AI Knowledge Capture: Keep the Systems Your Team Builds

Somewhere on your team, one person has quietly become the AI department. They've built the prompts, the scripts, the automations that make their corner of the company run faster than the rest. Now give them a job offer from an AI lab at three times the salary. If they take it, you won't be able to figure out what they built for you — where it lives, how it works, or how to run it tomorrow.

The takeaway: AI work leaves traces — chats, prompts, skills, scripts, tooling. Companies that capture those traces turn one employee's experiments into durable assets; companies that don't are renting their AI capability from whoever happens to stay. Write the capture requirement into your AI policy before the first resignation, not after.

What follows are notes from this week's Executive AI Roundtable discussion, shared under the Chatham House Rule.

The Traces Are the Asset

Employees leaving with know-how in their heads is an old story. What's new is that AI work writes itself down as it happens. As one attendee put it: "All the stuff that you're doing with AI creates traces." The chats, the prompts that worked, the skills files, the little tools — every session leaves a record that could become a company asset, if anyone bothered to keep it.

The same attendee pushed the point further: companies that keep those traces "have the capability to clone their employees." Not the person — the system the person built. The prompt library that handles renewals, the script that triages support, the context files that make the AI useful in your business instead of generically smart. Capture converts hours of one employee's labor into something the next employee inherits.

Most companies capture none of it. The workflows live in personal Claude accounts, on personal laptops, in chat histories the company has never seen. Leadership knows AI is being used; nobody can say where the value is accumulated. You don't need the resignation letter to be losing the asset — you're losing it every week it stays uncaptured. The fix costs almost nothing before that letter arrives: a policy line that says AI work lands in company accounts and repositories, plus somewhere shared for it to land.

Rules Need Incentives

The policy conversation usually starts and stops at control: approved tools, banned tools, and logging. One operator's warning about that instinct — in too many companies, "control is more important than value creation." Lock the tooling down hard enough, and your best people go home, use the better model on their personal account, come back with great work, and get celebrated for it. The work never touches the company infrastructure. The control you bought guaranteed the capture you lost.

Two fixes came out of the discussion. First, every policy needs an escape valve: an easy, fast path to say I need Claude Max or I want to try the new model and get a company credit card and a logged account the same week. The request channel matters less than the speed — a slow exception process is a ban with extra steps.

Second — "games are made up of rules and incentives," and the incentives get far less attention than the rules. A policy tells people what they may do; it doesn't make them want to capture what they build. I heard about a department at a large tech company that reportedly burned through its entire token budget in a couple of months. Celebrate usage and usage is what you get — spend without capture, motion without assets. Incentives should reward the artifact (the skill written down, the workflow that survives its author), not the token count.

Where you are in the adoption curve changes the calculus. "An existing company is necessarily a brownfield" — built terrain, not greenfield — the conservatism has to be broken before capture matters, and a usage push might even be right for a while. A younger company needs the discipline, not the push. As one operator summarized the underlying failure mode: "Most people don't manage to change. They manage to stability — to stasis."

If you don't yet have the written policy these rules live in, the free AI Policy Generator drafts one in about ten minutes — including the tool-request escape valve.

The Token Bill Comes Due

One attendee named the squeeze AI-dependent products are feeling this year: "The cost of using AI was significantly less, and now that it has to pay for itself, something that seemed profitable last year is no longer this year." Not everyone in the room bought the framing — per unit of capability, model prices have fallen; what keeps climbing is the appetite for bigger models doing bigger jobs. Either way, the bill lands on someone.

Descript came up as the case study. The promise is real — edit the transcript, and the video edits itself. The room's read: the economics leak. Skimp on transcription quality and the product's foundation cracks, while every minute of processed video burns tokens the customer's subscription may not cover. A product that burns tokens on every job doesn't earn software margins — and products like it are still priced at them.

One investor's prediction: within months, the vendors will stop eating that bill. "The vendors are no longer in the business of reselling tokens; the vendors are in the business of doing their software." Expect "log in with your Anthropic account" to become the default — the token cost moves to the customer, and the vendor has to win on workflow, not arbitrage. The deepest-pocketed players are solving it the other way: model companies absorbing the model-dependent tools, as in SpaceX's $60 billion acquisition of Cursor after its merger with xAI.

The same investor sees the flip side as the opportunity: plenty of former "startups" that stalled into small businesses — teams big enough to cover payroll, growth too slow to matter — are "cash-cow-level businesses that are significantly undervalued right now, because of AI." Replace enough of the labor line with AI and captured systems, and a break-even lifestyle business becomes a profitable one.

Step One Is Not the Destination

"Going from human labor to AI isn't the end. It's just step one." The sequence behind that line runs in three phases — a human understands the problem, then the AI is taught to understand it, then the understanding gets re-expressed as deterministic code that runs for close to nothing. Every phase drops the cost by orders of magnitude. That token bill is largely a symptom of stopping at phase two.

I use the same move constantly: have the AI build software tooling, and the tooling runs nearly free on your local machine forever. The AI's job was to write the calculator, not to be the calculator.

Local models extend the same logic to the model itself. I dictate through FluidVoice, which runs NVIDIA's speech model on my Mac — faster than the API versions, because there's no network round trip. Quantized builds of Whisper now ship in a 500 MB download. Local models aren't smart enough yet for general reasoning, but for transcription-shaped work, they're free, fast, and private — and the expectation in the room was that compute costs fall precipitously over the next 24 to 48 months as new fabs come online.

One attendee's rule: "If I'm using AI as a tool, I'm using it wrong." A tool keeps you in the inner loop — you steer every step. A system runs the outer loop and pulls you in only for judgment. That's the climb the four levels of AI adoption describes — chat to files to tools to triggers: the chat level is where work evaporates, and each level up is where it starts to compound.

Ship, Then Watch

"Merging a PR does not close the issue. It puts it into observation mode." Instead of blocking releases on ever-heavier review, one operator ships, then watches — borrowing Charity Majors' observability thinking and making it a first-class part of the AI workflow. Issues stay open after merge; an overnight review sweeps everything in observation mode and asks of each: roll back, raise a yellow flag, or close clean?

The goal is a specific ratio: "I want it to be able to make 50 decisions and only have to hold maybe two of them for me." And a specific discipline about the human's role — "I don't want to say LGTM anymore. If I'm going to say LGTM, it should just ship. It should not be blocking on me." If your review keeps producing rubber-stamp approvals, the approvals encode a framework you haven't written down yet. Write it down, automate it, and reserve yourself for the two decisions that genuinely need you.

What makes the fast shipping safe is knowing your quality bar. Two steps forward and one step back is okay to ship — for some products. A regression is acceptable as long as you're monitoring its impact in production and can respond immediately. Other products — medical, financial — have defect thresholds, with nothing shipped above the line. The bar is a choice the CEO makes, not a default the team guesses at.

Teach the System, Not the Session

When AI work needs correcting, the highest-leverage response is to make the fix teach. My loop: make the change myself, commit it, then tell the AI — I made some changes to your work. Look at what I changed; go update the skills; capture the context; see what you can learn from this. One diff becomes a permanent upgrade instead of a one-time correction.

An operator in the room framed the same instinct as a management question: "If you were working with a junior dev, would you just reach in and do the fix? Or would you be trying to teach them to do the fix?" Their version of the discipline is spending meeting time with the AI "talking about process, and not talking about the work" — seventy thousand dictated words in a week, most of it aimed at norms and frameworks rather than tasks. That's the capture problem, solved at the personal level: every correction lands in a file that the next session inherits.

For evaluating the work itself, I run a rubric-based review: have the main agent develop scoring criteria, spawn multiple agents to score from different perspectives, have the main agent synthesize, and make the call myself. Two details make it work. "The comments are more important than the scores," as one attendee put it. And the loop is honest — sometimes the system applies its own recommendations, rescores, and the score goes down. That's the review working, not failing.

Make the AI Do Its Homework

A film producer described using AI as an inside-baseball translator: in an industry where "sometimes people in the industry purport to know things that they don't actually know," the AI helped separate how things actually work from confident folklore. Their calibration was right, too — where they knew the domain cold, they could see the AI wasn't 100% correct, which is exactly the Gell-Mann amnesia warning: we notice AI's errors in our own field, then trust it everywhere else.

The fixes are mechanical. Demand quotes from original sources instead of assertions — "make sure it does homework, or else it's just participating in the gaslighting." Don't disclose your own position when you ask, because it will happily lean your way: "the model itself is sycophantic, usually to you, unless you tell it to be adversarial." The strongest version of adversarial runs a second model: from a coding agent, it's a command-line call to have a different model's agent review the work — models argue with each other more honestly than either argues with you. And these habits are capturable too — the demand-sources prompt, the adversarial-review setup — written down once, they join the system, and the next person inherits them.

Where to Start

• Inventory the uncaptured. List every AI workflow your team relies on and ask of each: if its author left Friday, could someone else run it Monday? Capture the ones that fail the test first.
• Add the escape valve to your policy. A fast, logged path to any tool an employee wants to try. If your policy is all control and no capture, revise it — the generator gives you a draft to react to.
• Reward artifacts, not activity. Incentivize the written-down skill and the reusable tool, not the token count.
• Convert repeated AI work into code. Anything the AI does the same way twice is a candidate for a deterministic tool that runs free.
• Ship with observation, not just approval. Keep issues open after merge, monitor what shipped, and reserve human review for the decisions that genuinely need judgment.
• Demand sources. Before acting on an AI claim outside your expertise, make it quote the original documents — and don't disclose your own position when you ask.

Resources From the Roundtable

• FluidVoice — local dictation on macOS running NVIDIA's speech model; faster than cloud APIs with no network round trip.
• Wispr Flow — high-volume dictation; the input side of the process-not-work pattern.
• Whisper — OpenAI's speech-recognition model; the quantized Whisper 3 Turbo build discussed ships in ~500 MB and runs locally.
• Claude Code — the coding-agent context where the multi-model adversarial review runs as a simple command-line call.
• Descript — transcript-based video editing; the session's case study in token economics meeting product pricing.
• Charity Majors on observability — the thinking behind treating post-merge observation as a first-class stage.
• SpaceX's acquisition of Cursor — the $60B deal cited as the model-companies-absorb-model-dependent-tools endgame.