Agents were the headline at dbt Summit, but a 0.2% reuse rate was the real lesson

This week's items all point at the same problem. A scheduler deciding whether to skip a rebuild, an agent deciding what "revenue" means and an auditor asking how a number was produced all depend on meaning that someone wrote into the pipelinepipelineAll active sales opportunities across the stages of the sales process, together with their combined potential value and probability of closing.View full definition → ahead of time. The dbt Summit keynote was four weeks ago, but its follow-up material kept arriving this week: dbt Labs' pricing explainer for dbt State, a Redshift adapter update and Fivetran's October product notes. It landed alongside O'Reilly's read on AI oversight and a KDnuggets piece on governed machine learning.

My thesis is that AI autonomy and cost pressure are making the same demand from opposite directions. For a CDO, the budget case and the AI trust case are now one piece of work: declared metadata (definitions, owners, freshness tolerances, tests) written into the transformation layer. Teams that treat this as documentation will get neither the compute savings nor agents they can trust.

dbt v2 and dbt State GA turn skipped rebuilds into a priced product

Here are the facts, all from the vendor's own release. Fivetran + dbt Labs announced the general availability of dbt v2 and dbt State on 16 September in Las Vegas. Fivetran + dbt Labs sells ingestion and transformation tooling, so treat every performance figure below as a vendor claim. According to the company, dbt v2 is a full Rust rewrite of the dbt engine built for the scale that teams run at today and for how agents write SQLSQLSales Qualified Lead: a prospect the sales team has validated as ready for direct outreach and a proposal, having passed clear qualification criteria.View full definition →, and it parses a 10,000-model project up to 10x faster than v1. Rittman Analytics, a consultancy writing up the event, says figures of 6.5x to 9x on comparable runs were quoted from internal testing at the partner keynote, so expect less than 10x in practice. The governance news is quieter: with this release, the two engine era of Core and Fusion ends, and both remain Apache 2.0-licensed and security-supported.

dbt State matters more to a CDO. It checks warehouse metadata and model SQL, works out whether the result would actually change, and then builds, skips, clones, or auto-defers to production. The customer claims, all published by dbt Labs, look like this:

• Virgin Media O2's Gordon Curzon credits it with 25% savings on both job run time and BigQuery compute costs.
• At CarGurus, the company saw a 9% compute reduction, with 35% fewer models built.
• RxBenefits' team has cut costs by 59% on scheduled jobs in the dbt platform running on a Snowflake adaptive warehouse.

The most useful number is the one the vendor didn't put in its headline. At Fanatics Betting and Gaming, turning dbt State on with no configuration gave them a model reuse rate of about 0.2%, essentially nothing. The effectiveness came after they configured source freshness and declared model-level SLAs, and that took the pilot from 0.2% to roughly 15% model reuse. The tool only paid off once the team had declared what "fresh enough" meant.

Worked example: does the reuse fee pay for itself?

The pricing is public. dbt State charges $0.094 for each daily active target table, with every test counted as a separate target table. So a model with two column tests costs $0.282 for each day it's reused, no matter how many runs reuse it that day.

Here is a hypothetical, arithmetic only. Take a project with 200 models, each carrying two tests: 600 target tables. At the 15% reuse Fanatics reached, that is 90 reused tables a day, or $8.46 a day, about $254 over a 30-day month. Now assume $8,000 a month of warehouse compute goes to dbt jobs. At CarGurus' 9%, you save $720 and net about $466. At the 15 to 30% band the vendor advertises, you save $1,200 to $2,400. Because the fee is tied to reuse, an unconfigured deployment costs almost nothing and saves almost nothing. The real cost sits in engineering time spent declaring freshness. Note too that test-heavy projects pay more per reused model, and that dbt State builds Python models but does not reuse them.

This is a decision for yourdata FinOps practice, and not something to sign off on a slide. Pilot it on your most overscheduled project, where Fanatics found the biggest reuse gains, and use the free trial window to measure. The same logic of not paying twice for the same bytes shows up in another dbt Labs post this week: dbt can now materialize models across Redshift data warehousedata warehouseA central repository that consolidates data from many source systems into a structured, query-optimized store designed for analytics, reporting, and business intelligence.View full definition → without copying data, and the adapter has migrated away from legacy Postgres metadata APIs to Redshift-native SVV_* system views and SHOW APIs. Redshift shops should test it. Everyone else can skip it.

What is a context layer, and does a CDO need one yet?

A context layer is a governed store of definitions, metadata and business knowledge that an agent reads before it answers. Most CDOs need the definitions now and can wait on the product. Fivetran's version, per the vendor's press release, unifies the data and metadata needed to give LLMs and AI agentsAI agentsAgentic AI refers to AI systems that pursue goals autonomously by planning, taking actions through tools, and adapting based on results, with minimal step-by-step human direction.View full definition → relevant context, building on dbt's structured context and adding unstructured knowledge, like docs and Slack threads. The design choice worth noting is openness. Fivetran says it publishes to Agents SchemaSchemaA schema is the formal blueprint that defines how data is structured, named, typed, and related within a database, file, or message.View full definition →: open, plain SQL tables in your own warehouse that any AI tool can query. Its pitch includes features to resolve "revenue" to one governed dataset out of the several that claim the name, and metadata connectors that pull from the dbt Semantic LayerSemantic LayerA shared business definitions layer that sits between raw data and reports, so everyone uses the same numbers for revenue, churn or margin.View full definition →, and from sources like Looker, Sigma, and Power BIBITechnologies and processes that turn raw data into actionable insights via reporting, dashboards and analysis, so teams can decide based on facts rather than intuition.View full definition →.

Pin the vendor down on status. The press release calls it Private Beta. Fivetran's own announcement blog says it is now in Limited Public Preview. That is an early-access product, whichever label you go by. dbt Charts sits alongside it in public beta: dashboards built as YAML, version-controlled right next to the models they depend on, so that charts an agent drafts can go through the same pull request as the SQL.

The best independent comment came from practitioners. Mark Rittman's view is that getting the context layer right is mostly hygiene rather than a technology purchase. If you already have dimensional models and a semantic layer over your gold tables, the remaining choice is about technology. Another post-Summit write-up said it would keep a wary eye on the joint Fivetran + dbt Context Layer initiative. My advice is to buy nothing yet. Pick your ten highest-stakes metrics and write each one up asa governed metric definition. DailyRevOps lists the fields well: semantic name, business owner, source models, grain, filters, time basis, currency or unit rules, freshness expectation and known exclusions. Any context layer you buy later will only be as good as those entries.

AI oversight stays voluntary, so auditability lands on the pipeline

In this week's This Week in AI, O'Reilly reported that the Trump administration announced a voluntary agreement with major AI companies that calls for internal safety monitoring, external audits, and independent board reviews, which carries no legal enforcement. Meanwhile, the Federal Trade Commission launched an investigation into OpenAI, Anthropic, and other AI companies focused on potential consumer risks. The O'Reilly hosts conclude that teams will need governance practices that can withstand outside review.

For a CDO, the practical reading is that nobody is going to hand you a compliance checklist for agents, so you need to be able to answer for yourself. KDnuggets made the engineering case on 8 October: enterprise AI teams cannot afford to treat governance as a last compliance review, and the requirements belong in training data selection and model behaviour from the start.

A dbt Labs post, written partly to argue against putting transformation logic inside Databricks, puts the audit question more sharply. It is competitive positioningpositioningThe mental space you want your brand to occupy in your target customer's mind relative to alternatives.View full definition →, so read it as such. Still, its test holds up: when your CFO asks why Q3 revenue was revised, tracing the answer requires Databricks platform access rather than a git diff. It also cites Databricks' own figure from the May 2025 Neon acquisition, that over 80% of the databases on the platform were being created by AI agents rather than by people. That figure is a year and a half old and covers one platform, but it shows where things are heading. When agents write models, being able to see version history for your definitions is the audit trail. Nothing needs buying here. Run the drill described below.

Why modeling for the tokentokenA token is the basic unit of text that language models process, often a word fragment, whole word, or punctuation mark rather than a single character.View full definition → matters more than the Context Layer

The underrated item this week is a dbt Labs blog post about its own internal plumbing, so it is self-reported. When staff began using Claude against Gong call data through an APIAPIApplication Programming Interface: a standardised interface that lets applications communicate and exchange data without knowing each other's internal workings.View full definition → wrapper, one account history query was burning 8,300 tokenstokensA token is the basic unit of text that language models process, often a word fragment, whole word, or punctuation mark rather than a single character.View full definition → per call transcript or around 240,000 tokens in context for one account history of ~30 calls, and it hit Gong's rate limits. The fix was ordinary data work. Fivetran already landed Gong in Snowflake, so the team modeled and summarized transcripts there and served them through the dbt MCPMCPAn open standard that lets AI assistants connect to your company's tools and data in a consistent, governed way, instead of one custom integration at a time.View full definition → server. Actively compressing transcripts shrank data volume by 20x or more, slashing a 60-minute call from 10,000+ tokens to just a few hundred.

It is underrated because it moves AI unit cost onto the data team's side of the ledger. Most firms budget tokens as a model-vendor line item, and procurement negotiates the rate per million. This case shows that volume per answer is set by how the data was modeled before it reached the context windowcontext windowThe context window is the maximum amount of text (measured in tokens) a language model can process at once, including both the input prompt and the generated output.View full definition →, and that can mean a 20x difference. A Context Layer still in early access will not fix a raw transcript fed straight to an agent. A dbt model that summarizes it will, today, with tools most teams already own.

The declaration gate: a rule for automating any model

Write this on the whiteboard:no scheduler, agent or dashboard may act on a model that lacks a named owner, a written definition, a freshness tolerance and at least one test.

Each failure mode this week maps to a missing declaration. Without freshness tolerances, dbt State reused nothing at Fanatics. Without definitions, an agent guesses which revenue table to trust. Without version-controlled logic, the CFO's question gets answered from memory. Without a modeled shape for the data, the token bill grows with every question asked. Apply the gate model by model and start with whatever your agents touch first.

What to do on monday

• Export 30 days of dbt job history and count the models rebuilt when neither upstream data nor code had changed. That number is your dbt State business case, or your reason to skip it.
• Run the DATT arithmetic above on your own model and test counts before any trial clock starts, and flag any Python models, which will not be reused.
• Pick the ten metrics your AI pilots are most likely to be asked about and fill in the DailyRevOps contract fields for each one, with a named business owner.
• Run the CFO drill: choose last quarter's most-revised KPIKPIKey Performance Indicator, a measurable value that shows how effectively you're achieving a specific objective, tracked over time against a target.View full definition → and time how long it takes to explain the change from version control alone.
• Ask the owner of your largest AI use case for tokens per answer, and find where unmodeled source data enters the context window.
• If you are evaluating Fivetran Context Layer, get its release status and pricing basis in writing, given the conflicting labels.

The Fanatics number is the one to bring to your next steering committee. Turning the tool on bought a 0.2% reuse rate, and declaring freshness and SLAs took it to roughly 15%. Agents will behave the same way: they can only use the meaning your team has already declared in the pipeline.