Due diligence is hundreds of documents, a deadline, and a small number of things that genuinely matter buried among them. It looks like the ideal application for AI, and in one sense it is — the search cost is enormous and mostly mechanical.
The reason it is dangerous is that a reviewer cannot re-check everything the system reported as fine. That is the entire time saving. So any confident false negative — a system saying nothing is wrong with a document when something is — is invisible by construction, and lands at the worst possible moment.
I designed the output to be evidence rather than conclusion: structured findings that each carry their source passage, as-of-date context attached rather than inferred, and organisation that supports item-by-item verification. The reviewer's job does not change — the cost of finding the items to review does.
Evidence-first output: structured findings, source attribution on every item, as-of-date awareness carried through, and tenant isolation throughout. The system shortens the search and leaves the judgement where it belongs.
Every due-diligence AI demo shows the system finding a problematic clause. The interesting question is what it does with two hundred documents where nothing is wrong, and one where something is — and whether the reviewer would notice if it missed that one.
That is not a question I can answer with a demo. It is a question about how I structure the output, and whether it invites verification or discourages it.
What it runs on