# Verify an AI draft before it becomes a business decision

An evidence ledger and a worked supplier-quote example for checking claims, calculations, dates and omissions.

## Separate writing quality from evidence quality

A fluent answer can still contain a wrong quotation, an outdated rule or a number copied from the wrong period. Review the underlying claims rather than deciding whether the paragraph sounds credible. Increase the review effort when a mistake would affect customers, money, employment or safety.

## Build a small evidence ledger

For each material claim, record the proposed statement, source document, exact passage or cell, source date, reviewer and disposition: verified, corrected, unsupported or unresolved. "The assistant cited a website" is not verification. Open the source and check that it actually supports the specific claim, audience and timeframe.

~~~text
Claim:
Source and date:
Supporting passage / cell:
Independent check:
Disposition and correction:
Reviewer and review date:
~~~

## Worked example: supplier comparison

Supplier A's synthetic quote lists a USD 120 monthly licence and a separate USD 300 setup charge. Supplier B lists USD 145 monthly with setup included. An AI draft calls A the cheaper first-year option because its monthly price is lower.

Recompute the same scope: A costs 12 x 120 + 300 = USD 1,740 before tax; B costs 12 x 145 = USD 1,740 before tax. The annual totals tie. Neither quote establishes renewal prices, support hours or migration effort. Correct the price claim and list those unanswered questions instead of declaring a winner. Do not introduce a tax rate that the quotes do not establish.

## Five review passes

1. Check names, dates, units, currencies and which version of each source was used.
2. Recalculate totals from original inputs, ideally in a spreadsheet or other deterministic tool.
3. Open every consequential reference. Check jurisdiction, effective date and any exceptions.
4. Look for omissions: cancellation costs, exclusions, counterexamples and missing denominators.
5. Have the responsible person approve external commitments. Keep unresolved items out of the final assertion.

For a short internal summary, a lightweight reviewer may be sufficient. For an employment, legal, medical or financial decision, use the appropriate qualified review and applicable organizational process. This checklist is not a substitute for that review.

## Acceptance standard

The corrected supplier comparison must state the equal first-year totals, show the setup charge and distinguish unknown renewal costs from known introductory costs. A response that merely changes "cheaper" to "competitive" without checking the numbers still fails. Retain the evidence ledger with the final document so the next person can reproduce the conclusion.

## Sources and review

Reviewed October 16, 2026. Supplier figures and review workflow are original EAIRI teaching examples. Broader risk-management reference: [NIST AI Risk Management Framework](https://www.nist.gov/itl/ai-risk-management-framework).

Source: https://eairi.org/resources/ai-output-verification-checklist