A growing number of companies are taking shortcuts in enterprise AI rollouts by wiring AI directly to vendor-provided MCPs and source data providers. According to dbt Labs, this approach is collectively costing billions in token consumption, vendor API charges, and loss of data ownership and portability. dbt Labs says it initially made the same mistake, using vendor MCPs until the hidden costs became clear.
The company's solution was to centralize data and engineer context directly within its data warehouse, rather than pulling raw data through APIs. When sales teams at dbt Labs connected Claude to Gong and Salesforce via native MCPs to analyze deals, costs escalated quickly. Analyzing a single call consumed up to 50,000 tokens, costing $0.25 per transcript read, with over 500 million tokens of Gong transcript data available.
By ingesting raw Gong data into the warehouse and using data modeling to summarize each call, dbt Labs shrank data volume by 20x and cut token consumption on a 60-minute call from tens of thousands to a few hundred. Serving call transcripts from warehouse summaries dropped token costs by roughly 98% while maintaining or improving context quality.
The company describes this as "context engineering" — modeling data for AI agents, not just for dashboards. Traditional analytics engineering focuses on quantitative data for human users viewing charts. Context engineering extends this to qualitative data like emails, transcripts, PDFs, support tickets, and chat logs, with agents as the primary consumers.
Key practices include compressing data to relevant forms, enriching it with related information, describing what it means, and governing what each agent can access. dbt Labs emphasizes not aggregating context but generating it using warehouse-native AI functions like parsing, chunking, and embedding. The recommended architecture is a single shared, governed context layer that all agents read from, with data read once in batch and served to every agent.
The company notes that direct MCP connections caused redundant token costs and API constraints, leading to the realization that existing data modeling processes could be applied to AI data. Following its merger with Fivetran, dbt Labs says moving and modeling unstructured data is now within its capabilities. The stack remains the same — Fivetran moves data, dbt models it — but now includes unstructured context. The analytics engineer, the company argues, is now the context engineer, with the data team owning AI and becoming mission-critical.
Source: dbt Labs Blog