dbt State and agent-ready data: why declared meaning comes first

0:05Host:MBA Training. Ten minutes on Agents were the headline at dbt Summit, but a 0.2% reuse rate was the real lesson. If a tool you paid for reused almost nothing on day one, who gets fired: the vendor or your team?

0:20Expert:Your team. And the vendor's own customer is the evidence.

0:25Host:Oh, come on. You're siding with the vendor?

0:28Expert:Not quite. I'm siding with the arithmetic. Four weeks ago at dbt Summit in Las Vegas, the conference run by the merged Fivetran and dbt Labs, everyone talked about agents. The useful lesson was a tiny number from a betting company.

0:45Host:Fanatics Betting and Gaming.

0:47Expert:Yes. They switched on dbt State with no configuration and got a model reuse rate of about 0.2%. That's essentially nothing.

0:57Host:Back up a step for the listener. What is dbt State?

1:01Expert:dbt is the tool that turns raw warehouse tables into clean, modeled ones. dbt State is a paid feature that checks whether a rebuild would change anything. If nothing would change, it skips the rebuild or reuses what's already there. Less rebuilding means a smaller warehouse bill.

1:20Host:And at Fanatics it skipped next to nothing.

1:23Expert:Out of the box, yes. Then they declared source freshness, meaning how stale each input is allowed to get. They also set model-level SLAs, the service-level agreements that say how current a table must be. After that, reuse went up to roughly 15%.

1:42Host:So the magic button is a form you have to fill in.

1:45Expert:Pretty much. The tool can't tell when the milk has gone off unless someone writes the date on the carton.

1:52Host:That's myth one, then. "Smart orchestration saves money automatically." Your verdict?

1:58Expert:Wrong. Plainly wrong. Fanatics, CarGurus, Virgin Media O2 and RxBenefits all have savings stories, and every one of them is published by dbt Labs. They sell the product, so cross-check before you quote them. The Fanatics story is the honest one because it shows the gap between installing the tool and configuring it.

2:21Host:Does the fee eat the savings?

2:23Expert:It's priced per "daily active target table" on a day it gets reused. That's a little under ten cents each, and every test counts as its own table.

2:33Host:Wait, the tests are billed separately?

2:36Expert:Each one. A model with two column tests costs you three charges for every day it's reused. Test-heavy projects pay more per reused model.

2:46Host:Wonderful. We get penalized for being careful.

2:50Expert:A little. The twist is that if you haven't configured anything, it reuses nothing, so you pay almost nothing. The cost you can't see on the invoice is engineering time spent declaring freshness. One more catch: it builds Python models but never reuses them.

3:08Host:Quick aside while we're on autonomy. SiliconANGLE covered Dell's AI Leadership Symposium this week. Speakers there said AI governance is moving to runtime as always-on agents spread.

3:21Expert:Runtime meaning while the agent is actually running, not in a review afterwards.

3:26Host:Right. And CIOs, the chief information officers, want agents and LLM calls kept inside their own virtual private clouds. LLM being the large language model, and a private cloud being a fenced-off network they control.

3:41Expert:That's sensible, and it's also incomplete. Fencing the agent in doesn't tell it which table is the right one. A perfectly contained agent can still pick the wrong revenue number.

3:53Host:Which brings me to myth two. "You need a context layer before you can trust agents."

3:58Expert:Half true. And the half people get wrong is the expensive half.

4:03Host:Define it first.

4:04Expert:A context layer is a governed store of definitions, metadata and business knowledge that an agent reads before it answers. Fivetran's version pulls in dbt's structured definitions plus unstructured material like docs and Slack threads. It's supposed to settle which of the several tables called "revenue" is the real one.

4:26Host:That sounds genuinely useful.

4:28Expert:It does. Now ask them what stage it's at. Their press release says Private Beta. Their own blog says Limited Public Preview.

4:37Host:Hang on, say that again.

4:39Expert:Two labels from the same company for the same product. Either way it's early access. If you evaluate it, get the release status and the pricing basis in writing.

4:50Host:I'll argue the other side. Get in early, shape the roadmap, and be ready when it matures.

4:56Expert:And spend the year teaching a beta product definitions you never wrote down.

5:01Host:Or you write them while you pilot. Both at once.

5:04Expert:In theory. In practice the pilot takes the meetings and the definitions never get written. Mark Rittman, the consultant, put it well. Getting the context layer right is mostly hygiene, not a technology purchase.

5:19Host:Mm. I'm not fully convinced. Some teams only find the discipline when there's a product deadline in front of them.

5:26Expert:Then they've bought a very expensive alarm clock.

5:29Host:We'll come back to that. To be fair, there is one design choice you like.

5:35Expert:Openness. It publishes to what they call Agents Schema, which is plain SQL tables in your own warehouse that any AI tool can query. No black box. That's the right instinct, even if I'd still wait to buy.

5:50Host:Then what does a definition look like when it's written properly?

5:54Expert:DailyRevOps lists the fields. Semantic name, business owner, source models, grain, filters, time basis, currency or unit rules, freshness expectation and known exclusions.

6:08Host:Grain meaning one row per what.

6:11Expert:Exactly. Per order, per customer, per day. Get that wrong and the agent counts everything twice and sounds completely sure of itself.

6:21Host:Like a confident intern.

6:22Expert:A confident intern with access to the board deck.

6:26Host:Before the third myth, auditing. There's no binding rulebook for agents yet.

6:31Expert:O'Reilly reported that the Trump administration's agreement with major AI companies is voluntary, with no legal enforcement. Separately, the FTC, the Federal Trade Commission, is investigating OpenAI, Anthropic and others over consumer risks. Nobody is going to hand you a checklist. You need to be able to answer for yourself.

6:54Host:Answer for yourself how?

6:56Expert:Picture the Monday meeting where the CFO asks why last quarter's revenue was revised. If the logic lives in git, the version-control system, you show a diff. If it lives inside a platform UI, someone ends up answering from memory.

7:12Host:That's a horrible spot to put an analyst in.

7:15Expert:It is. Someone's credibility gets spent on a question the pipeline should have answered for them. dbt Labs makes that argument against Databricks, so read it as competitive positioning. It also quotes Databricks itself saying over 80% of new databases on its platform were being created by AI agents.

7:37Host:That figure is old, though.

7:39Expert:About a year and a half old, and it covers one platform. It still shows the direction. When agents write the models, version history is your audit trail.

7:49Host:Myth three, then. "Token costs are a procurement problem." Tokens are the chunks of text a model bills you for.

7:57Expert:Wrong, and it's the most underrated story of the week. dbt Labs staff pointed Claude at Gong call transcripts through an API wrapper. Gong records sales calls, and an API is a programmatic connection. A single account history was eating huge amounts of context and hitting Gong's rate limits.

8:17Host:And the fix was a fancy new tool?

8:19Expert:Ordinary data work. Fivetran already loaded Gong into Snowflake. They modeled and summarized the transcripts there, then served them through the dbt MCP server. MCP is the Model Context Protocol, a standard way for AI tools to fetch data. Compression cut the volume by 20x or more.

8:41Host:Twenty times! From a summary table?

8:44Expert:A long call went from many thousands of tokens down to a few hundred. It's self-reported, so cross-check it. But procurement can haggle over the price per million tokens all year. How many tokens each answer uses is decided by how the data was modeled before it reached the model.

9:02Host:So the token bill really belongs to the data team.

9:05Expert:Whether they want it or not. And a context layer still in early access won't fix a raw transcript fed straight to an agent. A dbt model that summarizes it fixes that today.

9:17Host:Pull it together for us.

9:19Expert:It's the same lesson every time. Skipping a rebuild, choosing a revenue table, answering an auditor, keeping a token bill down. Each one depends on meaning someone declared ahead of time. My rule is that no scheduler, agent or dashboard acts on a model that lacks a named owner, a written definition, a freshness tolerance and at least one test.

9:43Host:Even I can't argue with that one. The beta question, I'll keep arguing.

9:47Expert:Keep arguing. Just write your definitions while you do.

9:51Host:Then give people one thing to do this week.

9:54Expert:Take the metric your AI pilot gets asked about most. Give it a named owner, a written definition, a freshness tolerance and one test, and merge it before Friday.

10:05Host:Write the date on the milk carton.

10:07Expert:Or the agent will happily drink it for you.

10:11Host:Sources: Fivetran, Rittmananalytics and Getdbt. Every link is in the show notes. We'll leave it there. The CDO self-assessment at mba-training.com will tell you where you really stand.