We already tried that: keeping lessons where AI can read them

Picture a new operations lead suggesting a discount for customers who pay by mobile money instead of bank transfer. It sounds sensible. Faster cash, lower fees.

Someone who has been there four years says, "We tried that." The discount got applied to customers who were going to pay that way anyway, the bank transfer customers did not switch, and the finance team spent a month untangling which invoices had the discount. Nobody wrote it down. The only record of the experiment is in the heads of the two people who were around for it, and one of them has since left.

Every firm has a few of these. Experiments that failed for good reasons, where the reasons now exist only as a sentence someone says in a meeting, if they happen to be in that meeting.

AI makes the forgetting worse

Ask an AI assistant for ideas to improve cash collection and it will very likely suggest an early-payment or mobile-money discount. It has no way of knowing you tried it. It will make a good case for it, too, because it is a good idea in general. It just did not work for you, for reasons specific to your customers.

This is one of the clearer findings in MIT NANDA's 2025 report on why so many enterprise AI projects stall. The researchers described the tools as ones that do not retain feedback, adapt to context or improve over time, and called it a learning gap. A tool that forgets every correction will keep making the same suggestion, and someone will keep having to explain why it does not fit.

Personal memory features in AI assistants help a little here. If you told your assistant last month that the discount failed, it might remember. Your colleague's assistant will not, and neither will the new operations lead's.

A learnings log

The fix is old-fashioned. Keep a written log of what the company has learned, in one place, and make it something both people and AI tools read before proposing changes.

Each entry needs four things:

• What was tried, or what happened.
• What the result was, with numbers if you have them.
• Why you think it turned out that way.
• What you would do differently, or what to check before trying it again.

The last line is the most useful one. "Do not try this again" is rarely the right lesson. "If you try this again, apply the discount only to customers who currently pay by transfer" is.

How we keep ours

At Provecta, every AI working session ends with a closing routine. It writes down what was done, what was decided and what was learned, in a dated decisions-and-learnings file. The next session reads the recent ones before it starts work.

Most entries are small and practical. One from 30 June is a good example of the kind of thing that would otherwise get lost: a destination for publishing work being set up and ready is not the same as permission to publish to it. That came out of a moment where the setup implied an obvious next step, and the right move was to stop and check with the owner instead. Written down, it applies to every session after, without anyone having to remember the original situation.

We do not claim this makes our AI wise. It makes it less likely to repeat a specific, recorded mistake. That is a narrower promise and an easier one to keep.

The part we have not solved

Lessons go out of date too. A discount that failed in 2024 might work in 2026, because your customers changed, or mobile money fees dropped, or you now have software that applies discounts properly. A log full of "we tried that" can make a team, and its AI, too cautious.

There is no clean rule for when a lesson should expire. Dating every entry helps. Reviewing old entries every few months helps more. But there is a real tension between remembering what failed and being willing to try it again under new conditions, and no format resolves that on its own.

Start one this week

Ask the three longest-serving people in your company to each name two things the firm tried that did not work. Write them up in the four-line format above, with whatever detail they remember. Date them.

Put the six entries at the top of a shared document called "What we have learned" and point your AI tools at it. Then ask your assistant for ideas on the problem one of those experiments was meant to solve, and see whether it still suggests the thing you already tried.

Book 20 minutes at pgco.world/book.