AI that asks instead of guessing

Every AI demo has a moment the vendor hopes you won’t probe: what happens when something is missing? Most systems answer the question anyway. The model fills the gap from its training data, phrases it fluently, and the output looks complete. In most industries that’s a quality problem. In legal work it’s disqualifying — an analysis of a reference the system never read isn’t a rough draft, it’s a liability with formatting.

We took the opposite position and built it into the workflow contract: when the workflow needs something it doesn’t have, it asks.

What asking looks like

Upload an office action and the file wrapper, and the workflow checks the record for completeness before analyzing it. Two situations come up constantly:

A cited reference isn’t in the upload. The examiner relied on three references; you uploaded two. The workflow does not proceed to “analyze” the third from the model’s memory of what patents like it tend to say. If patent retrieval is connected, it fetches the cited patent by its publication number — retrieval of a named document, not a search. If it can’t, the run stops with a question: upload the reference, or proceed without it. Either way, the choice is on the record and yours.

A document doesn’t classify cleanly. Prosecution documents have roles — office action, specification, claims listing, amendment, prior art. When an upload could be more than one thing (a file named ReplyBrief_draft_v3.docx, say), the workflow presents the candidate roles as buttons and waits. One click resolves it. What never happens: the system silently files it somewhere plausible and builds an analysis on top of the misfiling.

Why questions beat guesses

A question costs you seconds. A guess costs you at the worst possible moment — after the analysis is drafted, when the error has propagated into a chart, an argument, and a recommendation, all of which read as confidently as the parts that were grounded.

There’s also a subtler benefit. Because the workflow only proceeds on documents it actually has, the finished report can make a claim most AI output can’t: every source is accounted for. The report’s “Documents Considered” table lists what was used — and anything relied on but never obtained is explicitly marked Not reviewed, rather than quietly summarized from nowhere.

Attorneys don’t need AI that never has gaps. They need AI that knows where its gaps are and says so. That’s a design decision, not a model capability — and it’s one we made at the foundation.