Most commercial contracts are reviewed against a standard: a playbook that says what the company prefers, what fallback it accepts, and what it will not sign. The review is largely comparison. An attorney reads each clause, recalls or looks up the position, and notes where the draft deviates. For routine paper such as NDAs, order forms, vendor terms, and data processing addenda, the comparison is repetitive and the judgment is concentrated in a small number of deviations.
That shape suits a system that does the first pass and presents deviations for an attorney to resolve.
Turn the playbook into checks
The playbook is the product. For each clause type, write down three things: the preferred position, the acceptable fallbacks, and the red lines. Add the reason for each, in a sentence. Reasons let the system explain a deviation and let the attorney judge whether it applies. Store the playbook as structured entries, versioned, with an owner. A vague playbook gives vague reviews, and building the system often exposes where the playbook was never written down.
The design
- Contractany format, converted to text
- Classifysmall model: type, parties, governing law
- Segmentclauses, with locations
- Clause checksparallel, one per playbook entry
- Whole-document passcross-clause interactions
- Proposefallback language, cited
- Attorneyaccepts, edits, rejects
- Review recorddeviations, decisions
- Route by contract type. A small model classifies the document and selects the playbook. An NDA and a software subscription are checked against different entries.
- Check clauses in parallel. Each playbook entry is a separate, narrow check: find the relevant clause, compare to the position, report match, deviation, or missing. Anthropic's parallelization guidance notes that separate calls for separate considerations often outperform one call handling everything.
- Cite the text. Every deviation quotes the clause and its location. The attorney should never have to search the document to verify a claim.
- Add a whole-document pass. Clause-by-clause checking misses interactions: an uncapped indemnity that sits outside a limitation of liability, a definition that changes the meaning of a clause elsewhere. A final pass with the full contract in context looks specifically for these. The long context window makes this feasible for most agreements.
- Propose language. For each deviation, draft fallback wording from the playbook. A second model checks that the proposal does not introduce a new conflict with the rest of the document.
Limits to design around
Interactions. The whole-document pass helps, but it is where misses are most likely, so keep it in the evaluation as its own category.
Judgment. A deviation is not always a problem. Commercial context, negotiating position, and the relationship decide. The system reports and explains. It does not decide to accept a risk.
Confidentiality. Contracts are confidential and often privileged. Decide the processing environment up front. Anthropic's zero data retention arrangement covers the core Messages API and citations. Hosted features that store data, such as Batch, the Files API, and Agent Skills, fall outside it, so a legal build avoids them. The data handling article has the table.
Responsibility. The attorney remains responsible for the advice. Make that visible in the interface: the attorney's name and decision are attached to each review record.
How to evaluate it
- Attorney-labeled set. Take past contracts with the deviations an attorney actually flagged. Measure recall of those deviations, per clause type.
- False alarms. What share of reported deviations does the attorney dismiss? A noisy system gets ignored.
- Citation accuracy. Does each quoted clause exist at the stated location? Code can verify this by string match against the source text, which is a cheap and decisive check.
- Proposed language. Have attorneys rate acceptability on a short scale. Use a model judge to scale this up and sample the results with a person.
Start with one contract type that is high in volume and low in risk. Expand only as the measured recall supports it.
This is a reference design. It describes how we would build the system and cites Anthropic's documentation for platform facts. It is not a client case study, and the figures it uses are list prices or worked examples, not results.
Have a process like this?
Bring it to a 45-minute working session.