Contract review where counsel only reads the deviations

Trigger: watched folder · Pattern: extract clauses → playbook gate → file or escalate · Sample: examples/use-cases/contract-review.toml · Status: runs today (intel-remote,schema,trigger-fs-watch)

The problem

Sales wants the deal signed this week. Legal wants to read every line. Both are right: most inbound paper is your own template or a benign variant, and the one contract with an uncapped-liability clause buried in section 14.3 is precisely why legal reads everything. So counsel becomes the bottleneck for a stack of documents that are 90% routine — the worst possible use of the most expensive reading time in the company.

Legal teams formalized the answer years ago: the playbook — the list of clauses you care about and the positions you accept. What's been missing is a reader that applies the playbook tirelessly and knows what it's not allowed to decide.

What the agent does

  1. A contract (text export) lands in the intake folder; the workflow fires.
  2. One schema-enforced LLM step extracts the playbook clauses — liability cap, auto-renewal, governing law — and renders one verdict the schema forces to an enum: risk: "standard" | "review". The prompt's rule is strict: any deviation or any missing clause → review.
  3. standard → a clause memo is filed next to the contract, automatically.
  4. review → the run checkpoints for counsel. They read the memo — which names exactly which clause deviated and what it says — then resume to file, or kill the run and pick up the phone.

Counsel's queue shrinks to the documents that actually contain decisions.

Why this isn't "AI lawyering"

The line this design refuses to cross is the one that matters professionally: the model never approves anything. It extracts and compares; the only two outcomes it can trigger are "file a memo" and "wake up a lawyer." The asymmetry is structural —

[[edges]]
from = "gate"
when = "review"
to = "counsel_review"     # pause_for_approval — a human resumes

— and the failure direction is chosen deliberately: uncertain extraction fails toward counsel (schema repair → declared failure → human), never toward silent filing. A false "review" costs ten minutes; a false "standard" was the thing we built the gate to prevent.

The memo trail is its own win: every contract gets a structured record of what its key clauses said at intake, searchable later when someone asks "which agreements auto-renew in Q1?"

Honest limits