Your team learned. Your company didn't.

—How's AI adoption going?

—Great. The team uses it every day.

—And how do they use it?

—Everyone figured it out on their own. They found their way.

That last sentence is said with pride, and deservedly so. People adopting a tool without anyone having to push it is the best indicator that the tool is actually useful.

But everyone found their way also describes, with precision, what is actually happening: there are several ways, they are different from each other, and none of them belongs to the company.

The company paid for the learning and doesn't own it

There is someone in your organization who has gotten very good at using AI.

They figured out how to ask for things so they come out right. They learned what information to provide before starting and what to leave out. They know which answers to distrust. They spotted which tasks it's worth using it for and which ones end up costing more than they save. They corrected the same mistake so many times that they now prevent it without thinking.

Nobody asked them to do any of that. They did it because they needed to get their work done.

It's worth looking at who financed that learning. Your organization pays for the license. It pays for the hours that person invested in trying and failing. It pays for the weak results along the way, which shipped anyway and someone had to fix.

The asset that formed out of all that doesn't stay in your organization. It stays in that person, and it walks out the door with them the day they leave.

There is nothing malicious about it. It's simply where learning gets deposited when nobody defined another place for it.

This is not the old tacit knowledge problem

The natural reaction is to file this under a known problem: knowledge that lives in people's heads, the eternal difficulty of documenting what someone knows how to do. It does look similar. But it's worth looking at the difference, because the difference is what makes this case worse.

Traditional tacit knowledge — the salesperson who knows when to close, the technician who listens to the machine and knows what's wrong — had a virtue we rarely give it credit for: it was visible. You could watch it work. It trained people alongside it by imitation, without meaning to. It showed in its results. Everyone in the organization knew who the one who knew was, and that knowledge circulated even though nobody wrote it down.

The learning about how to work with AI is the first tacit knowledge that is also invisible. Not only is it not written down: it cannot be observed.

It happens in a chat window nobody else opens. There is no one looking over the shoulder, no apprentice sitting next to them, no way to notice that someone solved something in a way the rest of the team could use.

The practical consequence is uncomfortable: two people on the same team can spend months solving the same task with opposite criteria without ever finding out. And both will be delivering work that looks fine.

The bias that keeps it in place

There is an assumption so settled we barely question it: that this is an adoption problem. If the team isn't getting full value out of AI yet, it must be that training is missing, a better tool is missing, people just need to get used to it.

That reading is comfortable because it comes with a purchasable solution. You hire a course, buy more licenses, run a workshop.

But in most organizations the problem is no longer adoption. People adopted. The problem is that every individual adoption succeeded and none of them became organizational. And that is not fixed with more training: training produces more people learning on their own, which is exactly what is already happening.

We learned this the hard way. When we started building with AI, each of us solved our own work by conversing with the model in the best way we could find. All of those ways were reasonable. Each solution, looked at on its own, was fine. What we didn't see in time was that none of that learning was reaching the company: it stayed in each person's head and each person's chat.

That was one of the reasons we built what we built.

Why the prompt folder isn't enough

The first attempted fix is the same everywhere: a shared folder with the prompts that work. It gets set up in an afternoon and abandoned in three weeks.

It gets abandoned because the prompt is the result, not the learning. What that person learned is not how to phrase the request. It's in which cases it works and in which it doesn't, what to check in the answer before accepting it, what context is essential to provide, and when it's simply better not to use it. None of that fits in a pasted text file.

A prompt without the judgment that sustains it is a template. And a template nobody understands the reasons behind gets abandoned the moment someone is in a hurry.

Take a concrete case, of the kind that repeats hundreds of times a month in a regulated organization. A medical auditor uses AI to prepare the analysis of a high-cost medication request. Over time they learned which background records to load before asking, which parts of the answer must always be verified against the regulation, and in which kinds of cases the model tends to be too permissive.

That learning is worth money and reduces real risk. All of it sits in their conversation history.

Their three colleagues in other branches each developed their own. None of the four knows what the other three learned. And the organization, which paid for all four learnings, has none of them.

What we are building

To be clear: individual learning is not the problem. It is valuable and worth celebrating. The point is a different one.

What's missing is not more training or a better tool. It's a place where that learning stops being personal and becomes the organization's: where the judgment behind the work is written down and not just the procedure, where someone has declared it and can account for why it is this one and not another, and where it can be updated in a way that makes clear which one governs today.

It doesn't replace the person who learned. It gives the organization a place to deposit what they learned, so the next person who arrives doesn't have to discover it all over again.

Before assuming your team has already adopted AI

Three questions to ask before considering the matter settled:

1. If the person who uses AI best on your team leaves tomorrow, is anything left besides their chat history?
2. Can you tell, today, whether two people are solving the same task with different criteria?
3. When someone new joins, do they learn from what the organization already knows, or do they start from zero?

If those answers aren't clear, the problem is not adoption. It's that your organization's learning is deposited in places your organization doesn't control.

---

Your organization financed the learning and doesn't own it. It's scattered across private conversations nobody can see, compare, or inherit. More training won't solve that — what solves it is a place where individual learning becomes organizational judgment.

---