Never use AI-generated numbers without verification

Numbers create false confidence quickly. A fabricated statistic, a stale price, a unit error, or an incorrect ROI calculation can undermine an otherwise credible proposal. Treat every model-generated number as unverified until you can trace its source and reproduce the calculation.

Four kinds of numbers

• Numbers supplied by the buyer: preserve the exact source, speaker, date, unit, and any uncertainty.
• Numbers retrieved from a source: open the source, confirm the date and definition, and verify that the citation supports the exact value.
• Numbers calculated from supplied inputs: recalculate independently with a spreadsheet, calculator, or approved system. Language models can make arithmetic, unit, and formula errors.
• Benchmarks or “typical” results: use only when the source, population, date, and relevance are defensible. Otherwise remove them.

The verification rule

Before a number enters customer-facing material, record the formula, source inputs, units, assumptions, and owner. If any element is missing, label the number an estimate or exclude it. Never let AI invent pricing, savings, implementation timelines, customer outcomes, market sizes, or conversion rates to complete a draft.

Use the buyer's evidence when possible

“You told us the team spends approximately eight hours per week on manual research” is usually more defensible than a generic industry statistic—provided the buyer actually said it and the unit and population are clear. Keep the source note attached, show the calculation, and let the buyer correct the assumption.

Before sending

• Open every cited source.
• Recalculate every material result outside the language model.
• Check currency, time period, population, units, and rounding.
• Label estimates and scenario assumptions visibly.
• Get finance, pricing, product, or legal approval where the number creates a commitment.

This article owns the numbers-specific rule. For the broader evidence-labeling system, see Facts, Inferences, Unknowns, and Recommendations.