The Legal AI 'Clause Acceptance Rate' Benchmarking Report 2026: How Often AI-Suggested Contract Language Is Actually Accepted, Modified, or Rejected by Counterparties — and What That Tells Us About Real-World Tool Performance

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

Executive summary and what changed

The question in this briefing's title is a good one: how often does AI-suggested contract language survive negotiation? An earlier version answered it with acceptance patterns by clause type, deal size and tool category. Those were not measurements. We searched for public data and found none. No vendor, bar body, academic group or industry association that we could identify publishes counterparty acceptance rates for AI-drafted clauses. The earlier version's acceptance gradients, tool rankings and quoted counterparty remarks have been removed.

What does exist is a small body of public benchmarks of AI drafting and review quality, measured against lawyers in controlled settings. There is also long-standing evidence about which clauses parties fight over. This briefing summarizes that evidence, explains why it does not answer the acceptance question, and sets out how legal teams could measure acceptance themselves.

What the public benchmarks measure

LawGeex (2018). This is the most-cited early study. It compared an AI tool with 20 experienced lawyers reviewing five NDAs. The AI reached about 94% accuracy against about 85% for the lawyers, as the World Economic Forum reported in November 2018. It measured issue-spotting in review, not drafting, and not what counterparties accepted.

Vals Legal AI Report (February 2025). Vals tested several tools against a baseline of lawyers working without generative AI, across seven tasks. On redlining, the closest task to negotiation drafting, the lawyer baseline scored 79.7, ahead of Harvey Assistant at 65.0 and vLex's Vincent AI at 53.6. Only those two vendors opted into that task, and the report's narrative and results table give different figures for Harvey. On other tasks the tools often beat the lawyers. In document Q&A, for example, Harvey Assistant scored 94.8 and CoCounsel 89.6, against a lawyer baseline of 70.1. On data extraction, Harvey Assistant (75.1) and CoCounsel (73.2) edged the lawyer baseline (71.1). On document summarization, CoCounsel scored 77.2 against 50.3 for lawyers. The lesson is that tool performance varies sharply by task. Redlining, the task most like negotiation, was where lawyers held their lead.

LegalBenchmarks.ai contract drafting study (September 2025). As LawSites reported, this study assessed 450 outputs across 30 real-world drafting tasks from 13 AI tools, against in-house commercial lawyers. Human lawyers produced reliable first drafts 56.7% of the time; the AI average was 57%. The best tool, Gemini 2.5 Pro, scored 73.3%, against 70% for the best individual lawyer. In high-risk scenarios, specialized legal AI tools raised explicit risk warnings in 83% of outputs, compared with 55% for general-purpose tools. Speed was not close: the lawyers took nearly 13 minutes per task, while the tools produced output in seconds. The researchers noted that humans kept clear advantages on complex tasks needing commercial judgment, and that the results are a snapshot of fast-moving capabilities.

Stanford RegLab/HAI (May 2024). This study is about legal research, not contracts, but it is the main public evidence on tool reliability. It found that Lexis+ AI and Ask Practical Law AI each hallucinated on more than 17% of benchmark queries, and Westlaw AI-Assisted Research on more than 34%. Any drafting tool that cites authority carries the same verification burden.

None of these studies tracks a clause through to signature. Drafting quality judged by expert graders is not the same as acceptance by an opposing party with its own interests. The opposing party may reject excellent language for commercial reasons, or accept poor language because it doesn't care.

Which clauses get fought over

Long-running research on negotiation behaviour suggests acceptance would vary by clause type whoever drafted it. A 2009 survey by the International Association for Contract and Commercial Management (IACCM) found limitation of liability and indemnification were the two most negotiated terms. Price, IP rights, and confidentiality and data protection followed, as Out-Law (Pinsent Masons) reported. The data is old, but it supports a simple point. In any acceptance dataset, clause type would likely be a dominant variable, and contested clauses would show low acceptance whether a person or a model wrote them.

Why acceptance rate is a weak performance metric (analysis)

The earlier version's core argument still holds as analysis:

High acceptance can mean the clause didn't matter. Boilerplate on notices or termination for convenience may go unchallenged because neither side cares.
Counterparty sophistication confounds the result. An unrepresented or low-leverage counterparty may accept almost anything.
Deal pressure collapses negotiation. Late-stage clauses are often conceded to close.
The right outcome is sometimes rejection. A tool that drafts an aggressive opening position should see it negotiated down.

A better question is whether AI-assisted drafting moves outcomes toward the client's playbook position, faster, without new risk. Vendors are moving in this direction. Spellbook's October 2025 Series B announcement described a beta "Market Comparison" feature as a step toward suggestions grounded in real-time market data. Microsoft's message-center notice MC1388706 describes a Word legal agent that compares clauses to internal playbooks and proposes tracked changes. Comparing drafts to a playbook is measurable inside a team's own deal flow.

How to build your own acceptance data (practical recommendations)

Because no public benchmark exists, teams that want this data will have to build it. A minimal design:

Tag provenance at drafting time. Record whether each first-draft clause was AI-generated, AI-edited or human-drafted, and which tool and playbook version were used. Without this, redline histories cannot be attributed later.
Code outcomes at signature. Score each tagged clause as accepted as written, accepted with modification, or rejected. Record distance from the playbook position, not just acceptance.
Capture context variables. Record clause type, contract value band, counterparty representation and negotiation length. These are the likely confounders.
Track past execution. Disputes, amendments and claims involving the clause show whether accepted language actually performs.
Report against a human baseline. Compare AI-assisted and human-drafted clauses of the same type in similar deals. Treat early results as directional, given small samples. Revisit the coding scheme after the first quarter, once you can see which variables actually explain the variation in your own deals.

GCs evaluating vendors can ask whether a tool records clause provenance and outcome data, and whether that data stays under the customer's control. Firms advising clients on AI governance can apply the same measurement discipline to their own drafting.

Bottom line

There is no public evidence base for clause acceptance rates of AI-suggested contract language. Controlled benchmarks show AI tools now match or beat lawyers on some drafting and review tasks. Lawyers kept the lead on redlining in the one public head-to-head test, and judgment-heavy work still favors people. Until organizations instrument their own negotiations, claims about "acceptance rates," including any we might have published, should be treated as unsupported.

Sources

World Economic Forum, "This AI outperformed 20 corporate lawyers at legal work" (Nov 2018): https://weforum.org/stories/2018/11/this-ai-outperformed-20-corporate-lawyers-at-legal-work
Vals AI, "Vals Legal AI Report (VLAIR)" (27 Feb 2025): https://www.vals.ai/industry-reports/vlair-2-27-25
LawSites (LawNext), "AI tools match or exceed human lawyers in contract drafting benchmark study" (Sep 2025): https://www.lawnext.com/2025/09/ai-tools-match-or-exceed-human-lawyers-in-contract-drafting-benchmark-study.html
Stanford HAI, "AI on Trial: Legal Models Hallucinate in 1 out of 6 (or More) Benchmarking Queries" (23 May 2024, updated 30 May 2024): https://hai.stanford.edu/news/ai-trial-legal-models-hallucinate-1-out-6-or-more-benchmarking-queries
Out-Law (Pinsent Masons), "Defensive contract wranglings only prepare for failure, says negotiators' body" (16 Jun 2009): https://www.pinsentmasons.com/out-law/news/defensive-contract-wranglings-only-prepare-for-failure-says-negotiators-body
Business Wire / Spellbook, "Spellbook Raises $50M Series B to Expand AI Contract Review Platform" (9 Oct 2025): https://www.businesswire.com/news/home/20251009110230/en/Spellbook-Raises-$50M-Series-B-to-Expand-AI-Contract-Review-Platform
Microsoft 365 Message Center (via mc.merill.net), "MC1388706 – Microsoft 365 Copilot: Legal agent available in Frontier worldwide" (updated 7 Oct 2026): https://mc.merill.net/message/MC1388706