How Washington DC Buys AI

Executive orders and agency memos say how the government should use artificial intelligence. Contracts say how it actually does. A contract fixes what is being bought, from whom, for how much, and what the seller owes in return.

Since the start of fiscal year 2020, federal agencies have committed $2.44 billion in contracts whose descriptions mention artificial intelligence or machine learning. Annual spending has more than doubled since 2022.

The number of contracts peaked in 2024 at 542 and has fallen since, while the money kept rising. Agencies are making fewer, larger AI purchases.

The Pentagon buys most of it

The Department of Defense accounts for 70 percent of the spending. The Air Force and the Army alone account for more than half, about $1.33 billion between them. Every civilian agency combined spent less than half as much as the Pentagon.

The model makers barely appear

The companies that build frontier AI models are almost absent from these contracts. Among 1,685 AI contracts, none went directly to OpenAI, Anthropic or Google. Amazon Web Services and Microsoft appear a handful of times, for cloud-hosted AI tools at the Defense Information Systems Agency.

Instead the money flows to integrators and resellers. When the Treasury Department set up its ChatGPT licenses in September 2026, the contract went to Carahsoft, a government software reseller. The ten largest recipients took 28 percent of all AI spending.

How often anyone pays to check

A contract is the government's main chance to demand proof that an AI system works as promised. So I searched every contract description for the language of checking: testing and evaluation, verification, validation, assurance, red teaming, explainability, trustworthiness, bias, safety and risk management.

Only 85 of the 1,685 contracts use any of these words, about 1 in 20. Each square below is one contract.

Most of the checking that does get bought happens at the Pentagon, which holds 46 of the 85 contracts. The largest are Army contracts for testing and evaluating AI systems. Civilian agencies have bought a few programs of their own.

Generative AI arrives, quietly

Contracts that mention generative AI, large language models or chatbots first appear in 2023. There are 64 of them so far, worth about $57 million in total, a small share of the whole.

They show where agencies are trying the technology first: an Internal Revenue Service pilot of language models for taxpayer support, a chatbot for a Health Resources and Services Administration data warehouse, and an ambient medical scribe for Air Force readiness exams.

What this means

The government's AI spending is growing, concentrated in defense, and routed through intermediaries. Checking is rarely a line item. Only about 1 in 20 contracts mentions testing, evaluation or assurance at all, and most of those are at the Pentagon.

That matters because procurement is the one point where the government holds leverage over a seller. Requirements for evaluation, documentation and incident reporting can be written into a contract before it is signed. Once the system is deployed, that leverage is mostly gone.

A description is not the whole contract. Many agreements include testing requirements in documents that USAspending does not publish. But the descriptions are the public record of what the government says it is buying, and in that record, checking AI is rare.