AI business intelligence: what each BI tool's AI does, and how to check it

Read the documentation for eight BI tools' AI features in one sitting and a pattern appears. They describe different interfaces. They make the same promise and the same caveat.

• Microsoft: prepare the semantic model first, because without it "Copilot can misinterpret the data and return generic or inaccurate results" (Microsoft Learn, August 2026).
• Google: Looker's conversational analytics treats LookML as the "source of truth"; Looker, not the AI, writes the database query from the model (Google Cloud docs).
• Omni: "The AI does not generate raw SQL directly from text." It writes a query against Omni's semantic layer, which Omni turns into SQL.
• ThoughtSpot: Spotter compiles questions "into identical governed SQL on every run", on ThoughtSpot's semantic layer, Snowflake semantic views or dbt semantic models.
• Metabase: "The richer your semantic layer, the better AI performs."

The honest summary of AI business intelligence, then, is a language model translating questions into queries against definitions a human wrote. Where those definitions are good, it works well. Where they aren't, it answers anyway.

That makes the semantic layer, not the chat box, the thing to invest in. We've argued before that nobody wants to maintain two of them (semantic interoperability); if you already have one in dbt, choose a BI tool whose AI can read it. More in semantic layer for AI.

We read each vendor's own pages on 22 October 2026, and Count's row is ours. The table records what they say; where a page didn't say something, neither do we.

Two we left out. Julius returned an error to our browser, so we couldn't confirm anything from its own pages. Tableau's marketing pages also blocked us, so Tableau's row comes from its help documentation.

The column we'd weigh most heavily is the third. A BI answer you can't trace back to a query is a chatbot answer with a corporate logo.

Switching BI tools for an AI feature is rarely worth it. The AI is only as good as the model underneath, and your model lives in the tool you already have. So for most teams the choice is made: turn on the AI in your current BI tool, tidy the model, and limit it to governed data.

Power BI shops: Copilot is the obvious route, but check the capacity requirement before promising it to anyone. It needs paid Fabric (F2 or higher) or Premium (P1 or higher), and trial capacities don't count.

Tableau shops: Tableau Agent needs Tableau+ on Tableau Cloud. Its help pages are frank that it indexes field names and samples values, so badly named fields confuse it. Rename before you roll out.

Looker shops: conversational analytics inherits your LookML, which is the strongest starting point of the three. Hidden fields are ignored, so hide what you don't want it to touch.

Choosing fresh: pick the tool that shows its working. That's what we build Count to do, so weigh our view accordingly: it's a canvas that connects to your warehouse and business apps, runs SQL and Python, and has its own agent, and every claim in an answer comes with the reasoning chain behind it. Your team can work on the canvas or ask from Claude or ChatGPT through its MCP server. Omni, ThoughtSpot, Sigma and Hex were also built with AI near the centre and expose their logic: Omni and Metabase link to the SQL, Sigma shows logic and filters, Hex backs each thread with a notebook. Metabase is the cheapest place to start, with Metabot in every plan including open source.

Where Claude fits

Most of these tools now run an MCP server: ThoughtSpot, Omni, Sigma, Hex and Metabase all say so on their pages. That changes the shape of the stack. Your team can ask questions in Claude, where they already work, and Claude asks the BI tool, which answers from the governed model. You get Claude's writing and reasoning on top of definitions your data team controls, instead of Claude guessing at your schema. For one-off dashboards built in Claude itself, see AI dashboards.

If your company runs ChatGPT instead: ThoughtSpot and Hex both name ChatGPT alongside Claude, and custom MCP servers can be added in ChatGPT as plugins.

The vendors' documentation is more candid than their marketing, and it is the best list of failure modes we've found.

• Answers from memory. Microsoft says that when a question is about data in the semantic model, Copilot uses the model. "Otherwise, it answers from the large language model's (LLM) general knowledge." A business user can't always tell which happened.
• Plausible and wrong. Google warns Looker's AI "can generate output that seems plausible but is factually incorrect".
• Silent defaults. Tableau notes its agent's choices of date granularity and aggregation may not match your intent, and it can't ask clarifying questions when writing calculations.
• No memory of results. Snowflake's Cortex Analyst "doesn't have access to results from previous SQL queries", so a follow-up can't refer to the last answer's numbers.

Underneath all of them sits the text-to-SQL problem. The top BIRD system scores 82.95 against 92.96 for human data engineers (BIRD, September 2026), and an audit found annotation errors in 52.8% of BIRD Mini-Dev items, mostly from ambiguous questions (Jin et al., January 2026). A semantic layer narrows the ambiguity. It doesn't remove it. See text-to-SQL for more.

A two-minute check on any AI BI answer

1. Open the query. If the tool won't show one, treat the number as unverified.
2. Find the measure it used. Is "revenue" gross, net, booked or recognised? Is it the one in your glossary?
3. Ask a question you know the answer to (last month's customer count) and compare.
4. Rephrase the question and ask again. If the answer changes, the model is guessing at your meaning.
5. Look for a general-knowledge answer: anything that mentions benchmarks, markets or trends your data doesn't contain.

For answers that pass through Claude or ChatGPT on the way, Count does steps 1, 2 and 5 at once. Connect it to the assistant and ask it to audit the answer: each claim is labelled as a citation, an assumption, model knowledge or a hallucination, with the reasoning chain showing the query behind it.

Most of the work is in the model, not the AI settings. In the order we'd do it:

1. Write a glossary of your top 20 metrics with one definition each, and get finance to sign off the revenue ones. Our customer acquisition cost guide shows how many versions of one metric you can be choosing between.
2. Describe your fields. Looker reads labels and descriptions; Tableau indexes field names; Power BI recommends preparing the model for AI. A column called amt2 will be guessed at.
3. Limit the AI to endorsed data. Hex lets you focus Threads on "endorsed, modeled data"; Looker ignores hidden fields; Databricks Genie Agents let data teams set trusted datasets, metrics and business rules.
4. Add verified answers for common questions. Power BI apps and Databricks Genie both support author-provided answers that come back for matching questions.
5. Run evals before rollout. Omni, Hex and Snowflake all describe evaluations for testing answer quality. Write ten questions you know the answers to, and run them after every model change.
6. Connect it to Claude read-only through the BI tool's MCP server, if it has one, so people can ask from where they already work.

Step 5 is the one teams skip and regret. Without it, you find out the AI misreads "active customer" when the CFO does.

We'll declare our bias. Count has argued for years that BI tools are good at reporting and poor at decisions: they tell you what happened and leave you asking "now what?" (why your BI tool isn't driving decisions; the Tableau era is over).

AI in BI fixes the access problem. Anyone can now ask a question without filing a ticket, and that is a real gain, which we don't want to undersell. It doesn't fix the trust problem. A business user who gets an instant answer still can't see what was filtered out, which definition was used, or whether the sentence explaining the trend came from the data or from the model's general knowledge.

The tools that will last, in our view, are the ones that make the working visible next to the answer. Several in the table above already do. It's why Count draws the reasoning chain under every claim rather than handing back a number on its own.

What is AI business intelligence?

What are the best AI BI tools?

Does Power BI Copilot need a special licence?

Will AI replace BI tools?