What Federal Agency AI Use Policies Actually Say About AI-Assisted Legal Work Product: A Reading of Published OMB Guidance, DOJ AI Strategy Documents, and GSA Procurement Standards Through October 2026

Overview

Federal agencies deploy AI tools for an expanding range of legal functions—drafting regulatory preambles, reviewing procurement contracts, conducting litigation research, and analysing statutory text. Yet the core governance documents that govern this use were written primarily with operational, scientific, and benefits-adjudication AI in mind. The result is a significant structural gap: federal AI policy is increasingly specific about transparency, risk classification, and procurement integrity, but largely silent on the doctrinal questions that matter most when the government is a party to litigation or holds privileged communications with agency counsel.

This briefing reads the governing documents on their own terms and identifies where they speak, where they are ambiguous, and where they are silent on legal work product.

What OMB M-24-10 Actually Requires

OMB Memorandum M-24-10, issued in March 2024, established the foundational AI governance framework for civilian federal agencies; it was rescinded and replaced by OMB M-25-21 on April 3, 2025. It requires each agency to designate a Chief AI Officer, conduct annual inventories of AI use cases, publish minimum risk practices for AI affecting rights and safety, and expand the federal AI use case inventory that OMB maintains publicly.

M-24-10 required that agencies not deploy AI in ways that harm rights or safety without specific minimum practices in place—including ongoing monitoring, human oversight mechanisms, and the ability to opt out of AI-driven decisions in favour of human review. (These requirements were carried forward in modified form under the successor OMB M-25-21.) For AI tools affecting "rights-impacting" or "safety-impacting" categories, M-24-10 required agencies to provide notice to affected individuals, ensure meaningful human review of AI outputs, and conduct independent assessments. (OMB M-25-21, which replaced M-24-10 in April 2025, consolidated these into a single "high-impact AI" category.)

What M-24-10 does not do is distinguish between operational AI (processing benefit applications, flagging fraud in grant programmes) and legal AI (drafting enforcement memoranda, reviewing agency contracts, analysing regulatory comments). The minimum practices apply at the level of use-case category, not at the level of professional role. An agency deploying an AI tool to review procurement contracts is within scope if the review affects rights or safety; but the question of whether a lawyer in the agency's general counsel office using the same tool is subject to the same regime—or a different one—is left unaddressed. The memo does not reference the work product doctrine, the attorney-client privilege, or the Federal Rules of Evidence.

What the DOJ AI Strategy Documents Say

The Department of Justice has published AI use policies through its AI Strategy and, separately, through specific guidance from the National Security Division, the Civil Division, and the Executive Office for Immigration Review. DOJ's public AI strategy emphasises that AI use by DOJ attorneys must be consistent with applicable law, ethical obligations, and existing policies on the use of government information systems.

DOJ's guidance on AI in litigation contexts—to the extent it is publicly available through October 2026—concentrates on disclosure and accuracy obligations. DOJ attorneys, like all federal litigants, are bound by Rule 11 of the Federal Rules of Civil Procedure, which requires certification that submissions are not presented for improper purpose and are supported by the evidence and law. DOJ has issued internal reminders that AI-generated content requires human verification before submission to a court, consistent with the reasoning in Mata v. Avianca, Inc., No. 1:22-cv-01461 (S.D.N.Y.), where Judge P. Kevin Castel imposed $5,000 in sanctions on 22 June 2023 for submission of ChatGPT-generated fictitious citations. DOJ's published guidance treats this verification obligation as subsumed within existing professional responsibility requirements rather than as a freestanding AI-specific rule.

What is notably absent from DOJ's public-facing documents through October 2026 is any comprehensive guidance on how AI-generated drafts of legal memoranda or enforcement documents interact with the deliberative process privilege—the executive branch analogue to attorney-client privilege that protects pre-decisional, deliberative government communications. DOJ policy does not address whether an AI system's intermediate outputs during the drafting of a legal position constitute agency records subject to FOIA, or whether the use of a commercial AI vendor to process enforcement-related legal text affects the agency's ability to claim privilege over the underlying analysis.

GSA Procurement Standards and Vendor Assessment Frameworks

The General Services Administration has developed AI vendor assessment frameworks through its AI Center of Excellence and through the Federal Acquisition Regulation (FAR) supplement and related policy instruments. GSA's published AI procurement guidance emphasises that vendors supplying AI tools to federal agencies must comply with data handling requirements under the Federal Information Security Management Act (FISMA), satisfy FedRAMP authorisation where cloud services are involved, and adhere to OMB's minimum risk practices for rights- and safety-impacting use cases.

GSA's assessment frameworks for AI vendors focus on transparency of training data, explainability of model outputs, and the ability of agency personnel to audit decisions made with AI assistance. These are operationally coherent requirements for procurement AI, fraud detection, or workforce scheduling. Applied to legal AI tools used by agency counsel, they raise different questions. An explainability requirement for a tool that drafts a regulatory preamble or identifies analogous case law may be achievable. The same requirement applied to a generative AI system producing deliberative legal analysis—where the analytical path is iterative and probabilistic—interacts awkwardly with the privilege framework: a FOIA request for "explainability logs" generated by a legal AI tool could seek access to exactly the kind of pre-decisional reasoning the deliberative process privilege is designed to protect.

GSA procurement standards do not resolve this tension. They are silent on whether legal AI tools operated inside agency counsel offices require FedRAMP authorisation at the same level as operational agency systems, and they provide no guidance on contractual terms that federal agencies should require when commercial legal AI vendors process text that may contain privileged communications.

The Structural Gaps

First: The work product doctrine is not addressed. Federal AI policy documents treat AI outputs as government information system data. They do not engage with the question of whether AI-assisted legal analysis constitutes protected work product under Federal Rule of Civil Procedure 26(b)(3). When a DOJ litigator uses an AI tool to identify case weaknesses, draft a legal theory, or model litigation risk, the AI system's intermediate steps—cached queries, generated summaries, flagged risks—may exist as data on a government or vendor system. Whether those artifacts are protected as opinion work product, and whether production of technical access logs could constitute an implicit waiver, is a live question that current OMB and DOJ guidance does not answer.

Second: Vendor data handling creates unresolved privilege exposure. Commercial legal AI vendors typically require in their master service agreements that customers acknowledge model training exclusions, data retention limits, and logging practices. Federal agency procurement standards require data handling agreements under FISMA and FedRAMP frameworks, but they do not require vendors to contractually acknowledge that data submitted by agency counsel may be privileged and to treat it accordingly. This differs from what major commercial law firms have negotiated: law firm MSAs with legal AI vendors routinely include explicit acknowledgements that input data is confidential attorney-client or work product material, prohibitions on model training over client data, and specific breach notification timelines. GSA's procurement frameworks had not incorporated analogous legal-practice-specific terms through October 2026.

Third: Disclosure obligations may conflict with litigation strategy. M-24-10's transparency requirements—publishing AI use case inventories, disclosing minimum risk practices—were designed for public accountability. (Successor OMB M-25-21, in force since April 2025, retains AI use case inventory obligations.) When applied to the DOJ Civil Division's use of AI in active litigation, these transparency requirements create tension: disclosure that DOJ is using a specific AI system to conduct legal research or draft pleadings in a pending case could provide opposing counsel with information about litigation strategy. Agency AI use case inventories published under M-24-10 (and its successor M-25-21) had included some law enforcement and legal operations entries through 2025, with entries sometimes redacted for law enforcement sensitivity but without a consistent framework for treating legal AI use as categorically distinct.

What This Means in Practice

Federal agencies and DOJ litigators operating in this environment face a governance structure that was not designed for the intersection of AI and legal professional privilege. The responsible path—consistent with existing guidance and the professional obligations confirmed in ABA Formal Opinion 512 (issued 29 July 2024)—requires independent human verification of all AI-assisted legal submissions, conservative treatment of AI-generated legal analysis as potentially subject to FOIA unless covered by a recognised privilege, and explicit contractual protections for privileged material in any AI vendor agreement that federal procurement officers have not yet standardised. The policy documents speak clearly about operational AI. For legal work product, the silence is the story.