"The AI read every document." The line now appears in every second pitch for legal technology, and on its own it means very little. Reading is the cheap part. The question that matters is whether anything the system says afterwards can be traced back to what it read — finding by finding, clause by clause. That property has a name, traceability, and it is this property, not the reading itself, that changes how legal work is reviewed and relied on.

A finding without a source is an opinion

Take a due diligence memo that states: "several of the target's key contracts contain change-of-control clauses." The only question a board member or buyer can ask is: which ones, and how do you know? If the work was done by a junior associate, the answer is trust, plus whatever spot checks the partner had time for. If it was done by a system that cannot show its sources, the position is worse — there is no one to even ask.

The problem is not who produced the finding. It is the form of the finding: an assertion detached from its evidence. Traceability fixes the form.

What traceability means, concretely

Traceable output links every statement to the exact place it comes from: the clause, the page, the document — not "based on our review of the data room" but "clause 12.3 of the shareholders' agreement". Two further properties complete the picture. Coverage: the output states which documents were analysed and, just as importantly, which were not — corrupted files, missing schedules, documents in a language the process was not set up for. And granularity: claims are tied to passages, not to whole documents, because "somewhere in this eighty-page agreement" is not a source.

Review becomes "check the source"

Traceability changes the economics of supervision. Reviewing a conventional memo means either re-deriving the work or trusting it; in practice, most review is a blend — read closely where it feels risky, trust the rest. Reviewing traceable output is different in kind: open the source, read the clause, confirm or reject the finding. Seconds per finding, not hours per memo.

The reviewer's judgment then goes where it is actually needed — is this reading legally right, and does it matter for this transaction — instead of being spent on the prior question of whether the claim is even in the documents. Supervision becomes faster and stronger at the same time, which is a rare trade.

Sampling stops being the trade-off

Classic document review runs on sampling because it has to: read the largest contracts in full, skim the middle of the pile, leave the rest, and disclose the method in the report. Everyone involved knows where the risk sits — in the pile nobody read. When systems read every document at uniform depth and every finding is checkable at its source, that trade-off goes away. Coverage is complete, and verifying what was found is cheap.

What remains — and should be said plainly — is the risk of misreading: a document read but misunderstood. That risk is real, and it is exactly what the surrounding controls exist for: checklists by document type, a lawyer whose job is to challenge the findings, and escalation whenever the system is uncertain. Traceability does not remove the need for judgment. It removes the need for blind trust.

What to demand from any provider

Whether you are buying from us or from anyone else, four questions settle whether "the AI read everything" means something:

  • From any finding in the report, can I reach the underlying source in one step?
  • Do I receive a coverage list — including what could not be read, and why?
  • Are quoted passages verbatim, or paraphrased by the system?
  • Does the record show which findings a lawyer verified before delivery?

A provider with real traceability will happily demonstrate all four on a live matter. A provider who answers with accuracy percentages instead is telling you the findings cannot be checked, only believed. This is how we build every review we deliver — and if you want to see it on your own documents, we are glad to show you.