Teaching it once instead of saying it every time
The pitch for most AI estimators is "describe the job, get a price." Ours is closer to "correct it a few times, and it starts pricing the way you do." The difference only matters if the correction actually sticks — if it has to be said again on the next job, nothing was learned, only repeated.
When a correction generalizes — "always add 15% on occupied units" rather than a fact specific to one bathroom — it is saved as a rule and injected into every future estimate's prompt for that company. The job-specific fact stays with that one job; the generalizable one becomes standing knowledge.
A rule, not a sentence in a box
This used to be one long freeform text field, shared across every job. It worked, and it also had real problems: nothing in it could be retired individually, it carried no record of where it came from or when, it could not be scoped to a specific section of an estimate, and a contradiction between two lines in the box was invisible to the system.
Each learned rule is its own record now — text, when it was added, which estimate it came from, and the excerpt of the conversation it was learned from. A rule can be revised with its prior wording kept, retired without deleting its history, and restored if retiring it turns out to be wrong.
What happens when two rules disagree
The old box had no way to notice that 'occupied units +15%' and 'add 10% for occupied units' were both live at once. That is exactly the failure mode a growing rule set will eventually hit if nothing checks for it — corrections pile up faster than anyone reviews them by hand.
New rules are checked against the existing set for that kind of overlap before they are trusted alongside it, so a contradiction shows up as something to resolve rather than two rules quietly fighting over the same line on every future bid.
Whether it actually pays off, honestly
This is the part worth being direct about: rules accumulating is not automatically the same thing as estimates improving. It is possible to end up with a well-organized pile of corrections that still does not converge into a better price, and we track rule usage — which rules keep getting applied, which have never been used once — specifically because that is answerable, and worth answering before claiming more than the mechanism itself.
What we can say is narrower and more useful than a general promise: a correction here has a place to live, a way to be scoped, a way to be checked against what was said before, and a way to be taken back. Whether that adds up to consistently better pricing over a contractor's history is a measurement, not a slogan, and it's one we're watching rather than declaring settled.