Ask the model to argue both sides before you finalise an investigation summary

If you're using an AI tool to help structure notes from a workplace investigation - redacted and placeholder-named, per the usual rule - don't ask it to summarise straight into a conclusion. Ask it first to lay out the strongest reasonable version of each account, including the version you're inclined to believe less, before it drafts anything that looks like a finding.

The reason this catches something a straight summary won't: a summary written from notes that lean one way will read as balanced even when it isn't, because the model faithfully reflects whatever weighting is already in the input. Asking it to separately construct the other side's strongest case forces a genuine comparison onto the page, rather than one account dressed up as an objective account of both.

A useful follow-up once you've got both sides laid out: ask what evidence would be needed to resolve the specific points where the two accounts actually conflict, as opposed to the points where they agree and only the framing differs. That list is often shorter than the investigation felt like it needed, and it's a better use of the next interview than re-covering ground both sides already agree on.

This doesn't replace a competent, trained investigator's judgement - it's a way of pressure-testing a draft before it goes anywhere formal, not a substitute for doing the interviews properly in the first place.