Contract Scraper → CRM + Upsell Signal Engine (L4)

The steps

1. 01

   Ingest PDFs three ways for the model

   Tool: Self-hosted service

   For every closed-won contract: extract raw text + run OCR + render to PNG. Send all three to the model. Opus reads the image, the text, the OCR, accuracy goes up dramatically vs any single input.

2. 02

   Use a deterministic JS auditor, not just AI

   Tool: n8n + JavaScript

   Hard rules catch hallucinations cheaply. Example: if licenses = unlimited then seat_count must be null. If contract_value ≠ sum(line_items), flag. Anti-pattern called out in episode: stuffing PDFs into an agent and hoping.

3. 03

   Loop low-confidence cases back through with context

   Tool: n8n

   When the auditor finds a mismatch, push the error message + scraped data back to the agent: "X doesn't reconcile with Y, re-evaluate." Only if it still fails does it route to a human.

4. 04

   Human-in-loop only on the residual

   Tool: Slack notification

   Last audit before Salesforce write happens with a human, but only on the ~5% the auditor couldn't resolve. Push validated structured data into Salesforce custom fields.

5. 05

   Layer upsell signals on top

   Tool: Salesforce + agent

   Once contract terms are structured, an agent watches calls + product usage. "Mentioned HRIS hub on the call 2 weeks ago, not in their plan" → tee up for AM. This is where NRR moves.

Tools in this playbook

• Self-hosted service
• n8n + JavaScript
• n8n
• Slack notification
• Salesforce + agent

Next playbooks

• Voice-of-Customer → Product & Enablement Loop (L2)
• Zapier Agents for sales ops chores (L2)
• Agentic Nightly CRM Hygiene + Ownership Sync (L3)
• AI-driven price-quote acceleration with CPQ (L3)