How to check AI-generated work before you use it

A finished-looking answer still needs a check

An AI-generated brief can have clear headings, tidy numbers and a convincing conclusion. None of those tells you whether the conclusion follows from the evidence.

The useful question is more specific than “Does this look right?” Ask: “What would I need to check before using this to make a decision?”

For a comparison, that might be a price condition. For a research summary, it might be whether the quoted source supports the claim. For a product brief, it might be whether several interview comments really justify the proposed priority.

Microsoft's guidance distinguishes checking an output from improving its presentation. It also states that Copilot can help examine its own work without certifying its correctness. That is a useful boundary for any AI-assisted review. Read Microsoft's output-validation guidance. ↗

At Altiv, we recommend beginning with the parts that could change what you do next. The aim is to produce a usable decision and a short record of the evidence behind it.

A real check from Altiv's own research

While researching our content plan, we used AI to organise keyword data copied from Google Keyword Planner. The paste had lost its column boundaries. One row looked like this:

Under the expected column order, that could be recovered as 170 average monthly searches, 0% three-month change, a 65% year-on-year decline, and Medium advertiser competition. But a plausible interpretation was not enough to establish the reporting settings or column meanings.

We also had a separate forecast file set to India. It would have been wrong to assume the pasted results used that location just because the files arrived together.

Varun confirmed that the pasted results covered the United States, all languages, October 2025 through September 2026, with bids in INR. He also supplied a complete row for “AI adoption plan”: 10 monthly searches, 0% changes over both periods, High advertiser competition, and low/high top-of-page bids of INR 899.11 and INR 2,137.39. The copied row had lost its spacing and column separators, so human verification was essential.

That confirmation supported the parsing and the market label. It did not turn every part of the table into verified information. Two damaged percentage cells remained unresolved, and entries shown as dashes remained unavailable rather than becoming zero.

Here is the distinction that mattered:

Google defines Keyword Planner competition in terms of advertisers. A table can copy the value correctly and still use it to answer the wrong question. See Google's historical-metric definitions. ↗

This exercise produced a more usable research table. It did not establish that every recovered cell had been independently checked in Google's interface, or that a chosen keyword would deliver traffic.

Check the evidence behind the important claim

Choose the statement that carries the recommendation. Open its source and find the passage, record or input that supports it.

A link that opens is only the beginning. Check whether the source concerns the right people, task and period, and whether it establishes the claim being made. A study of one tool in one setting cannot automatically support a claim about every professional using AI.

The same applies to your own material. If a product brief says customers want a feature, return to the interview notes. Look for who asked, what problem they described and whether other evidence points elsewhere. This is Altiv's recommended application of the method, not a claim about a measured product team.

If the evidence supports a narrower statement, narrow the statement. If you cannot find the evidence, mark it unconfirmed and decide whether the work can proceed without it.

Check the number and what it means

A number can be copied accurately but interpreted incorrectly. Verify units, dates, filters and the population it describes before using it in a comparison.

For a calculated result, recompute the consequential calculation from the original inputs. For a converted or extracted table, compare source rows, including awkward cases such as blanks, ranges and negative changes. A successful check of one row supports that row and the tested parsing rule; it does not certify the whole dataset.

In our example, the market was US while the currency was INR. Changing the currency label to USD would have made the output look familiar to a US reader and made it wrong.

Decide the depth of checking from the intended use. A rough shortlist and a customer-facing financial recommendation need different levels of verification.

Look for the missing condition

Ask what information could reverse the recommendation.

• Does a supplier price depend on a minimum commitment?
• Does a claimed improvement exclude the time spent reviewing and correcting the output?
• Does a customer theme come mainly from one segment?
• Is a missing metric being treated as evidence that nothing happened?

You do not need to resolve every unknown before doing any work. You do need to preserve the unknowns that matter to the decision. In the keyword table, we could compare available estimates while leaving damaged percentage changes unresolved. Those percentages were not needed to support every comparison.

Enlarge image ↗Open original ↗

Close image ×

Use AI to help investigate, not just to reassure you

AI can help identify statements that need evidence, compare a draft with supplied material, or propose ways a recommendation could fail. Give it the actual material and a bounded checking task.

Compare this output with the source material below. Identify claims that are unsupported, numbers that do not match, and missing conditions that could change the recommendation. For each issue, point to the relevant source passage or record and suggest the next check. If the source does not settle a point, say so.

Output to review:
[Paste output]

Source material:
[Paste material you are permitted to share]

Treat the resulting review as a list to inspect. Open the cited passages, verify calculations and decide which findings matter. Agreement between two AI responses is not the same as confirmation from the underlying evidence.

Decide what is ready to use

Finish with a practical disposition:

These are Altiv's working categories, not a certification system. One output can contain usable sections and a recommendation that still needs to be held.

The useful record is short: what you checked, what you found and what happens next. Keep it with the version someone will actually use.