Vendor quote PDF to line items: an extraction playbook that survives production

Most quote extraction efforts fail in production for one of two reasons:

1. They treat extraction as a one-shot "AI magic" step (no review loop)
2. They cannot safely re-run when a document changes (duplicates, drift, broken audit trails)

Use this playbook instead:

• AI quote extraction
• Procurement Agent
• Trust and governance
• Agent runtime

Step 1: Treat extraction as draft generation

The output of extraction should be a draft that is reviewable:

• Keep the original document linked
• Keep the extracted structure visible (line items + totals)
• Keep a clear "what changed" story through review

Step 2: Put deterministic checks in front of fallbacks

Deterministic checks help you catch the obvious failures early:

• Totals math tolerance checks
• Minimum structure checks (rows/columns)
• Header coverage checks (vendor + currency + date/terms)

When these checks fail, fall back explicitly and label uncertainty.

Step 3: Route exceptions to a review queue

In procurement, exceptions are normal:

• Missing units
• Ambiguous line descriptions
• Different vendor formatting

The review queue is what makes the system usable, not the extraction model.

Step 4: Connect extraction to comparison (not to commitment)

Extraction should feed:

• Quote comparison

But it should not silently trigger:

• Award decisions
• Purchase order issuance

Those require approvals:

• PO approval workflow

Step 5: Make re-runs safe

Assume you will need to rerun:

• Updated PDFs
• Late vendor revisions
• Corrections after review feedback

Your system should be designed so reruns do not create duplicates and do not erase audit history.