Federal AI spending in the United States reached $7.2 billion in obligated contract funds in 2026, up 966% from $675 million in 2024. The potential value of awarded AI contracts rose even faster, from $4.6 billion to $91.8 billion over the same period, according to research from Brookings and the IBM Center for The Business of Government.
Those figures describe a market that has changed shape entirely in two years. They do not describe most agencies.
Why does federal AI spending look different from inside a civilian agency?
The Department of Defense accounted for 98.9% of all potential federal AI contract value in 2026, worth $90.7 billion, and held 1,319 of the 1,743 AI contracts in force. Every other department shared what was left. The Department of Commerce recorded $197 million, Health and Human Services $138 million, NASA $45 million, the Department of the Interior $1.5 million, and the Nuclear Regulatory Commission $890,000.
The number of agencies holding any AI contract at all rose from 17 in 2022 to 23 in 2024 and 28 in 2026, out of 441 federal agencies in total.
So a benefits administrator, a tax authority, or a permitting office reading that federal AI investment has grown tenfold is reading a defense procurement story. The budget headline and the operational reality sit a long way apart, which makes the headline a poor basis for judging what is available to a civilian agency or what it can realistically build.
What are federal agencies actually using AI for in 2026?
Adoption has grown much faster than civilian spending. The federal AI use case inventory expanded from 710 documented use cases in 2023 to more than 3,600 in 2025, covering fraud detection, case prioritization, border inspections, document summarization, and vendor research alongside general productivity work.
That gap between spending and adoption is the most useful signal in the data. Agencies are putting AI into daily operations without large contract vehicles behind them, which means the work is happening inside existing systems, existing teams, and existing budgets.
What changed in federal AI policy since 2025?
Three things reset the operating conditions for civilian agencies.
The General Services Administration launched USAi in August 2025, a shared generative AI evaluation platform that gives federal agencies access to leading commercial models at no cost to them, inside an environment built to federal security standards. FedRAMP prioritized authorization for conversational AI engines intended for routine use by federal workers, with its first authorizations under the 20x program targeted for January 2026. In March 2026, GSA partnered with NIST and CAISI to build consistent methods for testing and measuring AI systems before agencies deploy them in real settings.
The practical effect is that model access, security authorization, and evaluation tooling are no longer the first obstacles an agency hits. An agency can now try a model this quarter without a procurement cycle.
What determines whether a government AI pilot reaches production?
It comes down to the process the model sits inside.
A model can read a claim, summarize a case file, or flag an anomaly. What it cannot do on its own is route the exception to the right reviewer, hold a determination open while a document is chased, enforce who is allowed to approve what, or produce the audit trail that explains a decision two years after it was made. Those are orchestration problems, and they exist whether or not there is any AI in the process at all.
Brookings points to a shortage of AI-specialized talent, procurement and regulatory friction, and public skepticism about the technology as the recurring constraints on federal adoption. Each of those is a question about how work moves through an organization rather than a question about model quality.
Public sector systems are built to be accountable first, and accountability is expensive to change quickly. Agencies also experiment carefully, which is the rational response when a wrong output carries legal consequences for a citizen. The agencies making the most progress recognized both of these early and designed the surrounding process deliberately, with the model as one participant in it.
How does BP3 help government organizations move from pilot to production?
BP3 works with public sector organizations on the process layer that determines whether an AI pilot becomes an operational service.
When a government tax department's Lotus-based case management platform reached end of life, as every platform eventually does, the department engaged BP3 to design, build, and maintain a replacement case management application on IBM Business Automation Workflow. The technology choice was the straightforward part. The work was in modeling how cases actually move, who owns each decision, and what has to be true before a case can close.
BP3 supports ministries of justice, federal agencies, and other public sector organizations across social welfare programs, tax management, and healthcare services, applying process orchestration at scale so that automation and AI operate inside a defined, auditable flow rather than beside it.
Agencies that build that layer once can add AI to it repeatedly. Agencies that handle governance model by model rebuild the same work every time.
See how BP3 approaches process orchestration for the public sector, including the systems, standards, and delivery model behind it.