When AI reads my company's documents: accuracy from 26.7% to 86%

An AI given your contracts answers confidently and gets things wrong half the time when the document is long. Mistral measured the gap between guessing and searching. That figure also explains why scoping starts with your files.

An AI reading my company's documents gets things wrong more often than it appears: on FinanceBench, a set of questions about real financial documents, accuracy rises from 26.7% to 86% when the model searches in stages instead of swallowing everything at once (Mistral AI, Agentic Search announcement, August 20, 2026).

What the move from 26.7% to 86% measures

Mistral gave its model five tools to search a document instead of reading it straight through. On the same questions with the same model, accuracy more than tripled. The gain applies to office documents too: 51.9% on OfficeQA Pro, 45.6 points above direct reading. These measurements come from the publisher itself on public datasets, inviting us to read them as a demonstration of method rather than a market verdict.

Why the sales-meeting demonstration is always brilliant

A salesperson demonstrating AI on one of your documents chooses a question answered in a sentence on a page opened in front of you. The 26.7% describes the other situation: thirty questions about a two-hundred-page report, several requiring cross-referencing distant tables. Ask the question needing a cross-check and watch the response when information is missing. A system inventing a plausible answer costs more than one saying it does not know.

Your files are scoping material

The same lesson applies when having software built. How you work is already written somewhere: a quote template, route spreadsheet, meeting minutes, or the internal note explaining loyalty discounts. Describing all of it from memory in a conversation asks AI to guess. Attaching files during scoping asks it to search. That is the difference measured by the gap between 26.7% and 86%, transposed to your project.

What an agent team does with this material

In Maestro, you drop documents into the conversation, and Margaux, the agent scoping the need, uses them to ask questions instead of filling gaps. She returns a brief using your own terms: quote statuses, step names, rounding rules. You correct it, approve with 'Looks good to me', and the approved document becomes the reference for everything that follows. Attachments remain on your Mac like the rest of the project. To learn how to review this document without being a specialist, see reading a specification in ten minutes.

Where a document search tool serves you better

If your need ends at querying a document collection, contract database, or ten years of minutes, a staged search engine does the job and Maestro has nothing to add. The uses look similar from a distance but diverge in purpose: one answers questions, the other builds software. Choose according to what you want at the end, an answer or a product.

Prepare three documents before starting

Before your first scoping session, gather three files: the one you send customers, the one tracking daily activity, and the one you hate filling in. Those three contain most of your rules, including ones you never state. Attach them before describing your idea, and compare the resulting brief with the same conversation without files. The full method is in your specification in one hour, and the vocabulary you need not learn in I do not know the words.