Deliver reliable AI with the dbt Semantic Layer and dbt MCP Server

AI is changing how we use data at a pace we've never seen before. There's a new tool every day that requires us to redo entire workflows.

Data engineering teams that once focused solely on pipelines and transformations are now on the front lines of AI strategy. They're expected to deliver intelligent, reliable data products faster than ever before. And somehow, at the same time, they need to make sure those systems are trustworthy, scalable, and secure.

Just as dbt revolutionized data transformation in the cloud era by adding better processes around ETL - modular, version-controlled, testable workflows - we're now doing the same thing in the AI space. In this article, I’ll delve into how you can use the dbt Model Context Protocol (MCP) Server, combined with the dbt Semantic Layer, to deliver data to AI and ensure unparalleled accuracy.

The AI context gap

Our customers are moving incredibly fast, building everything from agentic workflows and internal copilots to customer-facing chatbots. However, the underlying AI stack is changing even faster.

Each week, there's a new breakthrough, a new model, a smarter framework, or even a new vector store. That velocity means AI and data teams are constantly adapting, moving from Claude to Cursor, from Pinecone to semantic layers. Many are retraining systems just to keep up.

Each of these shifts introduces friction: new integrations, rewrites, tuning, and duplication. There's a lot of time, risk, and slowing of innovation caused by this wide variety of tools.

At the end of the day, a lot of the fundamental questions are the same. What does this metric mean? Is this definition still up to date? Where did this number come from? Can we trust this data?

For enterprise AI, the bottleneck isn't compute power or data volume. It's context. Specifically, a structured, governed context.

Large language models can't reason over raw or fragmented data. They need a strong foundation. Without structured definitions and shared business logic, AI systems get stuck guessing. And when they guess, they start hallucinating. That's where you see all the data problems of asking a question and getting the wrong answer.

Take an internal chatbot, for example. A user asks, "What's our revenue for enterprise customers in Q2?" To answer that question, the model has to understand a few things:

• What defines an enterprise customer?
• Which up-to-date table or model includes revenue logic?
• Does that revenue table account for cancellations, ticket sales, or returns?
• Are there any other considerations for how you answer that question?

Without structured, governed context, the large language model might generate SQL that runs but returns the completely wrong answer with complete confidence. You would think the machine is correct without being able to validate it.

Or consider an agentic workflow generating an entire weekly order summary without a version-controlled definition of "order." It could pull that definition from the wrong table, double-count returns, or blatantly miss critical logic. That could result in incorrect results for a question as simple as "What was my revenue last quarter?"

dbt as the control plane for AI

What's needed isn't more tools. It's a unifying data control plane for your AI systems: one place where your business logic, transformations, tests, and documentation live. Structure it once, then apply it everywhere.

That's where we see dbt fitting in. With dbt as the control plane for your AI, you structure your data once and use AI everywhere.

AI systems need access to structured context. That’s primarily what dbt offers: the ability to create data about your data that can then be used for querying, for AI, for all these other use cases.

Those AI systems are going to pull that context so we can get to a more centralized, governed platform and workflow. The idea is that we can remove a lot of that context that exists in fragments or tribal knowledge and lock it down in code so it's reliable across your entire organization.

dbt is already the standard for creating high-quality, governed datasets from your warehouse. It captures rich metadata: model lineage, test coverage, and centralized metric definitions. It's also bringing performance and cost efficiency.

Rather than connecting each AI workflow to its own source, you can let dbt centralize those transformations, metrics, and documentation in one layer. This reduces queries, decreases compute spend through unified models and state-of-the-art orchestration, and provides faster response times for your end users.

With dbt, you get cross-platform flexibility. dbt works across all major data warehouse systems, including Snowflake, Databricks, and BigQuery. You're able to build fast across your entire stack without sacrificing consistency.

Another important consideration is security and governance, which is one of the big things holding back production grade AI projects. dbt helps close that gap. With every model, test, and transformation, we can validate that it's logged, versioned, and auditable. This means you can meet enterprise-grade requirements for compliance and data protection with AI.

Introducing the dbt MCP Server

Everything I've covered so far has been about providing the context that's required. So, how does dbt make it really easy to integrate your tools?

That's where the Model Context Protocol (MCP) fits in. Tools like LangChain and Semantic Kernel can directly query the dbt Semantic Layer, data lineage, and tested models via API so that your AI systems don't just access the data, they also understand it.

While the AI tools may change - GPT-4 to Claude, chatbots to agents - your foundation shouldn't change. Or, at least, it doesn't have to if you're using dbt.

The MCP unifies all dbt assets and AI applications. We make that accessible for you in two ways: both a local connection and a remote connection.

The local option is pretty self-explanatory. It runs on your laptop alongside your dbt project. It's a fantastic option for local development with tools like Cursor or Claude, and it empowers your agent to create your dbt code right on your machine.

The second option is to run it remotely. We have an incredibly straightforward setup where you can plug and play your dbt MCP Server with any of your AI tools and access them through any web application. It means you can connect multi-agent, multi-user systems more easily than ever before. You'll see a wide variety of MCP integrations with many of our partners in the future.

Real-world impact: the M1 Finance case study

A lot of what I've shared has been theoretical conversation, but I wanted to support that with a real case study.

By introducing dbt MCP, M1 Finance was able to reduce its engineering bottlenecks and dramatically improve its efficiency. Their teams didn't have to wait for specialized resources to move projects forward. It gave them a clear path to reduce some of the biggest blockers to adopting AI basically, hallucinations.

With structured, validated access to those systems, the AI could act on real authoritative data. With that additional context, they were finally confident that the answers and outputs they received were accurate, reliable, and safe. This lets them unlock true scale, not just a proof of concept or an experiment.

A customer journey: Galaxy's Edge Travel Company

To illustrate how this works in practice, let me walk you through a customer example we created.

Imagine Galaxy's Edge Travel Company, the largest tourism operator in the outer rim of space. They offer a wide range of services, including star cruiser vacation packages, droid-assisted lodging experiences, lightspeed-enabled transportation, and holotable concierge services.

Galaxy's Edge wanted to build an AI concierge—a holo guide—a much cooler version of Alexa or Siri. This holo guide can answer any questions instantly: pricing, availability, loyalty points, packaging rules, and bundling recommendations. You ask it what you want to know, and it spits out the answers immediately and always consistently.

The challenge? They have the data for all of these systems, but it's scattered across multiple platforms. They have star cruiser manifests in JSON events captured from their hyperspace travel system, droid service logs in semi-structured information, resort booking information stored in structured tables, and loyalty point balances that come through microservice API dumps.

While you likely don’t work for an intergalactic travel tourism operator, these systems probably resonate with systems in your environment.

Understanding the context gap

The holo guide doesn't understand the business. It's a brand new AI system. It can show up, but it doesn't know anything about their systems. This isn't because the AI is bad, it's because their enterprise context is a mess. If you take all this data in disparate systems and put an AI on top of it, it's not going to produce very good results.

For instance, calculations such as total trip cost vary across systems, producing different results. Room names could be "Deluxe Pod," "D-Pod," or "Pod Deluxe," and they could all refer to the same room, but AI doesn't know that unless we provide context. Loyalty tiers aren't joined correctly, causing the AI to hallucinate discounts and produce inconsistent pricing. Packaging availability depends on complex business logic.

All of these things have to be taken into account to ensure the holo guide consistently produces the right results. The organization recognizes that AI is only as good as the structured context we provide. What they need is a data control plane and a structured context layer built on something like dbt.

Building the solution: three key steps

To power their AI concierge, Galaxy's Edge needed to implement three critical components:

Data modeling. First, they created consistent facts and dimensions for trips, packages, customers, and droid services. They curated this data, breaking it up into facts and dimensions in an easy-to-understand manner.

The Semantic Layer. They defined metrics like total trip cost, occupancy rate, and loyalty-eligible balance that are always correct. The Semantic Layer also defines the model's joins to help it understand how joins are performed, adds business-friendly naming, and stores calculations and metrics.

The MCP Server. This is how they expose governed dbt context directly to the AI agent—in this case, the Holo Guide. When someone asks, "How many loyalty points will I earn if I add the Holocron Discovery Tour to my three-night star cruiser stay?" the system produces accurate results.

For AI to be reliable, it needs to know the definition of a Holocron Discovery Tour. It needs to pull semantic metrics on cost, loyalty accrual, and discount rules. It needs to join the customer profile or customer dimension, and it needs to apply real business logic.

To see a walkthrough of how to implement this in practice, sign up for our recent webinar and watch the full demo.

What's coming next: dbt agents

The dbt MCP Server is just one part of our AI strategy. We’re also working to build several agents directly inside the platform. We're developing tools such as an analyst agent, discovery agent, observability agent, developer agent, and even more over the upcoming months that will solve specific parts of the Analytics Development Lifecycle directly within dbt.

The first one I've tried is dbt Insights, which lets you use natural language querying to generate SQL and get results. These agents will be among the easiest and most powerful ways to use dbt AI directly within the dbt platform.

Conclusion: structured context is the foundation

AI systems are only as good as the structured context you provide them. Without that foundation - without clear definitions, governed metrics, documented transformations, and tested models - your AI tools will hallucinate, produce inconsistent results, and ultimately fail to deliver business value.

With dbt as your data control plane and the dbt MCP Server connecting your AI tools to that governed context, you can:

• Build reliable AI experiences that scale
• Reduce hallucinations by providing clear business logic
• Accelerate development by automating tedious tasks
• Deliver trusted data products that your business can depend on

Whether you're building internal chatbots, customer-facing AI concierges, or agentic workflows, the foundation is the same: structured, governed, reliable data. And that's exactly what dbt delivers.