AI Contract Review: What to Automate and What Stays With Your Lawyers

Most contract review is two jobs done by the same person. The first is mechanical: find the parties, the term, the renewal mechanism, the liability cap, the governing law, the data-protection obligations, and copy them into a checklist. The second is judgment: decide whether a non-standard indemnity is acceptable for this client, this counterparty and this deal, and what to push back on. Experienced reviewers spend a surprising share of their day on the first job, which is exactly the part a well-built system can take over.

That framing matters because the question buyers usually ask — can AI review our contracts? — is the wrong one. Asked that way, the honest answer is no, and the projects that try end up either unused or quietly dangerous. The better question is which parts of review are mechanical enough to automate, which need a lawyer, and how to make the handover between them fast and auditable. This guide walks through that split and how to put it into practice.

What AI does reliably today

Modern language models are good at reading long, inconsistently formatted documents and locating specific provisions, even when they are worded unusually or buried in a schedule. Combined with a defined list of fields and clear instructions, that makes several review tasks dependable enough to automate, provided every output links back to the exact source text.

• Clause and entity extraction: parties, dates, term and renewal, payment terms, caps, notice periods, governing law and jurisdiction.
• Checklist completion: filling the firm's or legal team's standard review template, with each answer citing the clause it came from.
• Playbook comparison: flagging where a clause deviates from your preferred or fallback position, and by how much.
• Missing-clause detection: noting when an expected provision — a data-protection schedule, a termination-for-convenience right — is absent altogether.
• Portfolio questions: across hundreds of existing agreements, which ones auto-renew next quarter, or which lack a change-of-control clause.

What should stay with lawyers

The work that should not be automated is the work that depends on context the document does not contain: the commercial relationship, the client's appetite for risk, the counterparty's leverage, what was agreed on a call last week. A system can tell you that an indemnity is uncapped and that your playbook prefers a cap of twelve months' fees. Whether to accept it on this deal is a decision for someone accountable for the advice.

• Accepting or rejecting a deviation from the playbook.
• Negotiation strategy and the order in which to concede points.
• Interpreting ambiguous drafting where reasonable readings differ.
• Anything that will be relied on as legal advice to a client or the business.

Assisted redlining sits on the boundary. A system can propose replacement language drawn from the playbook, which saves drafting time, but the edit should only be applied when a reviewer approves it. Treat suggestions as a starting point that a lawyer accepts, edits or discards — never as changes that flow into a draft on their own.

Encode your playbook before you buy or build anything

The quality of automated flags depends almost entirely on how clearly your positions are written down. Many legal teams have a playbook in someone's head, or a document that describes preferred language without stating fallbacks or walk-away points. Off-the-shelf tools then flag against generic market standards, which produces a stream of warnings your reviewers learn to ignore.

Before any technology decision, take the ten to fifteen clause types that cause the most review effort and, for each one, record the preferred position, acceptable fallbacks, positions that need escalation, and who can approve them. This is legal work, not IT work, and it typically takes a few weeks of a senior reviewer's time. It is also useful on its own: junior reviewers get a clearer reference, and the team's positions become consistent.

Measure accuracy against your own reviewers

Vendor accuracy figures are measured on someone else's contracts. What matters is how a system performs on yours: your templates, your counterparties' paper, your scanned legacy agreements. Build a test set of a hundred or so contracts that senior reviewers have already processed, and compare the system's extraction and flags against their checklists field by field.

• Measure per field, not overall. Party names may be near-perfect while renewal mechanics are not, and the fields that matter most deserve their own threshold.
• Track misses separately from false alarms. A missed uncapped liability is far more costly than an unnecessary flag.
• Include your hardest documents — scanned PDFs, heavily amended agreements, contracts with long schedules — not only clean templates.
• Re-run the test set whenever the model, prompts or playbook change, so quality is not assumed to hold.

A mid-sized legal services firm we worked with took this approach. Clause extraction was built over the firm's own playbooks and historical reviews, with a 92% accuracy target measured against senior-reviewer baselines before scope was expanded. Because the comparison used the firm's real work, reviewers could see where the system was strong and where it still needed checking, which did more for adoption than any demonstration.

Design the review screen, not just the model

A frequent reason contract AI goes unused is that its output arrives as another spreadsheet or report that reviewers must reconcile against the document by hand. The time saved in extraction is lost in verification. The interface matters as much as the model: reviewers should see the extracted fields, the flags and the source document side by side, with one click from any answer to the clause that supports it.

In the legal services engagement, a structured review workspace replaced the spreadsheets the team had used, bringing extraction, risk flags and documents into one screen. Attorneys started from a risk summary rather than a blank checklist. The result was an 80% reduction in manual review effort and roughly three times the review throughput per team, with attorneys still making every judgment call.

Handle confidentiality and data protection deliberately

Contracts are among the most sensitive documents a business holds, and legal teams are right to ask where they go. Confirm where documents are processed and stored, whether any provider retains or trains on them, and how access is controlled so that a reviewer only sees matters they are entitled to. For Indian organisations, data-protection obligations under the Digital Personal Data Protection Act also apply to the personal data contracts often contain. Log every extraction, flag and approval, so that the review trail can be reconstructed if advice is ever questioned.

A realistic rollout

• Weeks 1–3: pick one contract type with high volume and a stable playbook — NDAs, supplier agreements or customer MSAs are common starting points — and write down positions and fallbacks for its key clauses.
• Weeks 4–8: build or configure extraction and playbook flags, and measure them against a test set of contracts your senior reviewers have already reviewed.
• Weeks 9–12: run the system alongside reviewers on live work, with every output checked, and track time per contract and missed issues.
• After that: reduce checking on the fields that have proven reliable, add assisted redlining, and extend to the next contract type.

Hold the programme to measures your team already cares about: turnaround time, contracts reviewed per reviewer, issues caught at review rather than after signature. If you want to understand whether your documents, playbooks and processes are ready for this kind of automation, our AI readiness assessment covers those questions, and our team is happy to talk through where a first contract type might start.