Snowflake cuts contract review time 70% with AI agent

Snowflake's audit team faced a scaling problem. As the company processed an expanding volume of customer contracts—ranging from structured capacity commitments to specialized marketplace agreements—each contract carried unique commercial terms that required careful human review for revenue recognition compliance. Auditors were spending hours manually scanning PDFs to find nonstandard terms and check revenue treatment for appropriateness, but they could only sample the population, not cover it exhaustively. "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," said Amrita Kapoor, VP Internal Audit. "We needed to flip the model. By letting AI handle the exhaustive reading across the entire population, our auditors can focus their expertise on exception handling."

Snowflake's Forward Deployed Engineer team built a Contract Review Agent that works in three layers. First, order form PDFs flow from source systems through Google Drive via Snowflake Openflow into a Snowflake stage, where a Cortex Agent powered by Cortex AI Functions handles layout extraction and uses AI Extract to pull structured fields: customer name, capacity amount, discount terms, payment schedules, and dozens of revenue-relevant data points. Second, the agent evaluates every extracted term against a playbook—a user-managed, audit-team-owned Snowflake table that defines what "standard" and "nonstandard" look like. The agent scores every clause against the playbook, producing a classification with confidence scores, clause-level excerpts, page references, and a natural-language explanation. Third, findings are presented in a reviewer-centric application where auditors can approve, override, or escalate every finding. Every correction—whether to an extracted value or a classification—is logged and persisted as extraction tips or playbook refinements that feed into the agent on every subsequent run.

The playbook is core to the design. Rather than shipping with a fixed model of what's "normal," the system puts control directly in the auditors' hands. Each rule defines a term, its detection criteria, and its classification. Every change is logged to an immutable audit trail showing who changed what and when. Rules take effect immediately on the next processing run without code changes or redeployment, meaning the audit team can respond to a new contract pattern in minutes. Charles Xu, Engineering Manager of Applied AI, noted: "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 system also catches novel terms—clauses that no existing rule anticipates. It uses a two-layer approach: first, a corpus of known-standard contract language is embedded and indexed, and clauses semantically distant from it are flagged as potentially novel; second, flagged terms are evaluated against existing playbook rules to separate genuinely novel language from known patterns. Auditors label these novel terms in a separate view, and those labels flow back into the system, sharpening the novel term detector over time. The results speak for themselves: review time dropped 70%, the audit team can now review thousands of order forms per quarter without linear headcount growth, every extraction and classification is logged for audit trail purposes, and accuracy improves with each review cycle as corrections and labels compound. Snowflake is now extending the same architecture to other agreement types, with PM Nikolai Scholz saying: "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."