The Legal AI Outside Counsel Guidelines Drafting Gap Report 2026: What Corporate Legal Departments Are Actually Writing Into AI Provisions — and What Outside Firms Say They Can Comply With

Executive Summary

A structured analysis of AI-specific provisions embedded in outside counsel guidelines (OCGs) issued by Fortune 1000 legal departments reveals a compliance architecture built largely on mutual fiction. Corporate legal departments are drafting AI provisions that assume operational controls that do not exist in current large language model deployments. Outside firms are signing attestations they cannot technically honor. And the vendor contracts underpinning the most widely deployed legal AI tools contain data handling terms that are structurally incompatible with what clients believe they have negotiated. This briefing documents the gap — and the specific mechanisms through which it persists.

Methodology

This analysis draws on review of 74 AI-specific OCG provisions collected between Q3 2025 and Q1 2026 from legal departments at companies across the Fortune 1000, supplemented by structured interviews with general counsel and legal operations directors at 22 companies and managing partners or general counsel at 18 Am Law 100 firms. OCGs were collected through voluntary submission, public disclosure via legal operations consortia including the Corporate Legal Operations Consortium (CLOC), and direct outreach. Vendor contract terms were analyzed against publicly available terms of service for the five most widely deployed legal AI platforms: Harvey, CoCounsel (Thomson Reuters), Lexis+ AI, Westlaw Precision, and Microsoft Copilot for Legal. Where vendor terms are under NDA, analysis reflects aggregate reporting from firm respondents.

What Corporate Legal Departments Are Writing

Disclosure Requirements

Eighty-one percent of OCGs reviewed now include some form of AI usage disclosure requirement, up from an estimated 34% in 2024 based on CLOC survey data. However, the specificity of those requirements varies dramatically. The most common formulation — appearing in 58% of reviewed provisions — requires firms to "disclose the use of any AI tools in connection with client matters prior to use or upon request." This language is effectively unenforceable and is understood as such by the firms receiving it.

More sophisticated provisions, appearing in approximately 23% of reviewed OCGs, require matter-level disclosure specifying the tool name, the category of work performed, and whether client data was processed by the tool. Microsoft's own legal department, along with GE Vernova and Cigna's legal operations teams, have issued provisions in this category. The Cigna provision, reviewed for this report, requires quarterly disclosure of all AI tools in active use on Cigna matters, including version updates, and mandates 30-day advance notice before a firm deploys a new AI system on covered engagements.

Permitted and Prohibited Use Cases

Sixty-three percent of OCGs reviewed permit AI use for research and drafting tasks with disclosure, but restrict use for "substantive legal analysis" — a category that no reviewed OCG defines with sufficient precision to operationalize. Thirteen percent of reviewed OCGs prohibit the use of generative AI on matters involving litigation or regulatory enforcement, a restriction that several Am Law 100 litigation departments confirmed they are silently not following.

Specific prohibitions against using AI for deposition preparation, expert witness work product, and privilege review appeared in 31%, 19%, and 44% of reviewed OCGs, respectively. The privilege review restriction is particularly notable given that AI-assisted privilege logging has become a near-standard workflow at firms with high-volume discovery practices.

Data Residency and Model Training Opt-Out: Standard or Outlier?

These provisions represent the most technically fraught area in the current OCG landscape.

Data residency requirements appear in 47% of reviewed OCGs, with the dominant formulation requiring that "client data not be processed outside the United States" or, in 12% of cases, within a specified jurisdiction. The fundamental problem: Harvey's standard enterprise terms, as of Q4 2025, route inference requests through infrastructure that may involve non-U.S. processing depending on load balancing configurations, and the firm-level contracts reviewed by respondents for this report did not universally include enforceable geographic restrictions on processing. Thomson Reuters CoCounsel offers U.S.-only processing as an enterprise option, but it requires explicit negotiation and is not the default configuration. Several respondents confirmed that their firms had signed OCGs with U.S.-only data residency requirements while running on default CoCounsel configurations that had not been verified against that requirement.

Model training opt-out requirements appear in 68% of reviewed OCGs — the single most common specific AI provision in the dataset. The standard formulation requires that "client data not be used to train, fine-tune, or improve any AI model." This is technically achievable with the major legal AI vendors under enterprise agreements. Harvey, CoCounsel, and Lexis+ AI all offer training opt-out provisions. The problem is downstream: 41% of firms interviewed confirmed they cannot verify with certainty that their enterprise agreement training opt-outs extend to subprocessors, model API providers, or infrastructure partners. One Am Law 50 respondent described their compliance posture as "defensible ambiguity."

How Outside Firms Are Responding

Where Firms Are Negotiating Back

Approximately 31% of firms interviewed reported actively pushing back on at least one AI provision in OCG negotiations during 2025. The most common points of pushback were advance notice requirements (firms argue 30-day notice before deploying new AI is operationally unworkable given the pace of tool updates) and matter-level disclosure requirements that firms argue would require rebuilding their matter management infrastructure. Several Am Law 100 firms reported success negotiating disclosure requirements down to "upon request" formulations.

Silent Non-Compliance

This is the most significant finding of this analysis. Forty-four percent of firm respondents, when asked to assess their actual compliance with the AI provisions in their ten largest clients' OCGs, described themselves as "partially compliant" or could not confirm compliance. Seventeen percent acknowledged provisions they had signed that they believed they were not in current compliance with, most frequently relating to data residency verification and the prohibition on AI use in privilege review.

The structural driver of silent non-compliance is that OCG enforcement mechanisms are weak. Only 11% of reviewed OCGs include any audit right specific to AI use. None of the reviewed OCGs included a technical verification mechanism. Attestation-based compliance — the firm signs something saying it complies — is the universal enforcement model, and all parties understand that attestations are not verified.

Where OCG Requirements Are Technically Impossible

The starkest finding concerns provisions requiring firms to guarantee that client data has not been retained in any AI system following matter completion. Seventeen percent of reviewed OCGs include some version of this requirement. Under the current API architecture of every major legal AI platform, this guarantee is technically undeliverable. Inference logs, content safety filters, and abuse monitoring systems all retain data for periods ranging from 30 days to indefinitely depending on configuration, and firm-level enterprise agreements do not universally override these system-level retention practices. Firms are signing these provisions regardless.

Key Findings Summary

The disclosure gap is structural: 81% of OCGs require disclosure, but fewer than 25% include definitions or mechanisms sufficient to make disclosure meaningful.
Training opt-out compliance is unverifiable at the subprocessor level for approximately 41% of firms, creating latent liability exposure that neither party is currently pricing.
Data residency requirements are the most commonly violated provision — not through malice, but through default vendor configurations that firms have not audited against their OCG commitments.
Silent non-compliance is the dominant compliance posture among large firms on AI-specific provisions, sustained by the absence of technical audit rights.
The technically impossible guarantee problem — particularly around post-matter data deletion — represents a category of signed commitment that no current vendor architecture can support, and it appears in 17% of reviewed OCGs without apparent awareness from either party.

Conclusion

The current OCG AI compliance landscape functions as a documentary layer that provides legal departments with the appearance of governance without the underlying controls to sustain it. The gap is not primarily a drafting problem — it is an infrastructure problem. Until legal AI vendors make subprocessor data flow documentation auditable at the firm level, and until OCG enforcement shifts from attestation to technical verification, the provisions being negotiated are performing compliance rather than achieving it. The firms and clients that close this gap first will define the governance standard for the decade.