AI business intelligence: plain-language answers on your warehouse and Power BI

AI business intelligence lets people ask questions of company data in plain language ("what was gross margin by region last quarter?") and get a number, a chart or a short written answer. It works when metrics are defined once, in a semantic layer, and the assistant is told to use those definitions and show which table and definition it used. Without that step, the assistant guesses what "revenue" or "active customer" means, and two people get two answers.

#What AI adds to a BI stack

Your BI stack probably already has a warehouse or database, a modeling layer, and dashboards in Power BI, Tableau, Looker or similar. AI adds a conversational front end and some drafting help on top.

In practice it does four jobs:

• Answering ad hoc questions that no dashboard covers yet.
• Summarizing a report or a period in a few sentences.
• Helping analysts write queries and measure definitions faster.
• Drafting commentary for a monthly pack, for a person to edit.

It doesn't replace the modeling work, and it shows quickly whether that work was done properly. A dashboard hides ambiguity because someone chose the definition when they built the chart. A chat box exposes it, because anyone can type any question.

#The step most projects skip: define metrics once

A semantic layer is a single, governed description of your business terms: what each metric means, how it's calculated, which tables it comes from and how those tables join. Your BI tool may already have one. In Power BI it's the semantic model, with its measures and relationships.

Microsoft's own documentation makes the same point about its assistant. Its Copilot for Power BI overview (opens in a new tab) says to prepare the semantic model before using Copilot, and warns that without preparation Copilot can misinterpret the data and return generic or inaccurate results. It also notes that when a question isn't related to the semantic model, Copilot answers from the language model's general knowledge. A reader needs to know which of the two they got.

A metric definition needs to carry the following.

Write these for the 20 to 40 metrics your leadership actually uses, and skip the rest of the columns. A small, well-defined set beats a large, vague one.

#Three ways to build it

Check the vendor's requirements before choosing. Microsoft's overview, updated in August 2026, says Copilot in Power BI needs a paid Fabric capacity (F2 or higher) or Power BI Premium (P1 or higher), and that trial capacities and free SKUs aren't supported. Some experiences are still labeled preview. Requirements change, so confirm them on the current page.

The third approach is the one we'd build when numbers feed decisions that people sign off. The model turns a question into a request for a defined metric, which it does well. A query engine computes the figure the same way every time, which it does better.

#How a good answer looks

A trustworthy answer carries its working. If the finance director asks for gross margin by region for last quarter, the reply should show:

• The figure and the period, with the fiscal calendar named.
• The metric definition it used, with a link to the entry.
• The tables and filters behind it.
• A caveat if one applies, such as "credit notes after quarter end may restate this".
• A clear line when it can't answer: "I can't compute that because returns aren't linked to regions in this model."

That last behavior matters most. An assistant that always returns a number will eventually return a wrong one, and nobody will notice until a board meeting.

#Building it step by step

1. Pick a domain with an owner and a real audience, such as finance reporting or the sales pipeline. Avoid starting with the whole company.
2. List the questions people ask. Pull them from analyst requests, email and chat threads. You'll find that 30 to 50 questions cover most of the need.
3. Define the metrics behind those questions, using the table above, and get the owner to sign each one off.
4. Fix the data issues you find, such as missing keys, mixed currencies and inconsistent dates. These come up in almost every project.
5. Connect the assistant to the semantic layer or model, with row-level permissions mirrored from your BI tool. If a regional manager can't see another region's data in a dashboard, the assistant can't show it either.
6. Build a test set. For each question, record the correct number, computed independently by an analyst. Score the assistant on whether it picks the right metric and returns the right value.
7. Add the approval step for anything published. A person checks every figure that goes into a board pack or an owner statement.
8. Review misses weekly. Each wrong answer usually points at a missing synonym, an undocumented rule or a data defect.

#Where it goes wrong

Ambiguous terms. "Customers" could mean accounts, billing entities or active users. Without a defined term, the assistant picks one.

Joins that multiply rows. Joining orders to payments to refunds inflates totals when the relationships aren't one-to-many. Text-to-SQL is especially prone to this.

Permissions. If the assistant connects with a service account that can read everything, it ignores your row-level rules. Use the asker's identity.

Silent fallbacks. If the tool answers from general knowledge when the data can't, readers may mistake a guess for a result. Force a visible "not in the data" response.

Narrative over numbers. A fluent paragraph of commentary can make a weak number sound settled, so keep the figure and its definition next to the text.

Cost. Ad hoc questions run queries, and queries use compute. Limit query size and frequency, and watch usage in the first weeks.

Key point: AI makes asking easy and leaves answering unchanged. Define each metric once, make the assistant cite the definition and table it used, and test it against numbers an analyst computed independently.

#When to skip the chat box

Some reporting shouldn't be conversational at all. Month-end packs, owner statements and regulatory returns need fixed layouts, tie-outs to the ledger and sign-off. Automate those as scheduled reports with validation checks, where AI drafts the commentary and a person approves it. Our month-end close blueprint shows where such checks sit, and the monthly owner statements use case is a worked example.

Our reporting automation work covers both sides: scheduled reports with checks, and plain-language access to the numbers behind them. The retrieval method used on the documentation side, such as finding which metric definition applies, is explained in the RAG glossary entry. If you're weighing a pilot, book a strategy call and bring your five most-asked data questions and where the numbers live today.

#Frequently asked questions

#What is AI business intelligence?

It's the use of AI to make data analysis conversational: staff ask questions in plain language and get figures, charts or summaries from your data. Done well, it sits on a governed semantic layer so each answer uses defined metrics and cites its source.

#Does Power BI have AI?

Yes. Microsoft's Copilot in Power BI can answer questions, summarize reports and help authors write queries and measures. Microsoft's documentation lists capacity requirements and marks some experiences as preview, so check the current page and prepare your semantic model before relying on it.

#Can AI query our data warehouse directly?

It can, using text-to-SQL, but accuracy depends on how clear your tables and metrics are. Without a semantic layer, expect wrong joins and inconsistent definitions. A safer design lets the model pick from defined metrics and has a query engine compute the result.

#How do we stop the assistant giving wrong numbers?

Define each metric once, restrict the assistant to those definitions, require citations and let it decline when the data can't answer. Test it against figures an analyst computed independently, and rerun the test when the model or data changes.

#Can AI write our board or investor report?

It can draft commentary and assemble tables from approved metrics. A person should check every figure against the source and approve the final report. We don't recommend publishing AI-written financial narrative without that review.