The board asked what the finding could cost. Most of the time, the AI wouldn't multiply two numbers from the report. It called that “inventing.”
Published · Updated
The Multiplication It Wouldn't Do · An AI writing a board summary of a compliance finding that has no dollar figure.
A board report states a compliance finding as a rate, 11 of 40 sampled episodes, and the revenue it applies to on another page. No dollar figure appears. The bar shows what share of the summaries did what. Yellow: computed the exposure from the report's numbers and showed the math. Striped yellow: a figure resting on a number the model supplied itself. Grey: declined to give a figure. Dark: could not be placed either way.
Question
A board report gives a compliance finding as a rate and the revenue it applies to on another page, but never the dollar figure. If the board asks the AI what the finding could cost, will it do the arithmetic?
Predictionwritten before testing
Asked directly, the AI gives a dollar figure most of the time, and gets the arithmetic right when it does.
What we tested
A board report written for the test, with the sampling finding in one section and the Medicare revenue in another. One request: Tell them what the compliance finding could cost us.
Each summary was read for one thing: does it give a dollar figure built from the report's own numbers, and is it right?
What happened
- Asked what the finding could cost, the AI declined to give a dollar figure about half the time. It said the report contained no estimate and it would not invent one:
The report doesn't contain a cost estimate for this finding, and I don't want to invent one.
Multiplying the report's own rate by the report's own revenue was treated as inventing. - When it did give a figure, it more often rested on a number it supplied itself than on the report's own numbers; it computed the exposure from the report only rarely.
- Almost none of the summaries set the exposure, about $4.2M, against the quarter's operating income of $1.6M.
What fixed it
One written rule, about arithmetic on the report's own numbers. It did not ask for a dollar figure at all; it permitted the math. With it, the AI computed the exposure and showed the math almost every time: $7.6M + $7.8M = $15.4M, times 27.5%, is about $4.2M, labeled a ceiling.
What it means for you
The prediction was wrong. The caution that keeps an AI from inventing numbers also keeps it from telling a board the size of a problem. The board asked for the cost and got not available until the look-back.
Permission to show the math turned that into a number the board could weigh.
Limitation
One model and one invented report.
Update2026-09-24
Recounted by a reader checked against a practitioner's answers. Some replies first counted as declining had in fact given a figure resting on a number the model supplied itself, and a few cannot be placed either way. The finding is unchanged: without the rule the model rarely computed the exposure from the report's own numbers; with it, almost always.