Executive Summary

Key Numbers

Source: CNBC analysis of House disbursement records, April 2025 to March 2026

CNBC's window runs from April 1, 2025 to March 31, 2026. Pulling out only the AI spending with a vendor named on the line brought the total to $113,740. ChatGPT accounted for $100,580 of it across 798 transactions, and Anthropic's Claude for $13,160 across 37. In dollars the ratio is close to eight to one; in number of charges it passes twenty to one. ChatGPT looks like many offices each paying for their own subscription, Claude like a handful of offices paying in bulk.

The split ran along party lines as well. Forty-four Democratic offices spent $54,165 and twenty-seven Republican offices spent $15,782, a gap of more than three times. Which side speaks more loudly about regulating AI and which side pays for more of it do not always point the same way.

Add the two numbers and 71 offices left their name in the spending record. With 435 members in the House, that puts roughly one office in six paying for AI tools out of its own budget. Because free accounts and AI bundled into office software fall outside the count, this is a floor. More offices than that have already worked AI into their routines, and how many more is not something this method can tell you.

The roots of the concentration sit in 2023. The House Digital Service distributed 40 ChatGPT Plus licenses free of charge to a bipartisan group of staffers. OpenAI, working with the Congressional Management Foundation, ran training sessions for senior staff in both chambers. Free licenses put the tool in people's hands, and the familiar tool became the one their offices kept paying for. The 88 percent today is less a share that appeared suddenly than a delayed invoice for that distribution three years ago.

Over the same stretch, both companies were also spending toward the people who write the rules. By Issue One's analysis, OpenAI spent $1 million on federal lobbying in the first quarter of 2026 and Anthropic about $1.6 million, roughly 1.8 times and more than three times their spending in the same quarter a year earlier. An industry under active regulatory debate is also supplying tools to the offices designing that regulation.

The spending record only says which office paid how much. What the money bought became known through staffers themselves. Gathered together, the uses read at first glance like a list of ordinary productivity tasks.

• • Summarizing bills that run hundreds of pages and pulling out the contested provisions
• • Preparing questions and background memos to carry into a hearing
• • Reviewing policy research and organizing the arguments
• • Drafting replies to constituent mail from the district
• • Drafting social media posts that go out under the member's name

Scheduling a meeting or tightening an email leaves nothing behind that matters, whichever tool does it. These five are different. A bill summary feeds a member's vote. A constituent reply arrives as the government's official answer to a citizen. Hearing prep becomes the question put to a witness. The finished product goes out under a person's name, and nothing in the document marks which passage came from the model's summary.

The input side works the same way. To get a summary, the text of the bill has to go into the window; to draft a reply, so does whatever a constituent wrote in. How much of that is public document and how much is private correspondence, and whether anyone records the difference, is decided office by office. Eric Petry of the Brennan Center for Justice argues that lawmakers and staff need hands-on time with a range of AI products before they design rules for them. The point is fair, but what that hands-on time should leave behind is the part the argument does not address.

The House publishes each office's disbursements every quarter. Vendor names, amounts, and dates land in the accounting system, so a reporter could scrape them and count. That is where this statistic came from, and for the same reason the charge is the only thing there is to count.

When a staffer opens a ChatGPT window, what the system retains is that a subscription fee went out. Which bill went into the window, what the model handed back, how much of that answer survived into the final text: none of it is an accounting line, so none of it has a place to land. Split one such session into what it leaves behind and what it does not, and it looks like this.

That free accounts sit in the right-hand panel is the heart of this structure. No charge means no accounting entry, and no accounting entry means that use disappears from the statistics. It is the same reason CNBC's count could not capture AI bundled into office software or most Senate spending. Before the 88 percent says anything about vendor share, it says where the line between the countable and the uncountable has been drawn.

That line is about to blur further. In August 2025 the General Services Administration agreed through its OneGov program to supply ChatGPT Enterprise to participating federal agencies for $1 per year. Anthropic matched the terms and Google came in at 47 cents. More than two million federal employees can now reach frontier models at what amounts to no cost, and a charge of $1 means the signal left in the accounting system is worth $1 too.

One commentary in Federal News Network reads these deals as the doorway to vendor lock-in, citing the Department of Veterans Affairs enterprise agreement with Microsoft signed in April 2025 that has grown to roughly $930 million a year over three decades of accumulation. The more symbolic the unit price, the faster the adoption, and the faster the adoption, the more expensive switching becomes later. The House's 88 percent is a point on the early part of that curve.

The legal status is not blank. When the Department of Homeland Security approved generative AI for use, it revised its OpenAI terms of service to establish that AI-generated output constitutes a federal record and is subject to disclosure under the Freedom of Information Act. The ruling, in other words, is on the books.

The trouble is that whether anything can be retained depends on the type of account. Agency accounts offer a way to export and store conversation logs. When an employee uses the free tier on a personal account, the record lives only in that personal history and the browser cache. Public records accumulate in a place the agency cannot reach. It is the same reason free accounts sat outside both the count and the retention side of the diagram above.

At the state level this has already shown up in practice. In 2025 several cities in Washington State produced ChatGPT and Copilot conversation logs in response to public records requests from a reporter. That is evidence the thing is technically possible, and equally evidence that it surfaces only once a request arrives. Retaining records as standard procedure and assembling them after the fact are different activities.

A federal standard is not yet in sight. As far as public materials show, the National Archives and Records Administration has not issued a retention schedule specific to AI-generated records. Records management practitioners also note that existing schedules are built around business context rather than file format, which makes them hard to lay directly over AI output. The tool costs $1 to bring in, and how many years the records it produces must be kept has not been settled.

Seen from the data side, this case comes down to one point. Procurement statistics are being used as a proxy for data lineage. Nobody counts prompts, input documents, outputs, or how much of the output was adopted, so money is the only countable thing, and money is what reporters and researchers count. A proxy standing in for the real measure is a signal that the real measure does not exist.

A judgment that cannot be retraced cannot be audited or revisited either. Suppose someone asks two years from now why a particular paragraph of a bill analysis was framed the way it was. What that question needs is which model at which version received which input, what it returned, and how much of it survived into the final text. Without those coordinates, the only available answer is that the office happened to hold a ChatGPT subscription at the time.

Two things are checkable right now, whether in a public agency or a company. Both are questions that should be answered before a new tool comes in, and if adoption keeps outrunning the answers, the records pile up empty.

• • Does the document itself show where AI touched the work? A record that a tool was adopted and a record that a given paragraph came from a given input are not the same thing. Without the second one, there is no starting point for retracing how the output was reached.
• • Does that record stay with the organization regardless of account type? Use that leaks out through personal accounts and free tiers is caught by neither procurement nor retention. What has to be decided before any adoption policy is which paths of use count as organizational records.

What the House numbers tell us is not which vendor won. It is how far the rules for recording what passed through a channel have fallen behind the speed at which text close to public judgment is moving into that channel. Adoption is tallied quarterly, and the retention schedule has yet to arrive.

Primary Reporting

• 1.TechCrunch. (2026-08-03). "Congress' favorite AI tool? ChatGPT." techcrunch.com
• 2.CNBC. (2026-08-03). "ChatGPT dominates early AI spending in Congress as lawmakers weigh regulation." cnbc.com

Procurement & Policy

• 3.U.S. General Services Administration. (2025-08-06). "GSA Announces New Partnership with OpenAI, Delivering Deep Discount to ChatGPT Gov-Wide Through MAS." gsa.gov
• 4.Federal News Network. (2026-06). "The coming AI reckoning: Slouching toward vendor lock." federalnewsnetwork.com
• 5.Issue One. (2026). "Big Tech Spends Millions to Buy Influence in Washington in First Half of 2026." issueone.org

Records Management & Governance

• 6.FedScoop. (2024). "Generative AI could raise questions for federal records laws." fedscoop.com
• 7.Cascade PBS. (2025-08-31). "Washington city officials are using ChatGPT for government work." cascadepbs.org