Legal review is expensive because it is a skilled professional reading a great deal of text to find a small number of consequential things. That is a shape AI fits well — provided you design for the specific failure mode that matters in legal work.
Recall over precision
In most AI applications you balance false positives against false negatives. In legal review the asymmetry is severe: a flagged clause that turns out to be fine costs a lawyer thirty seconds. A missed change-of-control provision can cost a great deal more.
So legal pipelines should be tuned to over-flag. Design the review interface to make dismissing a false positive nearly free, then push recall as high as it will go.
Measure recall against a set of contracts a lawyer has fully marked up. If the pipeline misses clauses a human found, no amount of speed improvement matters yet.
What works well
Clause identification and extraction. Finding termination, indemnity, limitation of liability, assignment, change of control, governing law, auto-renewal, and payment terms across a contract portfolio. This is the core capability everything else builds on.
Playbook comparison. Checking each clause against your standard position and flagging deviations by severity. This is where in-house teams get the most leverage — it turns a two-hour review into a twenty-minute review of the exceptions.
Obligation and deadline registers. Extracting every commitment, notice period, renewal date, and reporting duty into a trackable register. Most organizations genuinely do not know what they have agreed to across thousands of contracts; this produces immediate, visible value.
Due diligence triage. Ranking thousands of documents in a data room by likely relevance so the expensive humans read the important ones first.
Precedent and prior-work retrieval. Finding how your firm handled a similar clause before.
First-draft generation from a playbook and a term sheet, for a lawyer to revise.
What requires a lawyer
Novel legal questions, strategy, negotiation, jurisdiction-specific advice, anything privileged or adversarial, and any final sign-off. The model assembles and proposes; a lawyer decides. This is both a quality and a professional-responsibility boundary.
Practical pipeline notes
Parsing is most of the battle. Contracts are structurally complex — numbered clause hierarchies, cross-references, defined terms, schedules, and amendments that modify clauses elsewhere. Parsing that preserves clause numbering and hierarchy is worth more than any model upgrade.
Defined terms matter enormously. "Confidential Information" means whatever clause 1.4 says it means. A pipeline that retrieves a clause without its definitions will misread it. Resolve defined terms as part of normalization.
Amendments change the operative text. The current position on a contract is the original plus every amendment. Systems that review only the base document produce confidently wrong answers.
Confidentiality is an architectural constraint. Client and counterparty material frequently cannot leave a defined boundary. Confirm data residency, retention, and training-use terms before any document moves. For many firms this determines deployment model entirely.
Frequently asked questions
Can AI replace a contract lawyer?
No, and the framing misleads. It replaces the reading, locating, and tabulating that consume most review hours. The judgement about whether a deviation is acceptable, what to concede, and how a clause interacts with the commercial deal remains legal work.
How accurate is AI clause extraction?
For common, well-defined clause types in reasonably standard contracts, recall of 90–97% is achievable. It degrades on unusual drafting, heavily negotiated bespoke agreements, and poorly scanned documents. Always validate against lawyer-marked contracts from your own portfolio rather than trusting a vendor benchmark.
Is it safe to send contracts to a commercial AI provider?
It depends entirely on the contractual terms you have with that provider — specifically data retention, training use, sub-processors, and residency — and on your own obligations to clients and counterparties. Enterprise agreements commonly exclude training on your data, but this must be verified rather than assumed, and some material simply cannot leave your environment.
What about privilege?
Sharing privileged material with a third-party processor raises genuine waiver questions that vary by jurisdiction. Most firms handle this with a deployment model that keeps privileged content inside their own boundary, plus documented processor terms. Take advice specific to your jurisdiction — this is not a settled area.
Where should a legal team start?
An obligation register over your existing executed contracts. It is bounded, measurable, does not touch live negotiation, and almost always surfaces commitments the organization had lost track of — which makes the business case self-evident.