AI contract extraction: why every fact needs a source

Separate the suggestion from the business record

Contract extraction turns document text into structured fields such as a counterparty, a value, an expiry date or a notice period. The model output is a proposal. It should not silently become the number used in a financial total or the date used for an operational commitment.

Keep the original value, the proposed value and the reviewer’s decision distinct. That gives the reviewer a clear task: inspect the supporting passage, confirm the interpretation and either accept or reject the suggestion. A confidence percentage alone cannot replace this check.

Ask what the citation actually proves

A page number helps a person find the relevant part of the document. An exact quote lets a system check that the passage appears on that page. Those checks help expose fabricated citations, but they do not prove that the interpretation is correct. A clause can contain multiple dates, exceptions or conditional amounts.

For example, a passage might mention both an initial term and a later renewal term. The words can be quoted accurately while the wrong date is proposed as the current expiry date. The reviewer needs enough surrounding text to understand which term applies. Keep access to the source document alongside the extracted passage.

Keep document text inside its data boundary

Uploaded text can contain instructions, including text that resembles a prompt. An extraction system should treat that material as document content and request only a narrow set of structured outputs. It should validate types, supported field names and document limits before storing candidates.

TheContractTool processes one organization-scoped document per extraction job. Its architecture validates candidate structure, checks quoted text against the cited page and stores source-generation references. It has no shared cross-company AI conversation or vector index. These controls are part of a layered approach, rather than a guarantee that every proposed fact will be correct.

Know the limits before you use the result

A scanned PDF may contain page images without selectable text. The current extraction architecture requires selectable PDF text or TXT input; it does not include OCR. An unsuccessful parse should be visible as a failed job, rather than quietly producing an apparently complete summary.

In this MVP, AI extraction is disabled by default. The review interface can be explored with synthetic examples. Activating live extraction requires a configured model, service permissions and an explicit operating decision about location and cost. Operational metadata remains manually entered or human approved.

Common questions