Snowflake cuts contract review time 70% with AI agent

Snowflake's Forward Deployed Engineer team built an internal AI-powered contract review agent to solve a scaling problem: as the company processes an ever-growing volume of customer order forms, MSAs, and amendments, each carrying unique commercial terms, the traditional model of auditors manually scanning PDFs for nonstandard clauses that affect revenue recognition stopped working. In the words of Amrita Kapoor, VP Internal Audit, the audit team was "spending hours manually scanning PDFs to find nonstandard terms and yet only providing assurance on a sample of the entire population." The solution was to "flip the model. By letting AI handle the exhaustive reading across the entire population, our auditors can focus their expertise on exception handling."

The Contract Review Agent works in three layers. First, order form PDFs flow from source systems through Google Drive into a Snowflake stage via Snowflake Openflow. A Cortex Agent powered by Cortex AI Functions extracts layout and structured fields—customer name, capacity amount, discount terms, payment schedules, and dozens of revenue-relevant data points—using AI Extract. Second, the agent evaluates every extracted term against a user-managed playbook, a governed Snowflake table that the audit team owns and edits directly without requiring an engineering ticket. The agent scores each clause as standard or nonstandard, producing confidence scores, clause excerpts, page references, and a natural-language explanation. Crucially, the system also applies novel term detection: it embeds a corpus of known-standard contract language and flags clauses semantically distant from it, then auditors label whether those anomalies are meaningful or routine noise—labels that flow back to sharpen future runs. Third, findings surface in a reviewer-centric application with portfolio-level dashboards and contract detail views. Auditors can approve, override, or escalate every finding; corrections are logged as extraction tips and stored in long-term memory, fed into the agent on subsequent runs. A parallel Snowflake CoWork agent gives stakeholders a natural-language interface to audit status without direct app access.

The system achieved a 70% reduction in review time—from days to hours—while processing thousands of order forms per quarter without scaling the audit team proportionally. Crucially, every extraction, classification, rule change, and correction is logged with full provenance, the kind of evidence trail external auditors expect. Engineering Manager Charles Xu noted that "the agent is not just parsing but reasoning. It explains why a term is considered nonstandard, cites the playbook rule and presents the contract excerpt, letting the reviewer confirm or correct. Every decision is logged." The playbook itself is auditor-controlled: rules take effect immediately on the next processing run without code redeployment, so when a novel discount structure appears in Q4 deals, auditors add a rule and the agent starts flagging it across the entire corpus. Snowflake is now extending the same architecture to other agreement types, with PM Nikolai Scholz observing that "without it, you grow your operational team linearly. With it, you keep the team lean while expanding coverage — and the system gets smarter every quarter because the experts are teaching it."