Tableau Knowledge is the layer that decides whether your AI agents can be trusted. It turns your metrics, business rules and the context behind them into a knowledge graph, and Tableau MCP makes that graph available to any agent. The dashboard is now one output of that graph, not the product itself.
Every serious conversation I have with a data leader this year ends in the same place. Not on the model. Not on the agent framework. On who owns the definitions.
We reorganised our own practice around that question well before Dreamforce 2026. Three of the twelve glossary terms describe it from three angles: how a system that reasons in language gets to use your numbers without lying about them. Salesforce calls them semantic understanding, knowledge for agents and MCP.
I call them the part of the stack where the margin went.
It is also the part where Tableau is furthest ahead, because it stopped shipping a semantic model and started shipping a knowledge graph.
What is Tableau Knowledge?
Tableau Knowledge is Tableau's knowledge layer for AI agents. It auto-generates a knowledge graph that links your data, semantics, business rules and documents by how they relate in your business, so agents retrieve governed context instead of guessing.
Tableau draws a useful line between three layers. They map neatly onto the three glossary terms in this article.
- Semantics: The business language already defined in published data sources, workbooks and the catalog
- Knowledge: Semantics plus structured and unstructured data (documents, PDFs, wikis), connected by relationship
- MCP: The open protocol that lets any AI agent query that knowledge with the user's own permissions
Tableau Knowledge is the foundation of the Tableau agentic analytics platform announced in May. Part 1 of this series covers the platform as a whole.
Semantic understanding: the metric has to mean one thing
Semantic understanding maps your metrics and fields to real business terms, so every model and query speaks your organisation's language. The glossary opens with a line I agree with completely: "Data means nothing without context."
The amt_net_2 problem
The practical version. A table has a column called amt_net_2. A human analyst learns over months that this is net revenue after returns but before partner rebates, valid only from the 2023 migration. That knowledge lives in her head and in a wiki page nobody updated.
A language model cannot learn that. It sees amt_net_2, reads "net," and answers a net revenue question with total sincerity and no idea about the rebates or the cutoff.
Semantic understanding is writing that knowledge down where the model reads it first: business labels, descriptions, governed calculations, valid ranges, ownership.
From semantic model to knowledge graph
Tableau has had a semantic model for years. What changed this year is that it became the centre of the platform rather than optional polish, and it grew into something bigger.
Tableau Knowledge holds not only the metrics and fields but the relationships between them, the business rules and the intent behind each definition.
The Auto Knowledge Graph, generally available since July, builds the first version of that graph from your existing sources and workbooks. It then keeps updating itself based on how people use it, rather than going stale like a data dictionary.
Behind it sits a knowledge engine that starts from 33 million semantic models built in Tableau over more than a decade. No other analytics vendor has that asset. It is why the automatic first draft of the graph is good enough to correct rather than rebuild.
Why Tableau's position in the stack leads
Every other term in the glossary reads from that graph. That is a design decision most competitors have not made.
Power BI's semantic model is strongest inside Microsoft's own ecosystem. dbt's metrics layer is excellent but has no native agent runtime or system of action behind it. Tableau's graph sits between the data, the agents and the CRM.
It is also open. Tableau co-leads the Open Semantic Interchange with Snowflake and dbt Labs, which has since entered the Apache Incubator as Apache Ossie with more than 50 backers. The definitions you put into Tableau are not locked in.
For a firm like ours, a Snowflake partner as much as a Salesforce one, that is the point that makes the whole architecture recommendable.
Knowledge for agents: definitions are not enough
Knowledge for agents is the business logic, exceptions and context that turn a correct number into a decision someone can trust. The glossary puts it simply: agents need more than data.
This is the step most teams skip, and the term I was most pleased to see named.
Business logic lives in people
A semantic layer tells the agent what net revenue is. It does not tell the agent that a drop in week 52 is normal. Or that the Nordic region reports two days late. Or that any comparison involving the acquired German entity excludes the first six months, or that the CFO hears about it before the board does.
That is business logic, and today it lives in people.
An agent without it produces technically correct nonsense: a beautifully reasoned alert about a Nordic revenue collapse that is really a Tuesday.
We have treated this as the real deliverable since early 2025. Sit with the people who know the exceptions. Write them down. Store them where the agent looks. The output is a body of encoded judgement that an agent inherits, and it outlives any particular front end.
Giving judgement a home in the graph
Tableau Knowledge gives that judgement a home inside the graph itself. Rules and relationships attach to the metrics they govern, rather than being buried in a prompt or a document the agent never reads.
The graph can also connect unstructured sources, such as the finance wiki or the PDF that explains the German carve-out. Exceptions that already live in documents do not need retyping. They need linking to the right metric.
The Auto Knowledge Graph gets you the skeleton. The exceptions are what you add, and they separate an agent that is technically correct from one the CFO trusts. Tableau naming this a first-class concept is the platform catching up with how the work actually gets done, and it makes the CFO conversation much easier.
Tableau MCP: the plug that makes it portable
Tableau MCP is Tableau's implementation of Model Context Protocol, the open standard, originally from Anthropic, for how an AI model finds and calls tools and data.
It exposes Tableau's governed semantic layer and knowledge to any MCP-compatible agent. This is the single decision that puts Tableau ahead of the market, and I want to be specific about why.
What Tableau MCP is
Tableau ships MCP servers for Tableau Next, Cloud and Server. On Tableau Cloud it now runs as a managed service, so there is no server to host.
Every user signs in with their own Tableau identity over OAuth 2.1, and their existing permissions apply automatically.
That last detail matters more than it sounds. Row-level security and certified definitions travel with the question. The agent sees what that person is allowed to see, defined the way the business defines it.
One definition, consumers everywhere
It means the governed metrics and rules in Tableau Knowledge are usable by Claude, by ChatGPT, by an Agentforce agent, inside Slack or Teams, by whatever your engineering team builds next quarter. Definition in one place. Consumers everywhere.
Most vendors talk about openness and then keep the semantic layer inside their own chat window. Tableau let it out. For the first time in my career, semantic work is genuinely reusable outside the tool it was built in. That changes what a client is paying for.
MCP distributes trust, it doesn't create it
It also means the reverse. Wrong definitions are now wrong everywhere at once, with the authority of a governed source behind them.
MCP does not add trust. It distributes whatever trust you already have. That is the whole argument for doing the semantic work properly.
How Tableau Knowledge changes the buying conversation
For more than a decade, clients bought Tableau for the front end and tolerated the semantic work as overhead. That has flipped, and I welcome it. The front end is now generated in minutes and polished by people. The knowledge graph, the encoded business logic and the MCP endpoint are the asset, and they are what a client is really buying.
Industry analysts see the same repositioning.
TechTarget's coverage of the launch framed it as Tableau shifting from a front-facing analysis tool to an underlying layer for AI, and flagged that the transition will challenge some users.
So when a client asks whether to keep Tableau, the answer depends on one question: will the governed knowledge graph live there?
If yes: Tableau just became more valuable, because every agent in the company will draw from it.
If the graph will live elsewhere: the dashboard tool is a detail, and the real conversation is about that other platform. As both a Snowflake and a Salesforce partner, we can have that conversation without defending a default.
What I refuse to recommend is building a tool-agnostic semantic layer from scratch to avoid deciding.
I have watched that project consume eighteen months more than once. With Open Semantic Interchange, the portability argument for a home-grown layer is mostly gone anyway. Pick the graph your agents will read from, put the definitions there, expose it over MCP, move on.
Conclusion: the knowledge graph is the product now
The dashboard was the whole deliverable. Now it is one output of the real product, the knowledge graph. Tableau Knowledge is the best current version of it, and Tableau made it open through MCP and the Open Semantic Interchange.
Everything else in this series follows from that.
If you want to know how much of your business logic is ready for agents, talk to our team. We start with your ten hardest questions.
.webp)
