Multi-agent workflow for procurement document processing

Problem

Every tender produced dozens of supplier documents in different formats. Specialists manually re-typed prices, deadlines and compliance items into spreadsheets, which was slow and error-prone.

Full automation was not acceptable: every decision had to stay with a human and be traceable and auditable.

Approach & architecture

• Modelled the process as a graph of specialised agents: document classification, structured extraction, rule-based compliance checks and summarisation — coordinated by a supervisor with explicit state.
• Every agent output is validated against a schema; anything uncertain is routed to a human review queue with the source snippet highlighted.
• Kept the system explainable: each field in the final sheet links to the page and passage it came from.

Result

• Processing time per tender went from ~2 days to ~3 hours of review.
• Extraction accuracy on the pilot set: 97% for prices and dates once reviewer corrections were fed back into the system.
• Complete audit trail for every automated step.

What made it robust

• Deterministic checks (rules engine) for everything that does not need an LLM.
• Retries, timeouts and idempotent steps so a failed document never blocks the batch.
• Deployed inside the organisation's network; only the approved, minimised text reaches the model provider, in the agreed region — or a locally hosted model where required.