Gerenciamento de documentos com IA e processamento inteligente

An incoming file is analysed and a category, folder and document type are proposed. An invoice arriving in a watched mailbox lands in the right place without a person opening it to work out what it was, which removes the sorting step that makes email-based intake unmanageable at volume.

The proposal is a proposal. It is presented for confirmation, and a user can override it; the override is recorded and improves nothing silently. We are deliberate about this because a classification that is wrong and unchallenged is worse than no classification: it puts a record where nobody will look for it.

Named fields are extracted into metadata — supplier, invoice number, amount, contract value, counterparty, renewal date, policy number, expiry date. Which fields depends on the document type, so an invoice and a lease are read for different things.

The source text for each extracted value is retained alongside it, so a reviewer can see where a figure came from without opening the document and hunting. That traceability is what makes extracted metadata safe to report from: a number you cannot trace is a number you have to re-verify, which defeats the purpose.

Long agreements and reports get a short structured summary: what the document is, what it commits the organisation to, and what is unusual about it. The purpose is triage — deciding whether this is the document that needs an hour of attention today.

A summary is not an approval basis and the interface does not present it as one. Approving a document you have not read remains a bad idea whether or not a model summarised it, and the audit trail records the approval against the document, not against the summary.

Semantic search matches meaning rather than strings, so "the lease that renews in the third quarter" can find an agreement whose text reads "term expires 30 September". Keyword search fails in exactly this way: you remember the substance and not the wording.

Semantic and keyword results are returned together rather than as separate modes, and both are permission-filtered before they reach the user. Making someone choose a retrieval algorithm is asking them to solve our problem.

Every AI action is written to the immutable audit trail as its own event type: what was proposed, what was accepted, what was overridden, and by whom. If a misfiled document later matters, you can establish whether a model proposed the location and a person accepted it, or a person chose it outright.

This is the difference between AI in a document system and AI in a consumer product. The requirement is not only that the suggestion is good; it is that the provenance of every value in a record remains attributable years later.