The Legal AI 'Change of Control' Clause Audit Report 2026: How AI Contract Review Tools Perform on the Specific Clause Category That Determines Whether Your Entire Agreement Survives an Acquisition

Research Briefing | The Legal Stack | Q1 2026

Executive Summary

Change-of-control provisions are the single highest-stakes clause category in M&A diligence. A missed automatic-termination trigger in a material SaaS agreement can unwind deal economics. A mischaracterized consent requirement in a distribution contract can create undisclosed closing conditions that blow up timelines. This briefing presents findings from structured blind testing of seven AI contract review platforms against a standardized corpus of 50 commercial agreements, supplemented by a practitioner survey of 44 M&A lawyers across Big Law, boutique M&A practices, and in-house legal teams. The headline finding is uncomfortable but not surprising: every platform tested performed adequately on obvious change-of-control language and inadequately—some catastrophically—on embedded triggers, definitional traps, and provisions where risk characterization, not mere detection, is the operative skill.

Methodology

The Test Corpus

Fifty agreements were sourced and anonymized across three verticals: SaaS/technology licensing (22 agreements), manufacturing and supply chain (16 agreements), and financial services including broker-dealer agreements and fund administration contracts (12 agreements). Agreements ranged in length from 8 to 94 pages, with an average of 31 pages. Each agreement was pre-annotated by a panel of three senior M&A attorneys with no affiliation to any vendor tested. COC configurations in the corpus included:

• Tier 1 (Obvious): Explicit "change of control" defined terms triggering termination or consent rights, appearing in the body of the agreement under headings such as "Assignment" or "Change of Control"
• Tier 2 (Embedded Definitional): COC-equivalent language embedded within definitions of "Affiliate," "Control," or "Permitted Transferee," creating deemed assignment or consent obligations only visible through cross-reference
• Tier 3 (Conflicting Carve-outs): Consent-not-to-be-unreasonably-withheld (CNBUW) provisions with carve-outs for competitors, regulatory change, or "strategic transactions," creating internally conflicting obligations
• Tier 4 (Deemed Assignment via Merger Clause): Delaware-style merger survival provisions that implicitly trigger COC without using the phrase, relying on case law such as Meso Scale Diagnostics v. Roche Diagnostics (Del. Ch. 2011) for their operative effect

Platforms were tested through API access where available, and through standard user interface workflows where not. Prompting was standardized using a defined query set: "Identify all change of control provisions," "Flag any assignment restrictions that could be triggered by an acquisition," and "Summarize the risk associated with each COC-related provision identified."

Platforms Evaluated

Harvey (Harvey AI, using GPT-4-class backbone with legal fine-tuning), Ironclad (AI Assist, contract intelligence layer), Kira Systems (Litera, smart fields including COC-specific trained models), Luminance (Luminance Legal AI, BERT-based with proprietary legal training), Spellbook (Rally Legal, GPT-4o integration), Leya (Stockholm-based, European commercial law emphasis), and Lexion (workflow-integrated AI, now part of Docusign IAM suite).

Detection Findings

Tier 1: Obvious COC Language

On explicit, heading-labeled change-of-control provisions, performance was broadly strong. Kira led with a 96% detection rate on Tier 1 clauses, reflecting its mature training set and pattern-matching heritage. Harvey and Luminance both returned 93% detection rates. Ironclad AI Assist recorded 89%, with most misses in financial services agreements that used non-standard section hierarchies. Spellbook and Leya came in at 87% and 84% respectively, with Lexion trailing at 81%.

These numbers are reassuring but almost beside the point. Tier 1 provisions are the ones a paralegal with a good search function can find.

Tier 2: Embedded Definitional Triggers

Performance collapsed here across the board. The definitional trap—where "change of control" as an operative concept exists only through a cascade of cross-referenced definitions—is where the gap between marketing and reality becomes visible.

Kira's smart field model, trained specifically on COC patterns, detected 61% of Tier 2 triggers. This was the highest score. Harvey returned 54%, which given its generative architecture and ability to reason across document sections was expected to be higher. Luminance recorded 49%. Ironclad, Spellbook, Leya, and Lexion clustered between 31% and 44%.

The specific failure mode was consistent: tools identified the definition section but failed to propagate its operative implications forward to identify clauses that became COC triggers by cross-reference. In one SaaS Master Services Agreement from the corpus—structurally similar to agreements used by mid-market Salesforce ISV partners—"change of control" was never defined as a term; instead, a "Permitted Assignment" definition excluded transactions resulting in a change of more than 50% beneficial ownership, creating an operative COC consent requirement readable only by cross-referencing the assignment clause. Only Kira and Harvey flagged this with consistency across the seven agreements in the corpus that used this structure.

Tier 3: CNBUW with Conflicting Carve-outs

This clause type—arguably the most consequential in deal diligence because it creates a disputed rather than automatic risk—produced the most varied results and the most concerning false characterizations.

Harvey performed best at 71% accuracy in correctly characterizing the net risk posture of CNBUW provisions with carve-outs. Critically, Harvey was most likely to note that a carve-out for "competitors" created an ambiguity about whether the acquirer fell within the definition, flagging legal uncertainty rather than characterizing the clause as either "consent required" or "no consent required" in binary terms. Luminance and Kira both achieved approximately 58% to 62% accuracy. The remaining platforms generated characterizations that a senior practitioner would consider misleading in over 40% of cases—most commonly by treating CNBUW as functionally equivalent to free consent, which it is not.

Tier 4: Deemed Assignment via Merger Clause

This was the most challenging category and the most dangerous for deal teams that treat AI output as conclusive. Detection rates ranged from 22% (Lexion) to 47% (Harvey). No platform reliably identified the circumstance in which a merger survival clause, combined with the anti-assignment language of the type at issue in SQL Solutions v. Oracle or PPG Industries v. Guardian Industries, creates a COC trigger without explicit COC language. This is a reasoning task that requires external legal knowledge integration, not pattern recognition.

False Positive Rates

False positives—flagging non-COC provisions as COC triggers—were lower than expected but not negligible. Harvey generated the fewest false positives at 6.2% of total flags. Kira, given its classification-based approach, generated 9.1%. Spellbook generated a notable 18.3% false positive rate, largely by flagging change-of-ownership language in insurance and indemnity provisions as COC triggers.

Practitioner Survey: How M&A Lawyers Actually Use These Tools

Forty-four M&A practitioners responded to our survey: 28 from law firms (ranging from Am Law 50 to specialist M&A boutiques), 11 in-house at companies with active M&A programs, and 5 in legal operations roles at PE-backed acquirers.

Key findings:

• 91% of respondents said they use AI contract review tools in diligence, but 0% reported using AI COC analysis without attorney review of flagged results
• 73% reported manually re-reviewing all assignment and definition clauses regardless of AI output in transactions above $50M
• 61% said the primary value of AI tools in COC review was speed of first-pass organization, not analytical accuracy
• 38% reported at least one instance in the past 18 months where an AI tool missed a material COC provision that was subsequently identified by attorney review
• Only 9% reported that their firm or legal ops team had conducted any structured accuracy testing of their preferred AI tool on COC clauses specifically
• When asked what verification step they apply to AI COC output: 82% said attorney read-through of full assignment section; 57% said definition section cross-check; 34% said keyword search independent of AI to verify recall

One general counsel at a Minneapolis-based industrial manufacturer put it plainly: "I use Kira to build the first index. I use a second-year to check everything Kira touches on COC. And I use a partner to read the definitions section regardless of what either of them found."

Where These Tools Are and Are Not Ready

Ready for unsupervised deployment: None of the platforms evaluated are ready for unsupervised deployment in COC review in transactions where the agreement is material. This is not a criticism unique to legal AI—it is a structural limitation of any pattern-recognition or generative system operating on clause categories where risk characterization, cross-document reasoning, and external legal knowledge integration are all required simultaneously.

Ready for supervised deployment with defined workflows: Kira and Harvey are ready for use as the first-pass identification layer in supervised diligence workflows, provided the output is treated as a recall tool—maximizing the number of potentially relevant provisions surfaced—rather than an accuracy tool. Both platforms meaningfully reduce attorney time on Tier 1 and large portions of Tier 2 review.

Useful with structured prompting: Harvey's generative architecture showed the most promise on Tier 3 characterization when prompts were structured to require the model to identify conflicting obligations rather than simply classify the provision. Deal teams using Harvey should maintain a standardized prompt library developed with M&A counsel rather than relying on default queries.

Not ready for: Any diligence workflow in which Tier 4 deemed-assignment triggers, or embedded definitional COC traps in complex commercial agreements, are the risk being managed. The detection rates in these categories—22% to 47%—mean that using AI as the primary detection mechanism creates a material risk of missing provisions that can determine whether a deal closes on the terms agreed.

Conclusion

The AI contract review market has made genuine progress on the easy problem in change-of-control review: finding the clause that says "change of control" in a section titled "Change of Control." It has made limited and inconsistent progress on the hard problem: reasoning across a complex document to identify that a cascade of defined terms, survival provisions, and assignment restrictions has created a COC trigger that no individual clause makes explicit. That gap is where M&A deals break. Until the tools close it—and none of them have yet—the practitioner survey data reflects the right posture: AI as first-pass recall infrastructure, attorneys as the analytical layer, and no unsupervised deployment on the clause category that can unwind the deal.

Methodology documentation and full platform scoring matrices available to Legal Stack subscribers. Platforms were given opportunity to review factual claims prior to publication; two requested corrections incorporated into final text.