Your files hold what you wrote, not what you decided

Say a firm tried a referral scheme two years ago. The proposal was written up properly: a six-page document with projected numbers, a launch plan and an enthusiastic summary. The scheme ran for four months, cost more than it brought in, and was quietly dropped after a short conversation in the owner's office. Nobody wrote that conversation down.

This year someone asks the firm's AI assistant, now connected to the shared drive, for ideas to grow referrals. It finds the six-page proposal, which is well written and on topic, and recommends it. With citations.

The assistant did exactly what it was built to do. It found the best document on the subject. The problem is that the most important fact about that document, that the firm tried it and it failed, was never a document.

Most of what a firm knows was never written

Documents record intentions: plans, proposals, policies, templates. They rarely record what happened next. Decisions get made in meetings, in calls, in a sentence at the end of a WhatsApp thread. Lessons get learned and stay with the people who learned them. Exceptions get agreed for one client and never written into the standard terms.

So when an AI searches your files, it searches a version of the firm that is tilted towards optimism and out of date. Plans look like current practice. Abandoned ideas look like live options. The policy from 2023 looks like it still applies, because nothing says it does not.

Atlassian's 2025 survey of 3,500 developers and managers found that finding information was the single biggest drain on their time. Search tools help people find documents faster. They do not fix the fact that the document is the wrong answer.

What a decision record looks like

The fix is unglamorous. Decisions, and the things that did not work, get written down as short entries of their own, separate from the documents they relate to.

At Provecta Group, the company memory our AI works from is built mostly of entries like this. Each one states a rule or a decision, then why, usually with the event that caused it and a date, then how to apply it. A few of the shapes we use:

• A decision. "We price this service by [unit], not by [other unit]. Decided on [date], because [what went wrong the other way]."
• A failure. "We tried this approach. It did not work, for this reason. Do not propose it again without new evidence."
• An exception. "Our default is to act without asking. Exception: when a client brief is ambiguous about the company or the direction of trade, ask one question first. This overrides the default."
• A fact with a gate. Last week we started a single list of the proof points we are allowed to use in public, each with its exact wording and where the evidence sits. The rule at the top reads: if a fact is not here, it is not proof yet.

The first two are templates, with the specifics left out. The format is the one we use.

Every AI session loads these entries before it starts work. It can still search documents for detail, but it reads the decisions first, so the abandoned referral scheme arrives already marked as abandoned.

Why this is the part AI is missing

MIT's 2025 report on business AI is mostly quoted for its failure rate. The more useful part was its explanation: many tools fall short because they do not retain feedback, adapt to context or improve over time. A search index gets bigger. It does not get wiser. Nothing it holds says "we tried this and it hurt".

A decision record is how a firm's feedback gets kept.

What it costs and what it misses

This is ongoing work, and someone has to own it. A decision only gets recorded if someone stops after the meeting and writes three lines. Most firms will not do that every time, and that is fine. The decisions that cost money when forgotten are the ones worth catching.

It also cannot record what nobody noticed. If a scheme failed and nobody connected the failure to the scheme, there is nothing to write. And a decision record can go stale like anything else. A failure from three years ago may not be a failure today. Each entry carries its date so a reader can judge.

Write three this week

Think back over the last year. Write down three things your firm decided, or tried and dropped, that are not in any document. One line for what, one for why, one for the date. Put them where your team, and any AI tool you use, will read them before they search anything else.

The operating system health check looks at where your firm's decisions and knowledge actually live. Book 20 minutes at pgco.world/book to go through it.