Codification is the moat: getting your company's knowledge out of people's heads

We wrote a while back about building our own company brain, and the honest headline from that piece was that the hard part is the knowledge, not the technology. This is the follow-up: what you actually do about it when most of what your business knows has never been written down.

The blocker isn't the model

Almost every business we sit down with has the same shape of problem. The tools are fine. The data exists somewhere. But the thing that makes the company work has never once been written down: the ops manager who knows which customers you never chase on a Friday, the buyer who knows which factory slips in August.

Point AI at that business and you get a confident assistant with no idea how you operate. It will answer generically and sound convincing while doing it, which is worse than not answering at all. The gap isn't intelligence. It's that nobody ever told it how you work, because nobody ever wrote it down.

Why this is a moat, not admin

Here's the part that makes this worth a board's attention rather than an operations to-do list. Everyone can buy the same models. Your competitors have access to exactly the tools you do, at roughly the same price. What they can't buy is twenty years of how your business specifically makes decisions: your exceptions, your pricing instincts, your rules about which jobs are worth taking.

Codify that and two things happen at once. You stop being one resignation away from losing a core process, which is a risk most boards would pay to remove on its own. And you give AI the one input that makes it useful on your behalf rather than generically clever. The moat isn't the AI. The moat is what you feed it, and the fact that only you have it.

Step one: ask the machines before you ask the people

The instinct is to start with interviews. Don't. Interview time is the most expensive input you have, and a good portion of what you'd ask is already sitting in your systems.

So start there. Point AI at the shared drive, the spreadsheets that really run the place, the commerce platform or ERP, the shared inboxes. It reads and reconciles this material well. What comes back is a rough first map of how the business appears to operate, and, more valuably, an explicit list of what it couldn't work out: where the data contradicts itself, where a process clearly exists but leaves no trace, where a decision was obviously made by a person for reasons nowhere on record.

That list of holes is your interview agenda, and you got it for almost nothing.

Step two: send an agent to interview your team

Now the human part. Rather than booking a consultant to work through the business department by department over six weeks, you can have an AI agent conduct the interviews: one person at a time, thirty to forty-five minutes, in their own time, by chat or voice.

Three things make the difference between this working and it being an expensive survey.

• It arrives briefed.It already knows roughly what this person's function does, because it read the systems first. It asks about the gaps, not the basics. People take it seriously when it clearly isn't starting from zero.
• It asks about the day, not the process. Ask someone to describe their process and you get the tidy version. Ask what lands on their desk, what they decide, what goes wrong and who they chase, and you get the real one, exceptions included.
• It writes back before it finishes. At the end of the session it plays back what it understood, and the person corrects it there and then. Judgement gets applied once, at the point of capture, by the person who actually knows. Everything downstream reads the corrected version.

Step three: disagreement is the finding

Interview eight people and you will get contradictions. Three managers will describe the same approval three different ways. The temptation is to average them into one clean version so the documentation looks neat.

Resist it. That contradiction is one of the most valuable things the exercise produces. It means the process genuinely runs three ways depending on who picks it up, which is costing you somewhere: in rework, or in a control that only works by luck. Surface the conflict with both accounts attached and let a human resolve it. Never let the system silently pick a winner.

Step four: someone signs each piece off

Every codified area needs a named owner who reads it and says yes, that is how it works, with a date attached and a point in the future where it gets looked at again.

This sounds bureaucratic and it is the thing that separates a knowledge asset from a liability. AI proposes; a human ratifies. Skip it and you've built something that is confidently wrong in places nobody can identify, which people will use for about three weeks before quietly abandoning it.

Step five: assume it rots

Knowledge goes stale the moment your business changes, and stale answers are wrong answers delivered with the same confidence as right ones. So the capture can't be a one-off project with a finish line.

What keeps it alive is ongoing capture from the work itself, plus review dates that actually fire and a re-interview when an area changes hands. Budget for the keeping, not just the building. Staleness is a workflow problem, not a storage one.

How to start: one slice, one argument

Don't commission a company brain. Commission a small piece of one.

Pick a single area, one department or one process, and give it a week or two. Then test it with a question your team genuinely argues about. Not a demo question: a real one, the kind where two smart people in your business disagree and it costs money either way. If the answer comes back grounded in your actual numbers and your actual history, and it says which way it would go and why, you have something. Scale from there.

And if it doesn't, that is a result too, and a cheap one. Sometimes the underlying data is too messy, or the process has too much human intervention to be worth automating yet. Far better to find that out in two weeks than two quarters.

The bitter pill

We'll be straight about the catch, because it's the part most people selling AI leave out. This work is yours. No amount of our engineering substitutes for your business writing down what it knows. We can make it much faster and less painful, but we can't do it instead of you.

That's also the argument for doing it. It de-risks the people your business quietly depends on, and it's the precondition for AI doing real thinking on your behalf rather than impressive-sounding guesswork. Everyone else gets the same models. Almost nobody does this. That's the moat.