How Generative AI Changed the Contract Review Bottleneck

Before generative AI tools became widely available to legal teams, the contract workflow bottleneck for most organizations was drafting — producing a first draft of a standard agreement took an attorney 45-90 minutes. That bottleneck has effectively been eliminated. Generative AI tools can produce serviceable first drafts of standard commercial agreements in minutes, and many legal teams are now using them for initial drafting of NDAs, vendor agreements, and SaaS subscription terms. The bottleneck has shifted downstream: from drafting to review. And that shift changes what contract review tools need to do.

The Volume Problem That Generative AI Created

When drafting was the bottleneck, the overall volume of contracts flowing through the review queue was constrained by drafting capacity. Legal teams had natural time to process incoming contracts simply because the outgoing drafts limited total deal velocity. Generative AI tools have removed that constraint for many organizations, and the result is a measurable increase in contract volume that arrives at the review stage.

This volume increase creates a specific problem for extraction and review systems: they need to be fast enough to keep up with the new pace of deal flow, not just accurate enough to perform well on a manageable queue. A clause extraction system that processes a 40-page MSA in 12 minutes was adequate when that MSA was one of three under review at any given time. It may be inadequate when the deal volume doubles because drafting friction was removed.

The more significant consequence of volume increase is that the case mix shifts toward the less-standard end. When deal volume increases, more of the incoming contracts are from newer counterparties, smaller relationships, and contexts where the counterparty is more likely to have used their own template rather than yours. The percentage of contracts that are genuine first-time drafts — where no precedent from a prior relationship exists — increases, making the extraction and deviation detection task harder on average across the portfolio.

Generative AI as a Source of Novel Clause Language

Generative AI drafting tools produce clause language that is statistically derived from patterns in training data — typically, large corpora of commercial contracts. The language produced is plausible and often legally coherent, but it can exhibit subtle variations in clause structure that differ from the specific patterns an extraction system was trained on.

This creates a specific problem for extraction systems trained on historically drafted contracts: the incoming contract population includes an increasing share of AI-drafted agreements whose clause structures are slightly different from the training distribution. An extraction system trained primarily on attorney-drafted agreements may have degraded performance on AI-drafted agreements with slightly unusual clause architectures — not dramatically different, but different enough to affect recall on specific clause types.

The practical indicator of this problem is a gradual increase in the percentage of flagged "unclassified" text blocks in incoming contracts — text that the extraction system processed but couldn't confidently assign to a known clause type. If your extraction system has been showing an increasing percentage of unclassified blocks over the past 12 months alongside an increase in AI-drafted incoming contracts, the two trends are likely related.

What Review Systems Need to Do Differently

The shift in bottleneck from drafting to review implies a few specific requirements for extraction and review tools that weren't as pressing when review was less time-constrained:

Triage before full extraction: When contract volume increases, not all incoming contracts have equal priority or complexity. A review system that runs full extraction on every incoming contract and presents results in queue order is less efficient than one that can perform a rapid triage pass — identifying contract type, estimated complexity, and rough risk tier — before full extraction, so that high-priority or high-complexity agreements are processed first.

Confidence-surfaced output: At higher review velocities, attorneys have less time to second-guess extraction results. A review interface that surfaces confidence scores prominently — so reviewers know which extractions are high-confidence and can be accepted quickly versus which require manual review — reduces per-contract review time significantly compared to one where everything is presented with equal apparent confidence.

Playbook comparison integrated into the first-pass view: In a high-volume review environment, requiring attorneys to separately look up playbook positions after reviewing extraction results adds friction that accumulates significantly across a large queue. Playbook comparison should be surfaced in the same view as the extracted clauses, so the attorney sees both the clause text and the deviation assessment simultaneously rather than sequentially.

The Verification Question for AI-Drafted Contracts

When a contract has been drafted by generative AI — the receiving party's, not yours — there's an additional verification consideration: does the AI-drafted agreement contain internal consistency errors that a normally-drafted agreement wouldn't have? Generative AI drafting tools occasionally produce agreements where defined terms are used inconsistently, where a cross-reference points to the wrong section, or where a numeric threshold defined in one section doesn't match a reference to that threshold in another section.

These internal consistency errors create legal ambiguity that may not be apparent from individual clause extraction results. A limitation of liability cap defined as "$500,000" in Section 10 and referenced as "the cap defined in Section 9" in Section 14 creates an ambiguity about which provision controls — an ambiguity that exists independently of whether either clause is well-drafted on its own terms. Cross-reference validation — checking that defined terms are used consistently and that cross-references point to existing sections — is a useful complement to clause extraction for AI-drafted agreements specifically.

What Doesn't Change

The fundamental requirements for reliable clause extraction don't change because the drafting tool changed. Recall still matters more than precision on high-stakes clause types. Playbook configuration still determines whether deviation detection is useful or just noise. The M&A due diligence limitations discussed in our article on extraction at scale still apply regardless of whether the contracts being reviewed were drafted by attorneys or AI tools. The underlying legal analysis required for context-dependent provisions is still attorney work.

What changes is the urgency of having extraction infrastructure that can scale with volume, triage intelligently, and surface results in a format that minimizes per-contract attorney time without sacrificing the review quality that legal risk management requires.

← Back to Insights