A CNBC analysis of House disbursement records landed this morning, and the number that jumped out at me is not the big one. It is how small the whole thing is.

Between April 1, 2025 and March 31, 2026, the House of Representatives — member offices, committees, institutional accounts — spent at least $113,740 on identifiable AI tools. ChatGPT accounted for $100,580 of that across 798 transactions. Anthropic's Claude was a distant second at $13,160 across 37 transactions. CNBC put OpenAI's share at about 88% of the dollars and 96% of the transactions; TechCrunch, writing it up from the same records, rounded it to roughly 90%.

A hundred and thirteen thousand dollars. That is one mid-level staffer's salary, spread across an entire chamber of government, for a full year.

What the records actually cover

Before anybody builds a theory on this, the caveats are big and CNBC is upfront about them. The disbursement records exclude free accounts. They exclude AI bundled into broader software contracts — which is most enterprise AI now, since it is baked into Microsoft 365, into Google Workspace, into every CRM. And they exclude most Senate use entirely.

So this is not a measure of how much AI Congress uses. It is a measure of how much AI Congress buys on a line item you can see. Those are very different numbers, and the second one is almost certainly a small fraction of the first.

What it does tell you cleanly: when a House office decides to pay for a chatbot out of its own budget, it overwhelmingly pays OpenAI. ChatGPT purchases showed up in at least 71 House member offices, roughly one in six.

Verdict: small dollars, real signal about default choice.

The party split is the interesting line

Democratic offices accounted for $54,165 in identifiable AI purchases. Republican offices spent $15,782. More than three to one.

I would not read ideology into that too fast. Office budgets are discretionary and small, staff sizes vary, and one enthusiastic chief of staff can move a district's number. But three-to-one across a chamber is not one office. Something structural is going on with how the two caucuses are approaching staff tooling, and I have not seen a satisfying explanation for it yet.

What the staff are doing with it is unremarkable and, honestly, sensible: summarizing legislation, drafting memos, preparing hearing materials, responding to constituents, sorting policy research. CNBC quotes Rep. Julie Fedorchak, R-N.D., saying her staff used to spend hours writing the prep memos for a dozen-plus meetings a day. That is exactly the kind of work a language model is genuinely good at — compression of documents somebody already has.

Verdict: this is document triage, not policy-by-robot.

Why a rounding error still matters

Here is the part I actually care about. The same institution using these tools is the one deciding how AI gets purchased, used, and regulated across the country. OpenAI, Anthropic, Google, and Microsoft are all working Washington, and there is a specific kind of advantage in being the tool that a legislative aide opens at 9am to make sense of a 400-page bill.

That is not a conspiracy claim. Nobody is being bribed with a $20 seat. It is a much more ordinary thing: whichever tool people use every day becomes the tool they picture when they write a rule. If your mental model of "AI" is a chat box that summarizes documents, you will write different legislation than someone whose mental model is an autonomous agent holding live credentials to a production system. Those are both AI in 2026, and only one of them is on the average Hill desk.

That gap worries me more than the spending totals. The risks I spend my time on — an agent with a token it should not have, a tool call nobody logged, a model acting on a system of record — are not visible from inside a chat window. If you want the concrete version of that, the checks in the MCP server security checklist are the sort of thing that does not occur to you until you have wired an agent into something real.

The vendor-concentration problem

Eighty-eight percent to one vendor is a concentration number I would flag in any organization, not just this one. If a private company told me a single unreviewed SaaS vendor held 88% of a category's spend and 96% of its transactions, I would ask who signed off, what the data handling terms are, and what happens when the price changes.

For congressional offices, the data question is the sharp one. Constituent correspondence, draft legislation, unreleased hearing prep — that is sensitive material by any normal standard, and it is being pasted into a third-party service under whatever terms the office clicked through. The records do not show what tier of account these are or what the retention settings look like. Nobody has published that, so I am not going to assert a problem exists. I am saying it is the question I would ask, and I have not seen it answered.

The broader pattern here is the same one showing up everywhere: adoption ran ahead of procurement discipline, because a $20 seat clears every approval threshold there is. It is the same dynamic behind the breach wave hitting mid-sized businesses — not a dramatic failure, just tools arriving faster than anyone's process for reviewing them.

What I take from it

Three things. Congress is using AI for the boring, useful stuff, which is fine. The visible spending is tiny, which means the real usage is invisible and bundled, which means nobody has a full picture including the people inside the building. And one vendor owns the mental model of what AI is, at the exact moment the rules for AI are being written.

None of that is scandal. It is just worth knowing when the next AI bill shows up and reads like it was written by someone who has only ever used a chat box.

Sources: CNBC — ChatGPT dominates early AI spending in Congress as lawmakers weigh regulation · TechCrunch — Congress's favorite AI tool? ChatGPT