Every AI vendor quotes an accuracy figure and almost none of them explain what it measures. The number is usually true and frequently meaningless, because accuracy is only defined relative to a dataset, a task and a definition of correct. Four questions convert the claim into something you can act on, and any vendor doing serious work will welcome them.
First: accurate on what data? A figure produced on a public benchmark or on the vendor's own curated set tells you very little about your documents, which are messier, differently formatted and full of domain vocabulary. The only meaningful demonstration is on a sample of your material, chosen by you. If a vendor resists that, they are telling you something.
Second: what counts as correct, and who decided? Extraction tasks need a definition - is a partially correct field a pass? Is the right value in the wrong place a failure? For anything generative, someone had to judge, and if the vendor judged their own outputs the figure is self-graded. Ask how disagreements between reviewers were resolved.
Third: what is the base rate and what does failure cost? A classifier that is ninety-five per cent accurate on a problem where the majority class is ninety per cent is barely better than always guessing the majority. And accuracy hides the asymmetry that usually matters - a false negative and a false positive rarely cost the same. In fraud detection, invoice processing or clinical triage, ask for precision and recall separately, because a single number deliberately conceals the trade-off between them.
Fourth: how does it behave when it is wrong? This is the question that separates production-ready systems from demos. Does the system express uncertainty, or is it equally confident when correct and incorrect? Can low-confidence cases be routed to a human automatically? A system that is ninety per cent accurate and reliably flags its uncertain cases is far more valuable than one that is ninety-five per cent accurate and gives no signal, because the first can be deployed safely and the second cannot.
The way to settle all of this is a paid pilot on your data with success criteria agreed in writing beforehand. Assemble a hundred or so real cases with known-correct answers, hold them back, and measure. It costs a fraction of the annual licence and it replaces the vendor's number with your own. Any vendor confident in their product will agree to it; the reluctance itself is the most useful signal in the whole evaluation.
