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 — a review of public sources. The Legal Stack did not conduct a survey, benchmark or interviews for this piece.

What the public evidence does and doesn't show

This report was meant to assess how AI contract review tools perform on change-of-control (COC) provisions. No public, independent benchmark measures that. We found no published study that scores named platforms on COC detection, separates simple from embedded triggers, or reports false-positive rates for this clause type. An earlier version of this article ranked seven platforms tier by tier and described what practitioners commonly do. That material was qualitative judgement presented as if it were evidence, and we have removed it. That version also misdescribed two cases.

What can be documented is:

the case law that makes COC review hard
what vendors publicly say their tools detect
the general benchmark evidence on legal AI accuracy

From those, the practical conclusion follows. AI is a recall aid for COC review, and lawyers still have to do the analysis.

Why change-of-control review is a legal-reasoning problem

Whether a deal "triggers" a contract often depends on deal structure and governing law, not on whether the words "change of control" appear in the contract. Three cases show this.

Meso Scale Diagnostics v. Roche Diagnostics (Del. Ch. 22 Feb 2013). Vice Chancellor Parsons held that a reverse triangular merger generally does not effect an assignment "by operation of law." The anti-assignment clause in that case was therefore not breached. The court declined to follow SQL Solutions.
SQL Solutions, Inc. v. Oracle Corp. (N.D. Cal. 18 Dec 1991). Judge Patel's decision has been cited for the view that an assignment occurred when the licensee became another company's wholly owned subsidiary. It rested on federal copyright law's restriction on transferring licences, and it did not analyse non-assignment clauses. A California corporate law commentary notes that later courts have declined to follow it.
Cincom Systems, Inc. v. Novelis Corp. (6th Cir. 25 Sep 2009). Applying the Sixth Circuit's patent-licence precedent in PPG Industries v. Guardian Industries (1979), the court held that a forward merger of a licensee into a sister company transferred a non-transferable software licence without consent. It said: "A transfer is no less a transfer because it takes place by operation of law."

Analysis: the same anti-assignment clause can mean different things depending on deal structure (forward merger, reverse triangular merger or stock purchase), on whether a federal IP licence is involved, and on governing law. A tool that only finds clauses cannot answer whether this transaction triggers this clause. That takes knowledge of outside law and of the deal itself.

The earlier version described SQL Solutions and PPG as cases about "merger survival clauses" creating COC triggers. That was inaccurate. They concern whether a merger itself transfers a licence that cannot be assigned.

What vendors say their tools detect

Vendor documentation confirms that the leading extraction tools target this clause family directly. Litera says Kira identifies "over 1,400" common clauses, provisions and data points through built-in smart fields. Its M&A due diligence group includes change of control and assignment, and users can train custom fields with Quick Study. Litera describes the fields as trained by experienced lawyers using patented supervised machine learning. Those accuracy claims are the vendor's own, and we found no independent COC-specific verification.

Analysis: finding a clause headed "Assignment" or "Change of Control" is the easier task. COC risk is often hidden in other places:

definitions such as "Control," "Affiliate" or "Permitted Transferee"
consent-not-to-be-unreasonably-withheld (CNBUW) language with carve-outs
termination rights tied to ownership changes

Extraction models built to locate provisions may still miss these cross-referenced structures or describe them wrongly. We have no public data on how often that happens. Buyers should treat it as a hypothesis to test, not a known rate.

What general benchmarks tell us

Two public studies are relevant by analogy, though neither tests COC clauses.

Vals Legal AI Report (27 Feb 2025). The study evaluated Harvey, Thomson Reuters CoCounsel, vLex Vincent AI and Vecflow Oliver on seven tasks, including document extraction, document Q&A and redlining, according to LawSites. On document Q&A, Harvey scored 94.8% and CoCounsel 89.6%. The AI tools collectively surpassed the lawyer baseline on data extraction. On redlining, the lawyer baseline of 79.7% outperformed the AI tools. LexisNexis withdrew from most tasks.
Stanford RegLab/HAI study (May 2024). Using more than 200 open-ended legal queries, it found that Lexis+ AI and Ask Practical Law AI hallucinated more than 17% of the time, and Westlaw AI-Assisted Research more than 34%.

Analysis: the picture is strong performance on locating and answering questions about text, and weaker performance on tasks that require legal judgement or reliable citation to outside law. That is consistent with the COC problem described above. Strong extraction scores do not show that a tool can tell whether a reverse triangular merger triggers a Delaware-law anti-assignment clause.

The older 2018 LawGeex study, in which an AI tool scored about 94% against about 85% for 20 lawyers on five NDAs, is often cited. It concerned NDA issue-spotting, not M&A consent analysis.

Analysis: a supervised workflow for COC review

The following is The Legal Stack's recommended practice, not observed survey data.

Use AI for recall, not conclusions. Run the tool's COC and assignment fields across the data room, set to surface as many candidates as possible, and accept some false positives.
Always review definitions by hand. Have a lawyer read the "Control," "Affiliate," "Permitted Assignment" and similar definitions in every material contract, whatever the tool flagged.
Map triggers to deal structure. Before analysis, record the transaction form and the governing law of each material contract. Then apply Meso Scale-type reasoning (Delaware, reverse triangular merger) or Cincom/PPG-type reasoning (federal IP licences, mergers) as appropriate.
Characterise, don't classify. For CNBUW clauses and competitor carve-outs, ask the tool, and the associate, to set out the ambiguity rather than give a binary "consent required" answer.
Check recall independently. Run keyword searches separately from the AI ("control," "merger," "assign," "operation of law," "beneficial ownership") and reconcile the two sets of results.
Build your own test set. Collect past deals' COC provisions, including embedded and definitional examples, and measure each tool's recall before relying on it. This is the benchmark the market lacks, and a firm can build it privately. Re-run it whenever the vendor changes the underlying model or the smart-field version, because results on one release do not carry over automatically to the next.

The conclusion of the original article still stands on this firmer basis. Current tools speed up first-pass organisation, but the hard COC questions are legal-reasoning questions, and a lawyer has to answer them.

Sources

California Corporate & Securities Law, "Courts Consider Anti-Assignment Clauses And Reverse Triangular Mergers" (Oct 2013): https://www.calcorporatelaw.com/2013/10/courts-consider-anti-assignment-clauses-and-reverse-triangular-mergers
FindLaw, Cincom Systems, Inc. v. Novelis Corp., No. 07-4142 (6th Cir. 25 Sep 2009): https://caselaw.findlaw.com/us-6th-circuit/1397941.html
Litera, "Instantly start projects with Kira's out-of-the-box smart fields": https://prep.litera.com/products/legal2/kira/built-in-intelligence/
LawSites, "Legal AI tools show promise in first-of-its-kind benchmark study, with Harvey and CoCounsel leading the pack" (28 Feb 2025): https://www.lawnext.com/2025/02/legal-ai-tools-show-promise-in-first-of-its-kind-benchmark-study-with-harvey-and-cocounsel-leading-the-pack.html
Stanford HAI, "AI on Trial: Legal Models Hallucinate in 1 out of 6 (or More) Benchmarking Queries" (23 May 2024): https://hai.stanford.edu/news/ai-trial-legal-models-hallucinate-1-out-6-or-more-benchmarking-queries
Potter Anderson, "Meso Scale Diagnostics, LLC v. Roche Diagnostics GmbH, C.A. No. 5589-VCP (Del. Ch. Feb. 22, 2013) (Parsons, V.C.)" (22 Feb 2013): https://www.potteranderson.com/insights/cases/Meso-Scale-Diagnostics-LLC-v-Roche-Diagnostics-GmbH-C-A-No-5589-VCP-Del-Ch-Feb-22-2013-Parsons-V-C
AALS, "Study Finds AI Is More Accurate Than Lawyers for Reviewing Contracts" (2018): https://www.aals.org/study-finds-ai-is-more-accurate-than-lawyers-for-reviewing-contracts/