I recommend reading an AI draft with one question in mind: what will someone do because of this? That tells you where to put your attention. A convincing paragraph can still contain the wrong price or a commitment you never made.
Look for the facts people will act on.
In a customer email, check names, dates, prices, and promises. In a report, check the figures and the period they describe. In a meeting brief, check which decisions are settled and which are still open.
Go back to the original material for those details. Asking the same tool whether its answer is correct does not give you independent confirmation.
- What are we promising?
- Which number or date changes the decision?
- Who is expected to do something next?
Open the source.
A reference helps you find the evidence. You still need to read it. Check whether it supports the actual statement, including any qualifications.
For an illustrative example, a report might say that three customers raised an issue. A summary that calls it ‘a widespread customer concern’ has added an interpretation. You need to decide whether the evidence supports that wording.
Show me the source passage for each claim in this paragraph. Identify statements that go beyond what the source says. Keep observations, assumptions, and recommendations separate.
Tell it what it could not know.
A draft can be consistent with the documents and still be out of date. Perhaps you changed the deadline on a call this morning. Perhaps the recipient already knows the background and just needs the decision.
Add that context explicitly. Ask for the specific change you need. ‘Remove the background and ask Anna to confirm Thursday’ is a much clearer instruction than ‘make it better’. The revised draft still needs your review.
Get the right person to check it.
If you cannot assess a claim yourself, involve someone who can before using it. The person responsible for the customer relationship may need to review a promise. The person who owns the figures may need to confirm the comparison.
Save corrections that keep recurring. They tell you what your instructions or source material need to explain. If every answer needs extensive repair, look at the setup before making the process automatic.
Open an AI draft you’re about to use. Identify the three details most likely to affect what someone does next, then check them against the source.
Want me to work through this with you?
Book a call with me