The Contract Review Pipeline

Let me be direct about what Claude can and cannot do with contracts. It can read a 40-page master services agreement and extract every material clause in two minutes. It can flag indemnification language that shifts risk to your side. It can identify missing clauses that your standard template requires. It can compare a redlined version against your approved template and list every deviation. What it cannot do is practice law. It cannot tell you whether to sign. It cannot assess jurisdiction-specific enforceability. It cannot replace your general counsel's judgment.

The pipeline that works in production treats Claude as a paralegal, not a lawyer. Stage one: extraction. Claude reads the contract and produces a structured summary — parties, term, termination provisions, liability caps, indemnification, IP ownership, confidentiality, payment terms, and any non-standard clauses. Stage two: risk flagging. Claude compares the extracted terms against a checklist of your organization's risk tolerance — clauses that deviate from your standard positions get flagged with severity levels. Stage three: human review. Your legal counsel reviews only the flagged items, not the entire contract.

The economics are compelling. A 40-page contract that takes a lawyer two hours to review takes Claude two minutes for extraction and flagging. The lawyer then spends 20 minutes reviewing the flagged items instead of two hours reading the whole document. You have reduced legal review time by 80% without removing legal judgment from the process.

The risk checklist is the key artifact. It encodes your organization's specific positions: "Indemnification must be mutual." "Liability cap must be at least 2x annual contract value." "IP created during engagement must be assigned to client." "Termination for convenience requires 30-day notice." Each item becomes a pass/fail check that Claude applies mechanically. The checklist evolves as your legal team encounters new patterns — every contract that surfaces a previously unchecked risk becomes a new checklist item.

1. 1. Build Your Risk Checklist Work with legal counsel to define 15-25 specific checks: indemnification symmetry, liability caps, IP assignment, termination provisions, payment terms, data handling, jurisdiction, force majeure. Each check has a pass condition and a severity level (acceptable, review, critical).
2. 2. Create the Extraction Skill Build a skill that reads a contract and produces a structured JSON output: parties, term dates, key clauses by category, and any non-standard language. The structured output is what the risk flagging stage consumes. Unstructured summaries are not useful for automated checking.
3. 3. Build the Risk Flagging Stage A second agent reads the extraction output and evaluates each item against the risk checklist. For every deviation, it produces: the clause text, the expected standard, the actual language, the severity level, and a plain-English explanation of why it matters.
4. 4. Format for Legal Review The output to legal counsel should be a one-page summary with a risk heat map: green items (standard), yellow items (review recommended), red items (must negotiate). Lawyers do not want a 10-page AI analysis. They want a prioritized list of what needs attention.