· 10 min read

Congress Spent $100K on ChatGPT and $13K on Claude. That 8-to-1 Split Is a Lesson for Anyone Selling AI to Institutions

AI Building in Public Lessons

CNBC pulled House disbursement records for the year ending March 31 and found something that should worry anyone pitching AI tools to institutional buyers: ChatGPT captured about $100,580 across 798 transactions, Claude captured $13,160 across 37 transactions, and that's out of at least $113,740 in total identifiable AI spending across House offices, committees, and institutional accounts. OpenAI took roughly 90% of the dollars. Not because ChatGPT tested better on some staffer's eval suite. Because it's the name everyone already knew.

If your business is consulting, integrating, or building tools for government, universities, hospitals, or any other buyer that answers to a budget committee, that ratio is worth sitting with for a minute, because it's not really a story about model quality. It's a story about how institutions actually buy software, and it applies just as much to your next SOW as it does to a congressional office's OpenAI subscription.

The numbers, and what they don't tell you

TechCrunch, reporting on the same CNBC analysis, put it plainly: OpenAI received roughly 90% of all spending on AI tools by House offices, committees, and institutional accounts during the year ending March 31. Do the actual division on the confirmed figures and it comes out to 88.4% of the dollars ($100,580 of $113,740), which is close enough to "nearly 90%" that I'm comfortable using CNBC's own framing. Claude was a distant second. Every other vendor combined didn't clear the noise floor.

Transaction count tells the same story from a different angle. 798 ChatGPT purchases against 37 for Claude is a 21-to-1 gap, wider than the dollar gap. That matters because it means ChatGPT wasn't winning a handful of big enterprise deals while Claude picked up smaller ones. ChatGPT was the default that individual offices reached for over and over, in small transactions, the way you'd expect a habit to show up in spending data rather than a procurement decision.

One caveat worth stating clearly: this is House-side spending only, and it excludes free-tier usage and AI bundled into broader software contracts (think Microsoft Copilot riding inside an existing Office 365 deal). The real gap between how much Congress uses ChatGPT versus Claude is probably smaller than the paid-spend numbers alone suggest, since bundled and free usage doesn't show up as a line item at all. But the paid-transaction data is the closest thing to a revealed-preference signal we have, and revealed preference is what you're actually selling against.

Why this isn't a benchmarks story

Here's the part that should sting a little if you've spent any time arguing about which model is "better." Congressional staffers buying AI tools for memo drafting, bill summarization, hearing prep, and constituent replies are not running comparative evals before they expense a subscription. They're buying the name they already have a mental model for: the one their kid uses for homework, the one that showed up in a 60 Minutes segment, the one procurement defaulted to because someone two years ago picked it and nobody's revisited the decision since.

That's not a knock on the staffers. It's a completely rational way to buy software when the downside of a wrong pick (a public data mishap, a bad headline, a colleague asking "wait, why are we using some AI startup nobody's heard of") outweighs the upside of picking the objectively stronger tool. Anthropic's Claude models score competitively, and by plenty of measures better, on the kinds of reasoning and writing benchmarks that would actually matter for legislative research and drafting. None of that shows up in an 8-to-1 spending ratio, because the buying decision isn't being made on the benchmark axis at all. It's being made on the "nobody gets fired for this" axis, and ChatGPT currently owns that axis in a way no amount of eval-table superiority undoes.

I've watched a version of this play out with clients who are nowhere near the scale of a congressional office. A five-person nonprofit board will greenlight "add ChatGPT to our workflow" in one meeting and spend three months on a subcommittee to evaluate "an AI vendor" if the vendor isn't a name they've heard on the news. The size of the institution changes, the psychology doesn't.

The partisan split is a smaller story, but it's real

CNBC's data also showed Democratic House offices spent roughly $54,165 on AI tools, about 3.4 times the $15,782 spent by Republican offices. That's worth a sentence, not a section. It's a real gap and probably reflects staffing differences and adoption timing more than anything ideological about the tools themselves, and I'd be cautious reading much more into it than "some offices adopted these tools faster than others." The headline number, the 8-to-1 vendor split, is the one that generalizes to your business. The partisan number mostly doesn't.

What this means if you sell into institutions

If your work involves pitching AI consulting, integration, or tooling into government, universities, healthcare systems, or any other buyer with a compliance department and a risk-averse procurement process, the incumbent-recognition tax is not a temporary market inefficiency you can out-argue. It's the actual shape of the market. A better benchmark deck doesn't fix it, because the buyer isn't evaluating on that axis. Neither does a lower price, because price isn't the objection either. The objection is "I don't know what I don't know about this vendor, and picking the familiar one costs me nothing politically."

The practical move isn't to compete head-on for the first purchase decision. It's to stop asking the institution to bet on you instead of the incumbent, and instead position your work as compatible with whatever they already run. "Works alongside your existing ChatGPT deployment," "integrates with what you already have," "adds the capability your current tool is missing" reads completely differently to a risk-averse buyer than "switch to this instead." You're not asking them to defend a new vendor decision to their boss. You're asking them to approve an addition to a decision they already made and already feel fine about.

What I'd actually do

If I were pitching AI work into an institutional buyer this quarter, I would stop leading with capability comparisons entirely. I've made that mistake myself, walking into a pitch with a slide that says "here's why our stack outperforms the obvious choice," and watching the room's attention drop the second it registers as a bet on an unfamiliar name. It's a losing frame no matter how true the numbers are.

Instead, I'd lead with the integration story: here's what breaks or is missing in your current ChatGPT (or Copilot, or whatever they already run) workflow, and here's how we plug that gap without asking you to rip anything out. That framing lets the benchmark case do quiet supporting work later, once trust exists, instead of carrying the whole pitch upfront. It's a less satisfying way to sell if you genuinely believe your tool is better, and often it is. But "better" isn't what wins institutional budget. Familiar, low-risk, and additive is what wins it, and the Congress numbers are about as clean a confirmation of that as I've seen.

Author

Lukas

Sources

More AI articles →

Stay in the Loop

Get new posts delivered to your inbox. No spam, unsubscribe anytime.

Newsletter coming soon. Set PUBLIC_CONVERTKIT_FORM_ID in .env to activate.

Related Posts