How do I stop AI from inventing numbers in my business plan?

A revenue plan I built with AI came back with a blank where the total should have been. That blank was the best line in the document.

Before I ran the prompt, I labeled every fact I handed the model. CONFIRMED, with the source named. DRAFT, meaning it came from an internal note nobody had checked. OPEN, meaning the client had never answered. Then I added one sentence: "Do not produce a revenue total from any number I marked OPEN. Write the formula and name the missing variable instead."

The model obeyed. Instead of a confident twelve-month projection built on a list size that does not exist, I got a formula with one variable circled and a question at the top asking for it.

Why does AI make up numbers?

An AI will fill any blank you leave, and it fills it with something plausible. Plausible is the dangerous part. A made-up number looks identical to a real one on the page, and a founder will build a quarter on it. Labeling the inputs moves the model from guessing to asking.

Timing makes this more pressing. The models got good enough that their mistakes stopped looking like mistakes. Two years ago a bad output was obviously bad. Today it is a clean table with a wrong number in row four. The fix sits upstream, in how you hand over the facts.

How the setup works

Anyone can run this on a planning prompt they already use. Split the background into what you know and can source, what you believe but have not checked, and what you do not know at all. Add the math guard sentence above.

Then cap the model's questions at six, ranked by which one blocks the most work. Without the cap, a careful model can stall you with twenty.

If the prompt has a second step (a 90-day plan, a calendar), paste the first output in whole. The diagnosis becomes the source of truth for everything after it, so the model cannot re-diagnose halfway through and hand you a plan that contradicts its own audit.

What happens when two facts disagree?

The labels surfaced one more behavior I did not plan for. Two of my CONFIRMED facts contradicted each other: two documents, two prices for the same product. The model listed both versions with both sources and stopped. It did not pick.

That is the behavior you want from a planner. A prompt that says "assume reasonable numbers" gives the model permission to choose whichever price reads better, and you never learn the conflict existed. A planner that stops and shows you both prices hands you a decision you now get to make on purpose.

Do this today

1. Open the last plan an AI wrote for you.
2. Highlight every number in it.
3. Next to each one, write CONFIRMED and the source, DRAFT, or OPEN.
4. Count the OPENs.
5. Send every recommendation that rests on an OPEN back to the pile until you answer it.
6. Before your next planning prompt, add the math guard sentence and the six-question cap.