Before Congress writes a single rule for artificial intelligence, it has already picked a supplier. In the year to the end of March, House offices spent nine of every ten visible AI dollars on one product: OpenAI’s ChatGPT.

The figures come from a CNBC analysis of House disbursement records. Of $113,740 spent on named AI vendors, ChatGPT took $100,580, about 88% of the money and 96% of the transactions. Anthropic’s Claude was a distant second at $13,160. ChatGPT now appears in at least 71 member offices, roughly one in six.

The work is exactly what you would expect. Staffers use the tools to summarise bills, draft memos, prepare hearing materials and answer constituents. Representative Julie Fedorchak, a North Dakota Republican, said a ChatGPT-built database of her legislative work was “saving dozens and dozens of hours every week” of staff time.

The skeptics are the biggest buyers

The politics are the awkward part. Democratic offices spent $54,165 on identifiable AI tools, more than three times the $15,782 from Republicans. ChatGPT turned up in 44 Democratic offices and 27 Republican ones. That is the same party whose members have warned loudest about AI’s risks to workers, privacy and elections, now leading the spending on it.

None of this is neutral ground. OpenAI, Anthropic, Google and Microsoft are racing for influence across Washington as lawmakers decide how AI gets bought, used and regulated. The head start was structural. In 2023 the House handed out its first 40 ChatGPT licences, and by the time restrictions arrived in 2024 the name had stuck.

“ChatGPT was almost synonymous with AI,” said Marci Harris of the POPVOX Foundation, which trains congressional staff. She compared it to Kleenex and tissues, or Band-Aid and bandages. OpenAI reinforced the lead. It co-led training sessions for senior staff through a nonpartisan Hill group. It also spent $1m lobbying Washington in the first quarter, up 82.5% on a year earlier.

Why the vendor matters

Claude’s absence has a colder explanation than brand recognition. Interest in rivals rose only after Anthropic’s fight with the Pentagon, which named the company a supply-chain risk in March. The Senate bars Claude for official use outright. When your model is blocked, you do not show up in the spending records.

That is why the vendor question is not trivial. Eric Petry of the Brennan Center said the goal, “not to reward political favorites,” should be giving government the best tools available. His worry is that offices are standardising on one model before lawmakers understand how any of them behave.

The caveats are real, and CNBC names them. The records miss free accounts, AI folded into broader software like Microsoft Copilot, reimbursed purchases and almost all of the Senate. The true figure is certainly higher, and $113,740 is rounding error in a federal budget. What the numbers capture is not scale but habit.

And habit is the point. The same institution that will write the rules is learning what AI is from one company’s product. It learns on that company’s training, too. Meanwhile that company keeps tripling down on lobbying. Regulators tend to keep the tools they already know. OpenAI appears to understand that better than anyone.

Get the TNW newsletter

Get the most important tech news in your inbox each week.