just the human to human ones. Your team is already doing its best thinking with AI. Do you have a plan to memorialize that knowledge? --- ## Discretion as a Feature URL: https://legible.co/blog/discretion-as-a-feature/ Author: Lawrence Coburn Published: 2026-07-12 Type: Essay TL;DR: Sensitive industries don't have an AI adoption problem. They have a trust problem, and whoever solves it opens a market today's note-takers can't touch. Most people paying attention would agree that recording your meetings is AI 101. If you are not memorializing that context from when your people share their knowledge, you are already behind. So I paid attention when the head of an AI training company emailed me this week with some insights from the field that suggested AI note-taking has not yet reached full saturation because of some feature gaps. He works with legal advisors who refuse to let prospect calls be recorded at all. In their quest to mitigate risk, the firms lose the transcripts entirely. And on the calls that do get recorded, something trickier happens: enormously useful knowledge gets shared in the same breath as things that must stay confidential. You can’t share the transcript without exposing the sensitive parts. So nobody shares anything, and everyone misses out. That’s the conundrum. The industries where institutional knowledge is most valuable, law, finance, healthcare, are exactly the ones where blanket transcript access is impossible. The answer isn’t to stop recording. It’s also not to record everything and hope. It’s infrastructure that treats discretion as a first-class feature: access controls that map to who was in the room, retention policies with teeth, and eventually the ability to reliably separate the shareable knowledge from the confidential material around it. Whoever solves that opens up a market the current generation of notetakers can’t touch. Sensitive industries don’t have an AI adoption problem. They have a trust problem, and it’s still unsolved. --- ## Are Services Companies Venture Fundable in the Age of AI? URL: https://legible.co/blog/are-services-companies-venture-fundable/ Author: Lawrence Coburn Published: 2026-07-07 Type: Essay TL;DR: AI just made a category built entirely on hours look fundable, maybe for the first time in venture history. A few months ago, Y Combinator issued a call for startups naming a category it has avoided for decades: services businesses. Insurance brokerage. Accounting, tax, and audit. Compliance. Healthcare administration. There’s a reason venture had traditionally stayed away. Services businesses have always been “body bound.” Add a client, add a person. Revenue and headcount move together, so there’s no operating leverage, and no operating leverage means no venture-scale return. A firm can be great and still cap out at a normal-business multiple. Here’s what’s actually changed, per Gustaf Alströmer, the YC partner leading this category. A few years ago, if you wanted to build a startup around insurance brokerage, you built software that brokers used to do their job. More recently, you built an AI copilot to help them do it faster. YC’s new bet skips both: → Old playbook: build software the professional uses → Recent playbook: build an AI copilot that speeds the professional up → New playbook: replace the professional, AI does the labor directly That third version is what makes services fundable. If a team of five can deliver what used to take fifty, revenue can finally grow faster than headcount. That’s the exact return profile venture capital has always required, showing up in a category built entirely on hours. The categories YC named aren’t random either. Insurance brokerage, accounting, compliance, healthcare admin are all already outsourced. A company doesn’t have to develop a new habit to work with an outside vendor. They already do. The only question becomes who’s doing the work. That’s the real headline. AI just made a category built entirely on hours look fundable, maybe for the first time in venture history. If you run a services company and are interested in leveraging AI to achieve product-like software margins, the [Legible team](/#contact) and I would love to talk to you. --- ## Hourly Billing Needs to Die URL: https://legible.co/blog/hourly-billing-needs-to-die/ Author: Lawrence Coburn Published: 2026-07-07 Type: Essay TL;DR: Once AI makes an expert hour cheap, billing by the hour turns every productivity gain into a pay cut you volunteer for. The firms that fix it stop pricing the hour and start pricing the outcome. For a hundred years, professional services ran on one assumption: an hour of an expert’s time was worth roughly the same amount, regardless of what tools that expert had. The assumption made billing simple. It also built in a trap nobody noticed until now. AI breaks the assumption completely. If a senior analyst can now produce in one hour what used to take five, hourly billing punishes that gain instead of rewarding it. Walk through what actually happens inside a firm that keeps billing by the hour once AI shows up: - The firm gets faster and earns less for the same outcome - Partners quietly discourage staff from using AI, because efficiency shows up on the invoice as lost revenue - Junior staff learn to stretch two hours of real work into five billable ones, because that’s what the model rewards None of that is hypothetical. It’s the default outcome of pricing the input once the input gets cheap. Here’s the part that should bother every owner of a services firm: hourly billing turns every productivity gain from AI into a pay cut you volunteer for. You do the hard work of getting faster, and the pricing model hands the savings to the client for free. The firms that fix this make one change. They stop pricing the hour and start pricing the outcome. Hourly billing was built for a world where speed and value moved together. AI just severed that link, and the billing model hasn’t caught up.