The US federal government adopted AI across 2025 and into 2026 along two tracks, with tensions emerging between the adoption push and existing institutional guardrails. The two tracks were a centralized procurement architecture run through the General Services Administration, and a set of agency-specific deployments at DHS, DOD, FDA, ICE, CBP, DOT, and DOGE. Oversight responses — congressional letters, state attorney-general inquiries, court challenges, and outside-group critiques — addressed security, accuracy, vendor lock-in, and the enforceability of vendor use policies against government pressure.
Strategic framework
America's AI Action Plan (July 2025) organized federal AI policy around three pillars: innovation, infrastructure, and international/security. Federal adoption is embedded throughout the plan rather than treated as a separate workstream.
Procurement architecture
The central procurement vehicle is GSA OneGov + the USAi Platform. In an August 2025 round of announcements, GSA introduced USAi.gov, a free evaluation platform for agencies, alongside $1-per-agency deals with major frontier vendors: OpenAI (ChatGPT), Anthropic (Claude for Government, at FedRAMP High), and Google (Gemini). The same effort included up to $1 billion in federal IT savings via AWS and a Box AI workflow-automation arrangement. Separately, Palantir's federal footprint expanded through a $10 billion Army arrangement in July 2025.
America's AI Action Plan also catalyzed GSA's Procurement Ecosystem Initiative (August 2025), a request for information (RFI) toward an AI-incorporated procurement ecosystem.
The two GSA vehicles are sequenced rather than parallel: agencies use USAi to test models before procurement, then obtain access to the chosen model at near-zero prices through OneGov deals (Source: fedscoop.com). GSA launched the OneGov strategy in April 2025 to offer agencies discounted rates on selected private-sector technology and software services (Source: nextgov.com). USAi followed on August 14, 2025, announced as a secure generative AI evaluation suite available to all federal agencies at no cost, offering chat-based AI, code generation and document summarization in what GSA described as a standards-aligned environment, together with dashboards and usage analytics for tracking performance and measuring adoption maturity. Deputy Administrator Stephen Ehikian, Federal Acquisition Service Commissioner Josh Gruenbaum and Chief Information Officer David Shive each framed the launch as implementing the administration's AI strategy; Gruenbaum described the platform as a secure environment for exploring models "prior to making procurement decisions" (Source: gsa.gov).
A third GSA lever operates on authorization rather than price. GSA's end-of-2025 account of its AI work describes FedRAMP 20x as eliminating the prior requirement that an AI cloud provider offering be sponsored by an agency before entering the FedRAMP authorization pipeline, alongside measures to allow widespread access to AI services across government and to streamline AI procurement (Source: gsa.gov).
Return on investment remains unquantified. Zach Whitman, GSA's chief data scientist and chief AI officer, said the agency has seen "a lot of success so far in terms of general adoption" while acknowledging it has yet to determine what leaders can expect for return on investment, describing the effort as a work in progress. USAi functions as a central hub where agencies can rerun model evaluations and view, through the provided dashboard, how a model performs on particular use cases and how likely it is to disclose information it should not (Source: fedscoop.com).
Measured adoption across agencies
The federal AI use case inventories, published annually by agencies, provide the only governmentwide count. Valerie Wirtschafter's April 2026 Brookings analysis of the 2023, 2024 and 2025 inventories records growth from 710 reported use cases in 2023 to more than 3,600 in 2025, with nine agencies reporting use cases for the first time in 2025 and 26 agencies increasing their reporting between 2024 and 2025 (Brookings — Assessing the state of AI adoption across the federal government (April 2026)).
Adoption is concentrated in large agencies. Those with more than 15,000 employees contributed 69 percent of all reported use cases in 2024 and 76 percent in 2025, while midsize (1,000–14,999 employees) and small agencies combined fell from 31 percent to 24 percent over the same period despite a roughly steady number of participating agencies. Averaged per agency in 2025, large agencies reported 211 use cases, midsize agencies 48, and small agencies five, against 114, 32 and four respectively in 2024. The eleven small agencies that reported in 2025 submitted 60 use cases in total, 2 percent of the inventory. Brookings attributes part of the divergence to differing agency missions and part to uneven capacity and resources (Brookings — Assessing the state of AI adoption across the federal government (April 2026)).
The composition of each agency's inventory tracks its mission. In 2025, 52 percent (17) of Social Security Administration use cases supported service delivery and benefits processing; 36 percent (86) of DHS and 54 percent (170) of Department of Justice use cases supported law enforcement; and 20 percent (89) of Department of Health and Human Services and 45 percent (166) of Department of Veterans Affairs use cases facilitated health and medical services (Brookings — Assessing the state of AI adoption across the federal government (April 2026)).
Reclassification of high-risk use cases
The Trump administration's OMB memorandum M-25-21, "Accelerating Federal Use of AI through Innovation, Governance, and Public Trust," replaced the prior "rights- and safety-impacting" category with "high-impact" use cases, defined as those used as "the principal basis" for an outcome. Of the more than 1,000 use cases Brookings could match across the 2024 and 2025 inventories by agency and name, 85 percent of those previously flagged as rights- or safety-impacting retained a high-impact designation, covering 175 use cases; 32 were downgraded, including nine Department of Justice law-enforcement use cases such as machine-learning triage of threats at FBI field offices, with the stated rationale in more than half the cases being that the system did not form the principal basis for a decision; and eight received a high-impact designation for the first time, among them use cases relating to health-care billing and diagnostic decisionmaking. More than 85 percent of all high-impact deployed use cases in 2025 lack some required information about risk-mitigation measures, despite OMB's requirements (Brookings — Assessing the state of AI adoption across the federal government (April 2026)).
Agency-specific deployments
Adoption beyond the central platform proceeded agency by agency:
- DHS is described as a leading law-enforcement and border AI use case (MIT Technology Review, January 2026) and maintains an AI Use Case Inventory Library.
- DOD runs Task Force Lima and an AI Rapid Capabilities Cell (AI RCC), a joint effort of the Chief Digital and Artificial Intelligence Office (CDAO) and the Defense Innovation Unit (DIU); it deploys Anthropic's Claude Gov, the subject of the dispute described below.
- FDA began an agency-wide AI rollout (June 2025) and ran an AI-Assisted Scientific Review pilot (May 2025).
- ICE uses AI from Palantir and OpenAI.
- CBP signed an Altana AI contract for Section 232 enforcement (October 31, 2025).
- DOT, under the Trump administration, plans to use Google's Gemini to write regulations (ProPublica, January 26, 2026).
- DOGE developed error-prone AI used to cancel VA contracts (ProPublica, June 2025).
- VA — congressional bills introduced during the week of July 24, 2026 would require measurement of AI efficacy at the Department of Veterans Affairs (Source: nextgov.com).
Tensions
Security
The Grassley letter documents that the CISA Acting Director uploaded at least 4 "for official use only" documents to the public version of ChatGPT. CISA is the agency responsible for federal cybersecurity. The episode illustrates a failure mode of the rapid-adoption push.
Accuracy and errors
A New York City AI chatbot told businesses to break the law and was set to be terminated by incoming Mayor Mamdani (The Markup, January 30, 2026). DOGE's error-prone AI at the VA (ProPublica, June 2025) is a second documented accuracy failure.
Corporate-capture concerns
Public Citizen (August 14, 2025) published "New Federal AI Platform Risks Corporate Capture, Worker Displacement, and Equity Blind Spots," arguing that $1 pricing binds agencies to vendors in unexamined ways.
Frontier-lab use-policy conflict
In the Anthropic–Pentagon dispute, the Pentagon's push for "all lawful use" of Claude Gov ran into Anthropic's refusal. Judge Lin's March 26, 2026 preliminary injunction blocked the resulting retaliatory ban and a supply-chain-risk designation. It is the first federal court ruling that the government cannot retaliate against AI vendors for stated use restrictions.
Evaluation standards
Through 2025 and into 2026 the adoption push ran ahead of any governmentwide method for evaluating what agencies were buying, with USAi supplying per-agency dashboards rather than a shared standard (Source: fedscoop.com). NIST's initial public draft of AI 200-2, the TEVV-Athlon framework, issued August 7, 2026 with a comment period closing October 6, 2026, is directed at that gap (NIST AI 200-2 ipd — The TEVV-Athlon Framework for Evaluating AI Systems (initial public draft, August 2026)); Ike Harris of the Frontier Security Institute described it as "the first step in standardizing the way the federal government evaluates AI systems both for itself and for its contractors" (Source: communicationstoday.co.in; see NIST AI 200-2 — The TEVV-Athlon Framework for Evaluating AI Systems).
Contract loss after vendor failure
After Grok's antisemitic outbursts and the "MechaHitler" incident, xAI reportedly lost a major government contract (WIRED, August 14, 2025). The episode, which surfaced in the same August 14, 2025 reporting window as the Public Citizen critique, runs counter to the corporate-capture concern, in that a demonstrable vendor failure produced a contract loss.
Canonical cases
| Vendor | Agency | Outcome |
|---|---|---|
| Anthropic | DOD (Claude Gov) | Court ruled against government ban; contract protected |
| OpenAI | ICE + broad federal OneGov | Deployed, controversial |
| DOT (writing regulations) | Deployed, controversial | |
| Palantir | ICE + Army ($10B, July 2025) | Entrenched |
| xAI | (Various) | Lost after Grok incidents |
| Altana | CBP | Active 2-year contract |
| Box | Multiple (workflow) | OneGov active |
Relationships
- consolidates: GSA OneGov Program and USAi Platform (August 2025), Government AI Procurement.
- depends-on: Brookings — Assessing the state of AI adoption across the federal government (April 2026) — the governmentwide inventory analysis behind the measured-adoption section; America's AI Action Plan, Sen. Grassley Letter to CISA re ChatGPT (2026-02-05), Anthropic Claude Gov + Pentagon Dispute (2025–2026), Anthropic v. United States (Pentagon ban challenge).
- deploys-in: DHS — Department of Homeland Security (AI Deployer), DOD — Department of Defense (AI Deployer), FDA — Food and Drug Administration (AI Deployer), ICE — Immigration and Customs Enforcement (AI Deployer), CBP — Customs and Border Protection (AI Deployer), DOT — Department of Transportation (AI Deployer), DOGE — Department of Government Efficiency (AI Deployer), GSA — General Services Administration (AI Deployer).
- related: OMB Memorandum M-24-10, OMB Memorandum M-24-18, CSET + DOJ + CSIS — US-China AI National-Security Axis (composite source summary) (DoD AI RCC context), NIST AI 200-2 — The TEVV-Athlon Framework for Evaluating AI Systems, NIST AI 200-2 ipd — The TEVV-Athlon Framework for Evaluating AI Systems (initial public draft, August 2026), OMB M-25-21 — Accelerating Federal Use of AI through Innovation, Governance, and Public Trust, AI Pre-Release Vetting.