Three jobs that need a different approach
A fluent answer can conceal the exact error that matters. I change the approach when the task depends on exact calculations, claims about real people, or judgment the reader cannot independently check.
What you will leave with
A decision to verify, use a different tool, or involve a qualified reviewer.
Practice time: about 15 minutes1. Calculations and business records
AI tools can use code, calculators, and spreadsheets. That makes "AI cannot do arithmetic" the wrong rule. The useful question is whether the inputs, calculation, and result are visible and can be checked independently.
Use a spreadsheet, calculator, or validated system for the calculation. AI can help draft a formula or explain a result, but verify the range, units, assumptions, and treatment of blanks and duplicates. A correct calculation on the wrong rows is still the wrong answer.
Worked example / Fictional practice material
A list contains 12, 8, and 5 units. The total is 25. If the last row was imported twice, a formula can accurately total the duplicated data and give 30. Checking arithmetic alone misses the source error.
Better request: Identify the input rows, show the calculation, and list possible duplicates or missing values. Do not change the source records.
2. Claims about real people
Do not turn generated biography, reputation, or inferred intent into a fact about a customer, applicant, colleague, or vendor. Require an appropriate source and check identity, date, and context. A shared name or an old page can point to the wrong person.
For sorting, use explicit work-related categories you can inspect. "Billing question" is a category in a message. "Likely unreliable customer" is an unsupported judgment about a person. The latter needs a different process, not a stronger prompt.
If a tool cannot locate supporting evidence, the useful output is "not established." A second AI agreeing with the first is not independent verification.
3. Work you cannot competently review
AI can help organize questions for a qualified professional or summarize an approved source. It should not settle a consequential legal, tax, medical, or personnel decision merely because the explanation sounds convincing.
Ask who could recognize a wrong answer and what evidence they would need. If no one in the process can check it, pause that use. You can still use AI to prepare a neutral list of open questions using information approved for the tool.
Make the checking route part of the request
- Identify the claim: Which part would matter if it were wrong?
- Find independent evidence: Original records, a reproducible calculation, current authoritative guidance, or a qualified reviewer.
- Check the match: Does the evidence support this exact claim for this person, date, and situation?
- Record the decision: Verified, corrected, unresolved, or outside the tool's role.
The NIST Generative AI Profile describes confidently false output as a risk of generative AI. My practical response is to design the check before relying on the answer. Asking the model to be careful helps express your intent; it does not supply the missing evidence.
Keep this
A verification decision sheet
Copy this into your own document and fill in the brackets. Use approved or fictional material when trying it with AI.
OUTPUT: [What does the tool propose?]
CONSEQUENTIAL CLAIMS: [Numbers, identities, commitments, or judgments.]
EVIDENCE NEEDED: [Original source / independent calculation / expert review.]
SOURCE MATCH: [Correct person, period, units, version, and context?]
REVIEWER: [Person able to recognize the relevant error.]
RESULT: [Verified / corrected / unresolved / outside the tool's role.]
STOP CONDITION: [What prevents this output from being used?]
NEXT ACTION: [Who resolves the gap, and by when?]Want help applying this to your organization? Bring the job you have in mind, and we can work out a useful next step.