AI Contract Review and Risk Analysis

We configure per contract type. Most deployments start with customer MSAs, NDAs, and vendor agreements because they're the highest-volume types. From there we add DPAs (data processing addenda), SaaS order forms, employment agreements, real estate leases, loan documents, and reseller agreements. Each contract type gets its own playbook YAML, its own extraction schema, and its own completeness check. Setting up a new contract type after the first one takes 3-5 days once we have your playbook for it. We do not support M&A transaction documents, complex financings, or litigation-adjacent agreements: the judgment density is too high and you'd spend more time reviewing the agent's output than doing it yourself.

On well-defined playbooks and high-volume contract types, clause identification accuracy runs 92-96% in our deployments based on blind test sets of 100+ contracts against attorney gold-standard review. Accuracy is highest on clauses with clear structural language (LOL caps, specific indemnification patterns, term and termination) and lower (88-92%) on clauses that use highly varied language for the same concept (e.g. force majeure definitions, change-of-control triggers). We test accuracy against your actual contracts before go-live using a test set your senior attorney reviews, and we publish the accuracy by clause type so the team knows where to focus second-pass review.

Yes. Output is a .docx file with tracked changes generated via python-docx, preserving the contract's original formatting and numbering. The file opens cleanly in Word, Google Docs, iManage, or Docusign CLM. Suggested changes are based on your playbook fallback language, not generic alternatives, and the agent adapts its drafting style to match the document's voice rather than pasting stock language. Attorneys accept, modify, or reject each suggestion exactly as they would with a colleague's redlines. For CLM deployments, we push the redline version directly into the CLM workflow so it shows up in the attorney's queue without a separate download step.

Three mechanisms. First, a completeness check: for each contract type, we define the set of expected clauses and the agent flags anything missing rather than only flagging deviations in clauses that are present. Something missing from a contract is as important as something wrong. Second, mandatory human review on high-risk contract types (customer MSAs above a revenue threshold, DPAs for regulated data, any contract with named counterparties on your escalate list). The agent recommends but never auto-approves. Third, a monthly spot-audit: a senior attorney reviews 10 randomly selected AI-reviewed contracts in full and we track the agreement rate. If agreement drops below 90%, we stop and recalibrate rather than quietly accumulate drift.

Novel clause structures, unusual jurisdictions, or contract types outside the configured playbooks route to human review explicitly flagged as 'outside playbook coverage'. The agent never pretends confidence on novel ground. For a clause that's technically covered but uses unusual drafting (e.g. a liability cap expressed as a multiplier of a variable rather than a fixed amount), the agent flags with a 'novel drafting pattern' tag and shows both the extracted interpretation and the source language side by side. Attorney override decisions write back to a drafting-pattern library used to improve future extraction without changing the playbook rules themselves.

Wrong most often means either missing a flag (a deviation the agent classified as acceptable when it shouldn't have) or over-flagging (marking a clause as problematic when it's actually fine). Missing flags are caught by the monthly spot-audit and by attorney review of the summary, because the attorney still reads the actual clauses the agent extracted. Over-flagging is caught immediately when the attorney accepts the contract position rather than the agent's redline. Every override writes to a log with the reason, and that log drives quarterly playbook refinements. The system improves with use. We publish the false-positive and false-negative rates to your general counsel monthly.

Every review produces a machine-readable audit artifact: contract ID, timestamp, playbook version used, every extracted clause with its source location, every flag with the specific rule citation, the redline diff, every attorney override with reason, and final disposition. Audits export as JSON, CSV, or directly into iManage or your CLM's audit trail. For regulated industries (financial services under model risk guidance, healthcare under HIPAA) we produce an annual attestation report showing model version, training signal, accuracy metrics, and human-review rates. Auditors get direct read access to the log through a scoped view.

A first-contract-type deployment (typically customer MSA) runs 8-10 weeks. Weeks 1-2 convert the playbook to YAML with your senior attorney. Weeks 3-4 build the extraction pipeline and test on 50 historical contracts. Weeks 5-6 calibrate using attorney feedback on test-set outputs. Weeks 7-8 run shadow mode where the agent reviews live contracts in parallel with the human team and accuracy is compared daily. Weeks 9-10 cut over to production with mandatory human review retained. Adding subsequent contract types takes 3-5 days each once the first one is live. Full rollout across a typical in-house team's contract portfolio takes 4-6 months end to end.