Contract intake and obligation tracking with Document AI — confidence scoring and review routing
An operations team was keying contract terms into spreadsheets by hand and missing renewal and notice deadlines. They wanted extraction they could trust — which meant knowing exactly when not to trust it.
The challenge
Contracts arrived as PDFs, scans and email attachments in dozens of templates. Manual intake was slow and inconsistent, and obligations — renewal windows, notice periods, payment terms — were tracked in spreadsheets nobody fully trusted. A fully automated extractor would have been fast and quietly wrong; the team needed automation with a reliable human checkpoint.
The approach
- Classifycontract type, template
- Extractterms via Document AI
- Scoreconfidence per field
- Routelow-confidence → review
- Trackobligations → deadlines
- Classify first. Incoming documents are typed (MSA, SOW, NDA, renewal) so extraction uses the right schema.
- Extract with confidence scores. Document AI pulls parties, dates, values, notice periods and renewal terms; every field carries a confidence score.
- Route, don't guess. Fields below threshold go to a reviewer with the source passage highlighted; approved values flow on. Thresholds were tuned against a labelled set of real contracts.
- Turn terms into obligations. Extracted dates become tracked obligations with owners and reminders, reported in BigQuery dashboards.
What we built
- Document classification and schema-based extraction pipeline on Google Cloud
- Confidence scoring with tuned thresholds and a review queue showing source passages
- Obligation register with owners, deadlines and notifications
- Dashboards for intake volume, review rates and upcoming obligations
Results
- Contract intake moved from hours of manual keying to minutes of review per document
- Human attention concentrated on the fields that actually needed it
- Renewal and notice deadlines surfaced automatically instead of discovered late
- An auditable trail from every tracked obligation back to the clause it came from
What made it work
Confidence is the feature
Knowing which 10% of fields to check is worth more than a slightly better average accuracy.
Show the passage
Reviewers approve in seconds when the source text is highlighted next to the extracted value.
Close the loop
Review decisions fed back into threshold tuning, so the queue shrank over time.
Government and enterprise programmes are described under confidentiality agreements; client identities, figures and architectures are generalised where required. Marketing results are reported as the clients described them.