"Assessing the state of AI adoption across the federal government" is a report by Valerie Wirtschafter, a Fellow in Brookings's Foreign Policy program and its Artificial Intelligence and Emerging Technology Initiative, published on April 15, 2026 and running to a stated 47 minutes' reading time. It is the most detailed governmentwide quantification of US federal AI use located to date, and is the underlying source for the measured-adoption figures on Federal AI Adoption — Patterns and Tensions.
The report draws on five evidence bases: the federal AI use case inventories for 2023, 2024 and 2025; USAJobs historic job-announcement data; OMB memoranda from the Biden and second Trump administrations; submissions to the Office of Science and Technology Policy's Request for Information on Regulatory Reform on Artificial Intelligence (docket OSTP-TECH-2025-0067); and interviews with current and former federal technologists across eight agencies.
Its summary finding is that adoption "accelerated significantly over the past three years" while remaining "concentrated among a handful of large agencies," slowed by "workforce capacity constraints, a risk-averse culture, procurement and funding challenges, and low public trust in AI systems."
Measured adoption
| Inventory year | Reporting agencies | Total use cases |
|---|---|---|
| 2023 | 21 (13 large, 8 midsize, 0 small) | 710 |
| 2024 | — | 2,133 |
| 2025 | 41 (13 large, 17 midsize, 11 small) | 3,611 |
The 2025 total is stated as "more than 3,600 individual use cases, 69% above the total number reported in 2024 and five times the number reported in 2023." Nine agencies reported for the first time in 2025 and 26 increased their reporting from 2024.
Concentration is the report's central structural finding. For three consecutive years, five large agencies accounted for over half of all reported use cases. Large agencies — those with more than 15,000 employees under the Partnership for Public Service's classification — contributed 69% of reported use cases in 2024 and 76% in 2025, while midsize (1,000–14,999) and small (under 1,000) agencies combined fell from 31% to 24%. Per-agency averages in 2025 were 211 for large agencies, 48 for midsize and five for small, against 114, 32 and four respectively in 2024. The eleven small agencies reporting in 2025 submitted 60 use cases between them, 2% of the inventory.
Mission-critical rather than back-office application is documented by agency: 52% (17) of Social Security Administration use cases support service delivery and benefits processing; 36% (86) of Department of Homeland Security and 54% (170) of Department of Justice inventories support law enforcement; 20% (89) of Health and Human Services and 45% (166) of Veterans Affairs inventories facilitate health and medical services.
Roughly 60% of use cases with deployment information available are in the pilot or pre-deployment stage. Of the more than 1,600 use cases that state how they were developed, about 63% were built by contractors exclusively or alongside in-house resources; among fully deployed use cases contracting is involved in 72% against 28% purely in-house, while pre-deployment work splits roughly evenly. The report offers two readings of that imbalance — greater in-house prototyping capacity, or difficulty scaling without vendors — and does not choose between them.
The M-25-21 reclassification
The 2024 inventories, under OMB M-24-10, required agencies to flag "rights- or safety-impacting" use cases that "control or significantly influence" outcomes. OMB M-25-21, "Accelerating Federal Use of AI through Innovation, Governance, and Public Trust," replaced that with "high-impact" use cases that are "the principal basis" for outcomes. Rights- or safety-impacting cases were 16.5% of the 2024 inventory (351 use cases); high-impact cases were 12.3% of the 2025 inventory (445). The share fell while the absolute number rose.
The report is explicit that the comparison is limited by data structure rather than analysis: "the lack of a common use case ID in some cases makes it difficult to compare whether agencies have simply reclassified their rights- and safety-impacting use cases as high impact, or if they have taken a different approach." Of the more than 1,000 use cases matchable across the two inventories by agency and name, 85% of those previously flagged retained a high-impact designation (175 use cases). Thirty-two cases were downgraded, including nine Department of Justice law-enforcement use cases — among them machine-learning triage of threats at FBI field offices and investigative support — with the stated rationale in more than half of all downgrades being that the system did not form "the principal basis" for a decision. Eight moved the other way, gaining a high-impact designation for the first time, including health-care billing and diagnostic decisionmaking.
On compliance with those requirements the report finds: "More than 85% of all high-impact deployed use cases in 2025 lack some required information about risk mitigation measures in place, despite explicit requirements from OMB." Where a use case meets the high-impact classification, agencies are required to supply information on pre-deployment testing, impact assessments, monitoring or appeal processes.
Stated limitations of the inventory data
The report sets out four limitations before drawing on the inventories: they are self-reported and "some offices are unsure of what to report and in how much detail"; they "are never fully consistent, making comparison and standardization difficult across agencies and over time"; some agencies' use cases lack unique identifiers permitting tracking across years, and the consolidated OMB lists for 2023 and 2024 exclude those identifiers; and the inventories "do not account for any efforts to leverage AI deployed through DOGE." It also attributes part of the measured growth to clarified OMB reporting guidance in 2024 and 2025 rather than to expanded use alone.
Bottlenecks
Talent. Federal technical job listings referencing explicit AI capabilities rose from zero in 2016 to about 8% of technical postings in 2024, with posts nearly doubling from 184 in 2021 to 318 in 2024 — under 3% of all technical listings on the report's dictionary-based measure. Nearly 33% of AI-specific posts used expedited hiring pathways against 17% for other technical jobs. Since the start of 2025 the federal government has posted only 160 full-time listings explicitly requiring AI expertise, and the report notes that some postings designed to accelerate AI adoption "make no mention of AI or technology." It reproduces the April 2024 AI and Tech Talent Task Force figures — 172+ hires between October 30, 2023 and March 31, 2024, 72+ planned through August 2024, 500+ planned through September 2025, with the Department of Defense separately planning over 2,500 AI and AI-enabling hires by FY24 and over 9,000 by FY25 — and observes that 2025 workforce reductions may have fallen disproportionately on this cohort, since probationary employees hired under the talent surge were more easily dismissible and probationary periods can run up to three years.
Risk-averse culture. Technologists who scaled pilots emphasised having "explicit space to experiment" from a supportive supervisor, and the report notes that White House directives under both administrations supplied cover. Where leadership buy-in is absent — through competing funding priorities or unfamiliarity with the technology — "even talented technologists may default to safer, more conventional approaches." It adds that the administration's explicit linkage of AI deployment to workforce cuts through DOGE "may also reinforce this hesitancy."
Accountability, funding and regulation. AI systems can introduce "black box problems," fracturing the step-by-step traceability that earlier technologies allowed and complicating audits of adverse decisions. Capability turnover outpaces the planning horizon: "a pilot program launched today may be obsolete—or superseded by dramatically more capable systems—before it reaches scale." The federal budgeting process, which begins roughly eighteen months before the fiscal year, requires agencies to forecast AI capability on a timescale the report treats as impracticable; it points to the Technology Modernization Fund — $250 million authorised at its 2017 inception plus $1 billion from the American Rescue Plan, funding more than 60 projects across 34 agencies — as the flexible alternative, while noting criticism of its documentation-heavy application process.
Public trust. The report situates deployment risk against 16% of Americans saying they trust Washington to do what is right most or almost all of the time, and cites the Dutch childcare-benefits scandal — which contributed to the government's collapse — and Australia's Robodebt as the cases where algorithmic administration produced financial and psychological harm. It also cites OECD survey work finding that satisfaction with public-service delivery drives institutional trust, and concludes that "well-executed AI deployments focused on tangible service improvements could help rebuild confidence."
Recommendations
The report's stated recommendations are to expand support for technical talent and AI literacy across agencies; to continue addressing structural barriers in procurement, regulation and budgeting that hinder technology modernization generally; and to foster public trust "through stronger transparency practices, improved use case inventories, and a focus on high-impact, positive applications." It credits FedRAMP 20x and USAi with reducing authorization barriers and supplying a common experimentation platform.
Provenance and a correction to prior sourcing
Retrieved on August 9, 2026 from brookings.edu/articles/, the institution's own publication path rather than an aggregator. Charts are served from Datawrapper under a Brookings-themed template with "Get the data" links, and the footnote apparatus is present and numbered.
Two facts recorded as unresolved when this report was queued for ingest are settled by the page itself. The publication date is April 15, 2026, stated under the byline. The author is Valerie Wirtschafter alone; the multi-name list previously recorded as a collapsed byline (Sorelle Friedler, Cameron F. Kerry, Tom Wheeler, Niam Yaraghi and others) belongs to a different Brookings item linked in the same page's related-content rail, "What to make of the Trump administration's AI Action Plan," dated July 31, 2025.
The April 15, 2026 date bounds the freshness of every figure here: the inventory counts are drawn from the 2025 inventory and the hiring figures run to early 2026.
Relationships
- supports: Federal AI Adoption — Patterns and Tensions — the source of that page's measured-adoption section.
- related: OMB Memorandum M-24-10, OMB Memorandum M-24-18, OMB M-25-21 — Accelerating Federal Use of AI through Innovation, Governance, and Public Trust — the memoranda whose reporting requirements the inventories implement, and the one that redefined the flagged category.
- related: Government AI Procurement — the contractor-versus-in-house split and TMF discussion.
- related: GSA — General Services Administration (AI Deployer), DHS — Department of Homeland Security (AI Deployer), DOD — Department of Defense (AI Deployer), DOGE — Department of Government Efficiency (AI Deployer) — agencies whose inventories or programmes the report quantifies.
- related: Public Opinion on AI — the trust findings the report treats as an adoption constraint.
- related: 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) — the evaluation-method gap the report documents and NIST's draft answer to it.