AI agents querying your warehouse: measure before rebuilding

September 23, 2026 · 5 min · Free, with full transcript

Fivetran and dbt Labs used dbt Summit to argue that AI agents, not human analysts, are now the main consumer of enterprise data. The position here is that the architectural shift is real but the urgency is a sales pitch, and that re-platforming off a keynote is the expensive mistake. You come away able to test dbt Labs' claim that 30% of warehouse queries are machine-generated against your own logs, spot the failure mode where agents pass nonsense results downstream with no scar tissue, make the case for a written semantic layer over new tooling, and set a query budget before an agent touches production. References include O'Reilly Radar and MIT Sloan Management Review.

Listen to the podcast

4 min

Chapters

Key takeaways

  • Pull last month's query logs and calculate what share came from service accounts rather than named humans to find your real agent load.
  • Treat dbt Labs' 30% machine-generated query figure as directional, since the vendor sells the tooling that raises it.
  • If agent traffic is small, spend the quarter writing business definitions down instead of changing architecture.
  • Build the semantic layer so terms like active customer are defined in writing, because an agent has no memory to fill the gaps.
  • Set a spending cap on agent query volume before a pilot touches production data to avoid a five-figure compute surprise.
Read the full transcript

Host:This is Leaders Insights. On the table. If agents are the new primary consumer of your data, is your infrastructure built for the wrong audience?

Expert:The mistake I keep seeing this month is CDOs rebuilding their entire stack because a keynote told them agents are the new users.

Host:Panic-driven architecture. And the keynote was loud.

Expert:At DBT Summit last week, DBT Labs and 5tran rolled out a cluster of products all pointed at the same idea. That the thing querying your data warehouse is no longer a human analyst but an AI agent asking questions on its own.

Host:Translate that. What actually changes when the consumer is a bot instead of a person?

Expert:A human analyst runs maybe 40 queries a day and gets bored. An agent will fire off thousands, chain them together, and never sanity check a nonsense result, so the failure mode flips. A person notices when revenue comes back as negative. An agent just passes it downstream to the next agent.

Host:So this is real, not just a slide.

Expert:The shift is real. The urgency is being sold to you. Those are different things.

Host:Give me a number that tells me which is which.

Expert:DBT Labs put out a figure that something like 30% of the queries hitting their customers' warehouses are now machine-generated rather than human. Worth flagging, they sell the tooling that makes that number go up. So treat it as directional, not gospel. Cross-check against your own query logs before you believe your stack is under siege.

Host:And if I check my logs and it's 3% not 30?

Expert:Then you don't touch your architecture this quarter. You watch the trend line. The independent read matters here. O'Reilly's radar surveys through 2026 show agent adoption climbing fast in engineering but still thin in the data layer specifically. Plenty of companies are experimenting. Far fewer are in production.

Host:Here's the uncomfortable one. Vendors always say the audience changed so you need to buy the new thing. Why is this any different from the last five waves?

Expert:It mostly isn't, and that's the honest answer. But there's one piece that isn't marketing. The metadata problem. Metadata is just the labels on your data. What this column means. Where it came from. Whether it's trustworthy. A human analyst fills those gaps from memory. They know the revenue table excludes refunds because they've been burned before. An agent has no memory and no scar tissue. So if your definitions live in someone's head instead of written down, the agent guesses and it guesses confidently.

Host:So the real work isn't buying an agent-ready warehouse. It's documentation.

Expert:It's the semantic layer. The written down business definitions that say active customer means logged in within 30 days, not ever. MIT Sloan had a piece this year making exactly this point. The organizations getting value from agents aren't the ones with the fanciest models. They're the ones who did the boring work of defining their terms first. That's not a product you purchase. It's a project.

Host:Second number. What's the cost angle because thousands of automated queries can't be free.

Expert:That's the one nobody put on a slide. If agents run 10 to 100 times the query volume of humans, your cloud compute bill scales with it. These are walking into pilots with no cap on how much an agent can spend and getting a five-figure surprise. So before you let an agent loose on production data, you set a query budget the same way you'd set a spending limit on a corporate card.

Host:You've now told people not to rebuild, not to trust the vendor number, and to write documentation. That's a lot of not doing.

Expert:Because the expensive mistake is motion. Reacting to a keynote by re-platforming cost you a year and a budget. The cheap, useful move is small.

Host:Then land it. One thing my listener does Monday morning.

Expert:Pull last month's query logs and calculate what share came from a service account instead of a named human. That's your real agent load, not dbt's 30% yours. If it's small, you've got time to write your definitions properly. If it's already large and undocumented, you've found your fire. And it isn't the one the vendors were pointing at. Measure your own audience before you rebuild for someone else's.

Host:That'll do it. What we read for this one?

Expert:The new stack, dbt labs, vendor, data tooling, O'Reilly radar, MIT Sloan Management Review.

Host:End of episode. The CDO calculators are running at MBA-training.com.

Read and go deeper