dbt Labs tests context engineering: dbt MCP cut AI skill cost 48.5%

The problem: shared limits and rising token spend

In a blog post dated Oct 26, 2026, dbt Labs describes an internal library of AI skills, packaged instructions that tell an agent how to complete a task, such as building a recap PDF from a customer's sales calls. Teams across Fivetran and dbt Labs use them daily. Almost all of these skills pull context straight from vendor tools like Gong, which according to the company led to growing API constraints and higher token spend.

The post names three side effects:

• Shared API caps. Vendor connectors enforce org-wide rate limits, so heavy use by one team eats into everyone else's allowance.
• MCP tool parameter limits. A vendor's MCP exposes only the filters the vendor built. Gong's MCP, for example, has no parameter for finding calls by account name, so agents cross-reference other systems or bulk-read all calls on a date.
• Duplicated retrieval logic. Each skill author specifies how to find and fetch the same kinds of context, with no shared standard.

A single Gong call transcript can run past 70,000 characters, and every skill that pulled it read all of it, every time, according to dbt Labs.

The idea: digest context once in the warehouse

dbt Labs asks what would happen if the digesting were done once, upstream in the warehouse, with every skill then pulling the curated version through dbt's own MCP. The company frames this as the same "build it once, reuse it everywhere" pattern dbt pioneered for data transformation.

Two developments make this practical, the company says. Warehouses now ship AI functions that run inside SQL, and the dbt MCP server lets agents query transcripts, documents, and ticket and email bodies that sit alongside semantic layer metrics. Earlier posts in the series reported that compressing Gong transcripts this way shrank data volume by 20x and cut token use on transcript analysis from tens of thousands of tokens to a few hundred.

This post tests the idea more rigorously. The experiment used one skill, called sales-pdf-builder, which retrieves call transcripts through Gong's MCP and generates a branded recap PDF. The same skill and the same prompt were used in every variant, so the only change was the context-gathering step. In total there were 4 ways of gathering context, 80 runs and 3 measurements: cost, speed and output quality, all on real customer data.

Three hypotheses and four paths

The control used Gong's MCP to find and read transcripts. The skill was tightened so that the LLM's creative freedom would not confound results. The team then set three hypotheses, each changing one thing:

• Hypothesis 1: Swapping Gong's MCP for dbt MCP on full transcripts should leave cost, quality and speed roughly the same.
• Hypothesis 2: Retrieving a Gong brief, Gong's own AI summary ingested into the warehouse, should cut cost and speed sharply but lower quality, since the company cannot control what the vendor summary keeps.
• Hypothesis 3: Retrieving summaries written by a warehouse model that read the full transcripts, using the company's own rules, should balance cost, speed and quality.

Variant A points the agent at the warehouse's full transcript column through dbt MCP. The source text available here ends partway through the description of the variants, so details of the remaining paths and the full results table are not covered.

What the headline result means

According to dbt Labs, the headline result is that serving an engineered summary through dbt MCP cut cost 48.5% and run time 43.2%, while keeping most of the dates, numbers and next steps found in the full transcript. It missed decisions and job titles, and the company says it will show where in the full post.

For teams running agent skills against vendor connectors, the experiment points to a trade-off: curating context upstream can lower spend and ease pressure on shared API limits, but the summary definitions determine what is lost. Because the company controls those definitions in the warehouse, it can tune them to its business needs. These figures come from one skill on one company's data, and are dbt Labs' own claims.