AI Contract Review, With the Liability Kept Where It Belongs

Extraction yes, judgement no

Contract review has two parts: finding the relevant provisions and deciding whether they are acceptable. The first is a search and extraction problem and AI handles it well. The second is professional judgement with liability attached.

Every safe deployment we have seen keeps that line firmly. The system prepares; the person decides.

What it does well

• Clause identification — find the limitation of liability, the termination provisions, the payment terms, across hundreds of documents
• Extraction into a register — renewal dates, notice periods, values, counterparties
• Comparison against a standard — how does this differ from our template?
• Missing clause detection — what would we expect to see that is not here?
• Plain-language summaries for non-lawyers, clearly marked as summaries

The register use case alone justifies many projects. Most organisations cannot answer “which contracts auto-renew in the next 90 days” without someone reading files for a week.

What it should not do

Decide whether a term is acceptable. Advise on negotiating position. Approve a contract for signature. Interpret an ambiguous provision as settled.

The test we apply: if being wrong would create liability for the business, a qualified person must make the decision with the document in front of them. AI can tell them where to look and what changed; it cannot own the conclusion.

The workflow that works

1. Contract arrives and is classified by type
2. Key provisions extracted into a structured summary with links to the source clause
3. Comparison against your standard positions, with deviations flagged by severity
4. Reviewer sees the deviations first, with the original text alongside
5. Reviewer's decisions recorded, and the register updated automatically

The gain is that a reviewer spends their time on the five clauses that differ rather than reading forty pages to find them.

Accuracy and how to handle the gaps

Clause identification on standard commercial contracts is reliable; unusual drafting, heavily negotiated documents and scanned originals are less so. Design for the miss: reviewers should always have the full document available and should never be encouraged to trust the summary alone.

Track which clause types are missed and feed those into your evaluation set. Failures cluster by drafting style, which makes them addressable.

Confidentiality comes first

Contracts contain commercially sensitive material and often personal data. Decide where processing happens, what is retained, and whether any provider could use the content for training, before a single document is uploaded.

For many organisations this means regional processing under an enterprise agreement, plus internal access control so the contract register is not more widely visible than the contracts themselves.

Frequently asked questions