Verification and human review before work leaves your desk

Compare claims with the actual source

Review names, dates, amounts, units, responsibilities and status against the original material. Open cited sources and confirm they exist and support the claim. A plausible-looking link, quotation or citation can be invented or unrelated.

Ask the model to separate supported facts, interpretation and unknowns. Then inspect that separation yourself. Asking the same assistant whether its answer is correct is not independent verification.

• Is this a source fact or an inference?
• Is the source current, authoritative and within scope?
• Does a quoted passage match the original?
• Did an open question become a settled fact?
• Was a proposal described as an approved decision?

Verify numbers outside the prose

Check calculations with a trusted calculator, spreadsheet or approved system. Watch for mixed units, missing rows, duplicate entries, date assumptions and rounding. Even when an AI uses a calculation tool, its inputs or interpretation can be wrong.

For a summary of costs, confirm both the values and what they include. ‘Total’ may omit taxes, exclusions or a line the model failed to extract. Do not commit funds from an unchecked summary.

Match the reviewer to the risk

A routine draft may need your own factual and tone review. A legal response, financial commitment, personnel matter, security decision or safety issue needs the appropriate authorized reviewer. AI does not create approval authority.

Review recipient, attachments, confidentiality, wording and approval status before sending a message. In a connected workflow, tool access is not permission to act without the required review.

Keep the approved result and the unresolved items

Save the final approved result in the proper system of record. Keep enough source context to explain the decision under your organization's retention policy. Do not let an AI-generated summary silently replace the original record.

If an important mistake was found, check related outputs produced from the same faulty assumption. Correct the source or instructions, not just the one sentence. Keep unresolved issues visible until a person resolves them.