AI adoption in our company has moved quickly. Many of us now use an AI assistant not only to write, but to analyse. When we planned the next quarter, a colleague built our revenue forecast with one. It came with two scenarios: one based on the contracts we have, and one in which nobody cancels.

I asked a simple question. In one of our product lines, the notice period for the next renewal had already passed. Nobody could cancel anymore, so that line had to show the same number in both scenarios. It did not, because the second scenario had quietly kept a customer who had already cancelled.

The Model Did Not Lack Information

The cancellation was not hidden. It had been in our CRM and in our inboxes for weeks, and the first scenario already left the customer out. The second scenario overwrote a fact with an assumption: “nobody cancels” became “everybody stays”.

Looking back, I would not call this an AI mistake. The assistant answered the question it was given. Our own process has a similar blind spot: a contract is extended when no cancellation reaches the sales team, so silence counts as renewal.

Every Session Starts From Zero

The forecast itself was the smaller problem. Many of us now ask an AI assistant to collect and analyse data, and each session writes its own script. It decides again which sources count and defines again what an active contract is.

A correction made in one chat stays in that chat. The next forecast, built by someone else or by the same person a month later, starts from zero.

In The Accidental Fog-of-War Machine, I described how AI scales the standards and heuristics each person brings to their work. In When Building Gets Faster Than Learning, I argued that cheaper building moves the bottleneck to product decisions. Analysis follows the same pattern. Cheaper analysis moves the bottleneck to consistency.

For analysis, we looked at AI adoption through the lens of personal productivity. Each person got faster. The company did not get more consistent.

Researchers at MIT describe a similar problem at a much larger scale. In The GenAI Divide, Project NANDA’s report on generative AI in business, they name learning as the core barrier to scaling: “Most GenAI systems do not retain feedback, adapt to context, or improve over time.” In a small team, the effect is the same. If the learning stays in one chat, the organisation does not learn.

What Does AI Adoption Need Beyond Personal Productivity?

Beyond personal productivity, AI adoption needs shared skills: reusable packages of definitions, sources, scripts and checks that every colleague’s assistant uses for the same recurring question. Personal use makes each person faster. A shared skill makes the answers consistent, because a correction becomes a rule everyone uses instead of a fix that stays in one chat.

Anthropic, which introduced agent skills, describes them as “organized folders of instructions, scripts, and resources that agents can discover and load dynamically to perform better at specific tasks.” When the script lives inside the skill, nobody has to write it again, and every assistant calculates the same way.

Not everything belongs in a skill. Judgment is hard to write down, as I explored in Encoding Judgment Is Not a Writing Problem. But what went wrong in our forecast was not judgment. It was a definition, and definitions are the easy part.

For our forecast, a shared skill would hold four things:

• Definitions: A cancelled contract is out in every scenario. A passed notice period is fixed. Scenarios only vary what is still open.
• Sources: Before calculating, the assistant reads current communication in the CRM, the shared inbox and our chat. It marks each contract as decided, at risk or unremarkable, with a link to the source. Silence becomes a documented search: “checked, no cancellation found”.
• Checks: Numbers that must not move between scenarios are tested before anyone sees the table.
• Owner: One person keeps the skill current and decides when a definition changes.

Our engineers already work this way. Their AI pipelines live in one shared place, and their fixes come with tests that pin the values that must not move. Our analyses are not there yet.

Count the Second User

A shared skill only helps if someone other than its author uses it. When we reviewed the skills one of our teams had built, only a minority had ever been tried by a second person. That is the number I now pay attention to: not how many skills we build, but how many are used by at least two people for a real decision.

I am still learning where to draw the line. Many analyses are one-offs, and they should stay one-offs. My practical takeaway: if a question feeds a decision more than once a quarter, its definitions belong in one shared place, not in somebody’s chat history.

When analysis gets cheap, consistency gets expensive. The answer is not a smarter model, but fewer places where the same question is answered from scratch.