AI contract review: what actually works

Contract review works when the model has a defined playbook and must show its work. It fails when a team uploads an agreement, asks “is this safe?”, and treats fluent prose as approval. A five-minute pass can be an excellent first reader. It is not the signature decision.

What models catch reliably

With clean text and a specific checklist, models are useful at locating clauses, extracting parties and dates, comparing notice periods, identifying absent sections, and finding deviations from approved language. They can normalize a long document into a review table and draft a redline explanation. Repetition is an advantage: the role does not become bored on agreement number fifty.

Every finding should include the exact quoted text, page or clause location, playbook rule, severity, and proposed next step. For absent-clause findings, it should record where it searched and avoid inventing a quotation. These cited cells let a lawyer verify the result quickly.

What models still miss

Failure modes include definitions that alter a later clause, obligations spread across schedules, bad OCR, incorporated external documents, inconsistent numbering, and commercially unusual risks absent from the playbook. The role may also over-flag harmless language or miss the practical meaning of a remedy. Governing law and deal context can turn a familiar clause into a different question.

That is why the workflow includes document completeness checks, OCR confidence, cross-reference resolution, and escalation on uncertainty. Counsel evaluates interaction effects, enforceability, negotiation posture, and business appetite.

Build the review skill

The skill should contain contract-type checklists, approved and fallback language, risk definitions, examples, and a required output schema. Store it in git. Build an evaluation set from representative agreements, including messy scans and known traps. Score extraction, citation accuracy, missed material issues, and false alarms separately.

The legal solution produces a review file, not a chat answer. A lawyer accepts, edits, or rejects each material finding. Proposed redlines and matter-status updates default to Ask. Only the lawyer communicates the position and approves signature. That boundary gives the team speed without converting model confidence into legal authority.

Run it as a role, not a prompt

The durable implementation is a role inside the company operating system. Its instructions, checklists, examples, and connector definitions live in git, so every change has an author, review, and rollback path. Each run gets a sandbox and an auditable record. The team can use the best model for each step instead of tying the workflow to one vendor. Models can change; the role, tests, permissions, and history remain.

Connections are brokered server-side. CRM, document, mail, and accounting credentials never sit in a prompt or a browser extension. Read access is scoped to the records needed for the task. Writes default to Ask: the agent prepares the proposed update, message, or file, then an accountable person approves it. Self-hosting is available when policy, residency, or network boundaries require it.

Make the first pilot measurable

Choose one queue with enough volume to observe within two weeks. Record the current cycle time, rework rate, backlog, and escalation rate before the first run. Test historical cases, including awkward and incomplete ones, before touching live work. During the pilot, compare accepted outputs, corrected outputs, false escalations, and time returned to the team. A useful role becomes more reliable because corrections are committed back to its skill and evaluation set.

Keep the boundary explicit. The agent can collect evidence, apply a checklist, draft, route, and update systems after approval. A named employee owns exceptions and consequences. That division is what turns model capability into dependable operations without pretending that probability is judgment.

This work runs on Neuro OS. To scope a first role, get started.

Explore Neuro OS →