1. Is confidence reported per field or per document?
A document-level score is close to useless: if a bill scores 94%, you still do not know which field is the shaky one, so you check all of them and the automation saved you less than it looked like.
Ask to see a document where the tool was unsure about exactly one value, and watch where that shows up.
2. Can you get from a value back to the page it came from?
Not 'can you open the source document' — every tool can do that. Can you click a total and have the exact region it was read from highlighted?
Without it, verifying a field means finding it yourself, which is the work you were automating.
3. Where is the human gate, and can it be turned off?
Some tools auto-post above a confidence threshold. That demos brilliantly and it is the setting that produces the entry nobody remembers approving.
Ask whether the gate is a default or a guarantee.
4. Can the audit trail be edited?
If entries can be changed or deleted, the log's value as evidence is gone — including for the entries nobody touched, because a reader cannot tell which those are.
Ask what happens to the old value when someone corrects a mistake.
5. How are clients kept apart?
For a firm this is the one that matters. A filter over a shared pool and a genuine per-client boundary look identical in a demo and behave very differently when a staff member changes clients or leaves.
Ask what a user sees if the filter fails.