The Legal AI 'Playbook Divergence' Benchmarking Report 2026: How Much Do Internal Legal Department AI Playbooks Actually Differ From Outside Counsel AI Policies — and Where the Conflicts Are Creating Compliance Exposure

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

Executive summary

This report's title promises to measure how far corporate legal departments' AI requirements diverge from their outside counsel's internal AI policies. That measurement does not exist publicly. Outside counsel guidelines (OCGs) and law firm AI policies are almost always confidential. We found no published dataset that compares the two, and the percentages and sector breakdowns in earlier versions of this briefing had no verifiable basis. They have been removed.

The divergence is still real, and the reason is documented. ABA Formal Opinion 512 gives OCG disclosure terms ethical force. Surveys show that fewer than half of law firms have AI policies and most clients do not know how their firms use AI. Clients have been asking firms for AI disclosure since at least 2023. And 2026 court decisions on privilege and citation certification have raised the stakes. This briefing sets out that record and uses five conflict categories as a clearly labelled analytical framework, not as findings.

Why OCG AI terms now carry ethical weight

ABA Formal Opinion 512 (29 July 2024) is the anchor. On disclosure, it says lawyers "must disclose their GAI practices if asked by a client how they conducted their work." Disclosure is also required where "the client expressly requires disclosure under the terms of the engagement agreement," and the opinion extends the same point to the client's outside counsel guidelines. On confidentiality, it says "a client's informed consent is required prior to inputting information relating to the representation into" a self-learning generative AI tool. It adds that "merely adding general, boiler-plate provisions to engagement letters purporting to authorize the lawyer to use GAI" is not sufficient. It also requires client consultation "when [a tool's] output will influence a significant decision."

Analysis. A firm whose internal policy is "disclose if asked" is not necessarily non-compliant. But once a client's OCGs require disclosure, the firm's obligation follows the OCG, not the firm's default. The practical risk is less a contract dispute than a professional responsibility problem, which arises when a firm's policy is silent on something a client's OCG requires.

What is known about each side

Law firms. The Thomson Reuters 2025 Generative AI in Professional Services report found that 41% of law firms had policies governing generative AI use, either standalone or within a broader technology policy, according to LawSites. In the same research, 71% of corporate legal clients did not know whether their outside firms used generative AI. The Thomson Reuters Institute and Georgetown Law's State of the US Legal Market 2026 made the same point: most clients do not know whether or how their firms use generative AI, as reported by Legal IT Insider.

Legal departments. The ACC/Everlaw 2025 survey of 657 in-house professionals found that active generative AI use had risen to 52%, from 23% in 2024 (U.S. respondents). Only 9% said company policy prohibits generative AI use, down from 29%. 61% plan to push for changes in how legal services are delivered and priced, and only 24% were satisfied with how outside counsel are adopting the technology.

Client demands on firms. In November 2023, Bloomberg Law reported that some Paul Weiss clients had asked the firm not to use generative AI without first talking to them. Salesforce's chief legal officer said he expected lawyers to disclose generative AI use cases. In-house counsel were also asking about AI use in RFPs, and a Reveal executive advised in-house lawyers to "spell out AI policies in their outside counsel guidelines."

What OCG drafters are being told to include. A June 2026 practitioner guide published by Conventus Law recommends a dedicated OCG section covering four areas: - disclosure of AI use in work product - review and approval by a qualified lawyer before delivery - no client data sent to external AI systems without prior approval, with defined approved systems and storage locations - audit rights over how AI was used and who reviewed it

Analysis. Taken together, the picture is that clients are formalising AI expectations in OCGs and RFPs. Fewer than half of firms had any written AI policy in early 2025. And clients mostly cannot see firm practice. That combination is how divergence goes unnoticed until something goes wrong.

Five conflict categories (analytical framework)

The categories below are our framework for comparing an OCG with a firm policy. They are not survey results.

1. Data handling and model training. OCGs increasingly restrict client data to approved systems. Firm policies often rely on vendor enterprise terms. The difference matters more after United States v. Heppner (S.D.N.Y., February 2026). There, Judge Jed Rakoff held that a defendant's own conversations with a consumer AI chatbot were neither privileged nor work product. Part of his reasoning was that the consumer privacy terms allowed the provider to train on inputs and disclose them to third parties, according to Orrick. Orrick notes that negotiated enterprise terms may be treated differently but have not been tested. Courts are not uniform. In Warner v. Gilbarco, Inc. (E.D. Mich., February 2026), a magistrate judge treated a pro se litigant's ChatGPT materials as protected work product, according to Wilson Elser.

2. Disclosure. "Disclose if asked" firm policies and "disclose by default" OCGs conflict directly. Under Opinion 512, the OCG term controls the firm's disclosure duty.

3. Review and attestation. Most firm policies require lawyer review of AI output. Fewer turn that into a certification. Courts are moving toward certification anyway. In Mata v. Avianca, Inc. (S.D.N.Y. 22 June 2023), the court imposed a $5,000 Rule 11 sanction over ChatGPT-fabricated citations, according to Reed Smith. On 28 May 2026, the Supreme Court of Florida amended Rule 2.515, effective 15 June 2026, to require the signer of any filing to certify that "all legal authorities exist and are accurately cited." A simultaneous order preempted the patchwork of circuit-level AI disclosure orders, according to Jones Day. An OCG attestation requirement fits naturally alongside a court-level certification like this.

4. Notification triggers. OCGs may require notice before new tools or new uses. Firm policies often treat that as a relationship judgment. Opinion 512's consultation requirement for outputs that "influence a significant decision" gives clients a principled basis for defined triggers.

5. Prohibited uses. Typical candidates are consumer-grade tools, unreviewed final advice, and AI use on regulatory submissions without approval. A firm policy that does not separately prohibit what a client prohibits depends entirely on matter-level instructions reaching every timekeeper.

What happens when a conflict surfaces (analysis)

Without data, we can only describe likely pathways. Divergence most plausibly surfaces through a client question about AI use, an invoice review, a privilege dispute, or a court's citation certification. The likely consequences are re-review of work product, renegotiated engagement terms, and, in the worst case, a privilege or sanctions problem that neither policy anticipated.

Recommendations (analysis)

For GCs and legal ops leaders: - Put AI expectations into OCGs as specific, auditable terms covering disclosure, approved systems, review, notification and prohibited uses, rather than general principles. - Ask panel firms for their written AI policy in RFPs and annual reviews. - Remember that Opinion 512 obliges firms to answer when you ask how they did the work.

For law firm leaders: - Map your top clients' OCG AI provisions against firm policy and flag every term where the firm's default is weaker. - Push matter-level instructions into intake and time-entry systems so timekeepers see them. - Replace boilerplate engagement-letter AI consent, which Opinion 512 says is insufficient.

For both: - Agree in writing which platforms are approved and on what contractual terms, given Heppner's focus on provider terms. - Align attestation language with court certification rules such as Florida's amended Rule 2.515.

Sources

• American Bar Association, "Formal Opinion 512: Generative Artificial Intelligence Tools" (29 Jul 2024): https://www.americanbar.org/content/dam/aba/administrative/professional_responsibility/ethics-opinions/aba-formal-opinion-512.pdf
• LawSites, "Thomson Reuters Survey: Over 95% of Legal Professionals Expect Gen AI to Become Central to Workflow Within Five Years" (15 Apr 2025): https://www.lawnext.com/2025/04/thomson-reuters-survey-over-95-of-legal-professionals-expect-gen-ai-to-become-central-to-workflow-within-five-years
• Legal IT Insider, "Thomson Reuters' State of the US Legal Market report – Record profits and increasingly unstable ground" (7 Jan 2026): https://legaltechnology.com/thomson-reuters-state-of-the-us-legal-market-report-record-profits-and-increasingly-unstable-ground/
• Business Wire / ACC and Everlaw, "64% of In-House Counsel Expect GenAI to Reduce Reliance on Outside Counsel; 50% Expect Lower Outside Counsel Costs" (14 Oct 2025): https://www.businesswire.com/news/home/20251014382116/en/64-of-In-House-Counsel-Expect-GenAI-to-Reduce-Reliance-on-Outside-Counsel-50-Expect-Lower-Outside-Counsel-Costs-ACCEverlaw-Survey-Finds
• Bloomberg Law (Isabel Gottlieb), "Law Firms Wrestle With How Much to Tell Clients About AI Use" (29 Nov 2023): https://news.bloomberglaw.com/us-law-week/law-firms-wrestle-with-how-much-to-tell-clients-about-ai-use
• Conventus Law (Rohin Pujari), "How to Write Outside Counsel Guidelines That Govern AI-Assisted Work" (3 Jun 2026): https://conventuslaw.com/report/how-to-write-outside-counsel-guidelines-that-govern-ai-assisted-work/
• Orrick, "Court Rules AI Conversations Are Not Privileged: What United States v. Heppner Means for You" (Mar 2026): https://orrick.com/en/Insights/2026/03/Court-Rules-AI-Conversations-Are-Not-Privileged-What-United-States-v-Heppner-Means-for-You
• Wilson Elser, "Divergent: Two Federal Courts Chart Seemingly Opposite Paths on Work Product Protection for AI-Generated Materials" (5 Mar 2026): https://www.wilsonelser.com/publications/divergent-two-federal-courts-chart-seemingly-opposite-paths-on-work-product-protection-for-ai-generated-materials
• Reed Smith, "ChatGPT artificial intelligence tool generates fake opinions" (Mata v. Avianca, Inc., No. 22-cv-1461 (PKC) (S.D.N.Y. June 22, 2023)) (2023): https://www.reedsmith.com/our-insights/blogs/viewpoints/102ilzz/chatgpt-artificial-intelligence-tool-generates-fake-opinions/
• Jones Day, "Supreme Court of Florida Creates Statewide Standard for Artificial Intelligence Certification in Court Filings" (Jun 2026): https://www.jonesday.com/en/insights/2026/06/supreme-court-of-florida-creates-statewide-standard-for-artificial-intelligence-certification-in-court-filings