Anthropic Picked Its Database by Asking Claude

Three people were on Anthropic’s observability team when Claude 3.5 shipped and the database underneath it fell over. The engineer who had to fix that asked Claude which database to use instead, and Anthropic ended up on the one Claude named. If you sell to technical buyers, that’s what a shortlist looks like now: put together out of public technical writing, before anyone at the winning vendor knew there was a deal.

Observability is the software that tells an engineering team what its own systems are doing, and at Anthropic’s scale it generates an enormous amount of data.

Aaron Katz, Co-Founder & CEO of ClickHouse, Inc., told the story on 20VC in an episode published August 31. He wasn’t telling a customer story. He was answering a question about what seems insane today and will be ordinary in three to five years, and his answer was that AI agents will need an identity and a budget [00:16:22]. Then he reached for an example:

if you ask Anthropic how they chose to use Clickhouse, they’ll tell you they asked Claude, “What technology should we use for this specific observability use case?” And Claude suggested Clickhouse. [00:16:34]

Katz sells that database, so his account of why a customer bought it is marketing until somebody with no stake in it says the same thing. Here, somebody did, and it’s the customer. Maruth Goyal, a Member of Technical Staff at Anthropic, walked through the search in a talk at ClickHouse’s Open House conference. “It would be great if I could go ask a superintelligence what I should be using,” he joked, in ClickHouse’s writeup of the talk, published July 7, 2025. The next line is ClickHouse’s own narration rather than a quote from him: he asked Claude for its recommendation, and it suggested ClickHouse.

That’s the honest ceiling here. The customer is on the record in his own words, and every telling of it has still run through the vendor’s channel. Anthropic has published nothing about the decision on its own site. The talk and ClickHouse’s video of it both went out in 2025, so this isn’t a forecast about agentic buying. It happened roughly two years ago, and took until now to get described as procurement.

There’s no ad unit at the step where Claude answered that question. No demo request, no retargeting, no sales development rep. The only material the model had to work from was what ClickHouse had already published, and the only budget line that improves the outcome is the one that pays for documentation.

This landed on ClickHouse and not on a company with a bigger sales floor for the same reason Katz names his one regret. He has “about a 100 quota caring salespeople,” against a revenue base he described as “significantly more than 100 million,” while the competitors he named “have thousands of salespeople” [00:07:34]. Over two years, the thing he wishes he’d done differently was “increasing sales capacity” [00:08:16]. A company built that way has to be findable without a rep in the room. With an open-source product, he said, you can start “without any sort of vendor relationship,” then evaluate, deploy and scale “without a traditional enterprise sales process” [00:29:08].

A model reading your docs is that same buyer, doing the reading faster and with less patience.

We’ve covered the downstream half of this before. Research on B2B buying found AI agents shopping for software stall most often on the pricing page, then cite a competitor. Jess Graham’s “discovery tax” names what it costs a brand when the choosing gets delegated. An audit of 27,181 citations found answer engines quote people far more than company pages. This story sits above all three, with a named buyer attached.

For a 10-person B2B SaaS the translation is unglamorous. The comparison page nobody got around to writing, the benchmark that went into a Notion doc and stayed there, the integration guide that stops at step four. Those are what a model reads when a buyer asks what to use for a specific job. Treating those as support costs misreads what they now do. They’re the only inventory on the shelf at the moment a technical buyer forms a list.

The case against reading too much into this

Katz doesn’t think brand spend goes away, and he was asked directly. 20VC’s Harry Stebbings put it to him that developer relations and brand might matter less once agents are deciding. Katz said no. On a big data warehousing project at a bank or a telco there’s “still going to be a human that’s making that architectural decision,” and that holds “for the next decade” [00:31:10]. His reason is the sharpest line in the episode. “budgets are still controlled by people, right? And budgets correlate to technology decisions” [00:31:39]. He then described hiring the Chainsmokers to play a party at AWS re:Invent, which is no one’s documentation strategy.

The objection comes from the person with the most to gain from the opposite claim, which is what makes it worth printing. It sizes the story rather than cancelling it. Anthropic’s was one technical workload, picked by an engineer under pressure with a three-person team, and that’s the shape of purchase that gets delegated first. The bank’s warehouse migration goes last.

So the shortlist and the signature have come apart. A human still signs, and a model increasingly decides who gets considered, and those two jobs now want different budgets. Fund the party if you’re selling to the telco. Fund the docs either way, because the reading happens before anybody knows your name.