Narrowly, and on your worst data. Pick three real batches, including one with messy supplier documentation and one that failed something, and ask the vendor to load them and produce one finished output: a deviation write-up, a checked certificate transcription, an audit pack. Timebox it to two weeks.

A pilot on clean demo data tells you what the software does when nothing's wrong, which is never the state of a real quality record. What you're buying is the behaviour in the awkward cases.

Four, reliably. Show me the audit trail for something the AI generated last week. Show me what happens when two supplier certificates disagree about the same lot. Show me a failed result, then show me everything the system did with it. And export my data now, in front of me.

None of those are unfair, and all four are answerable in minutes by a system that works the way it's described. Hesitation on any of them is information.

Three vendors, not seven. Beyond three, evaluations stop being comparisons and become scheduling exercises, and the decision drifts to whoever presented most recently.

Head-to-head detail on the platforms most often shortlisted alongside us sits on the comparison hub, and the architectural question underneath the scoring is covered in AI-native versus AI bolted on.