By Jeffrey Shear, J Shear Consulting

In my last article, Claude, Power BI, or Both?, I made the case that most companies shouldn't choose. Power BI should be the governed foundation, Claude should be the fast, conversational layer on top, and clean data has to come before either one.

The most common follow-up I've received is a practical one: "Okay, but how do we actually set that up so it works?" This article is the playbook. It's written for business and technology leaders at mid-sized companies, with a short section at the end you can hand to your IT team.

The operating model in one picture

Think of your reporting in three layers. Each layer has one job, and problems start when a layer tries to do another's job.

The key principle: Claude Dashboards should read from the Power BI semantic model, not around it. When Claude uses the same definitions as your official reports, the numbers match. When it goes directly to raw tables or spreadsheet extracts, you've created a second version of the truth.

Step 1: Make your Power BI model AI-ready

This is where most of the value is won or lost, and it starts, again, with clean data. If customer records are duplicated or product codes are inconsistent, an AI assistant will reproduce those problems fluently and convincingly. Everything in Your Dashboards Are Lying to You applies here first.

Once the data is sound, the model itself needs to be understandable to an AI, much as it would need to be understandable to a new analyst:

• Create a measure for every core metric. If "Net Revenue" exists as an official measure, Claude can use it. If it doesn't, Claude will calculate its own version, and it may not match yours.
• Use plain business names. "Net Revenue" and "Active Customers," not "Measure 7" or "SUM_AMT_ADJ."
• Write descriptions. Add a one-sentence business definition to each key measure and table. These descriptions are what an AI reads to choose the right field.
• Hide what users shouldn't use. Hide technical columns, keys, and raw amount fields that could be summed incorrectly.
• Confirm relationships and security. Make sure relationships are correct and row-level security is in place, so each person sees only what they're entitled to.

Claude can help here, too. Paired with Microsoft's Power BI modeling tools, it can draft measure descriptions, flag inconsistent naming, and review a model against best practices, with a person approving every change.

Step 2: Connect securely, and connect live

Claude connects to Power BI through Microsoft's Power BI MCP servers, which let AI assistants query semantic models. Two choices matter most:

• Use the signed-in user's identity. When each person connects with their own Microsoft account, their existing Power BI permissions and row-level security apply automatically. A regional manager asking Claude a question gets only their region's data, just as in Power BI.
• Prefer live connections over extracts. A live connection queries the model each time, so answers stay current and security stays enforced. An extract is a snapshot: it goes stale and sits outside your security model.

Start with a small pilot group of five to ten people with a real need for faster answers, not the entire company.

Step 3: Give each tool a clear lane

Without clear guidance, Claude Dashboards can quietly become "shadow BI": dozens of unofficial dashboards nobody maintains. A simple rule set prevents that.

For the gray areas, use two tests: audience and lifespan. If a Claude dashboard is still being used by a wider group after about 30 days, it has earned a place in Power BI.

Step 4: Teach Claude to use your definitions

Even with a well-built model, how people ask matters. Encourage users to reference official measures by name, and set standing instructions for your team. A few examples:

• "Using the Sales semantic model, show Net Revenue by region for the last four quarters. Use the official Net Revenue measure; don't calculate revenue from raw tables."
• "Compare Gross Margin % this quarter to the same quarter last year, by product line, and explain the three biggest changes."
• "Which customers moved out of Active Customers in the last 90 days? Use the model's definition of an active customer."

Then build a verification habit. Claude Dashboards lets you click any number to see the query behind it. For anything important, especially early on, check that it used the official measure and that the figure matches your Power BI report.

Step 5: Create a promotion path

The real power of the two tools together is the path from idea to production:

1. Prototype in Claude. A leader asks a question; Claude builds a dashboard in minutes.
2. Validate with users. Is it used? Does it change decisions?
3. Reconcile the numbers against the official Power BI reports.
4. Rebuild it in Power BI. Claude can help draft the DAX and report layout for your Power BI developer to finalize.
5. Retire the prototype so there's only one version.

One practical note: there's currently no direct export from Claude Dashboards into Power BI, so promotion means a rebuild. Keep prototypes simple, and treat them as a requirements document rather than a finished product.

Step 6: Keep governance light but real

• Assign an owner for the Claude Dashboards program, usually whoever owns Power BI.
• Review sharing settings. Decide whether dashboards can be shared outside the company, and have your administrator set that policy.
• Review data handling. Confirm how data is processed and retained under your Claude plan and your compliance requirements.
• Hold a monthly review of which Claude dashboards are being used, what should be promoted, and what should be retired.

Common pitfalls

• Connecting Claude to raw tables or spreadsheet extracts instead of the governed model.
• Skipping model documentation, then blaming the AI for picking the wrong field.
• Treating beta and preview features as production. Pilot first; features and terms will change.
• Letting prototypes live forever and become a second, unofficial reporting system.

A composite example

Consider a 180-person regional healthcare services organization, a composite of situations I've seen. Power BI was well established, but department leaders waited weeks for answers to new questions, and row-level security was essential because of the sensitivity of the data.

They started with Step 1: about three weeks of cleaning provider and location data, adding descriptions to their 14 core measures, and hiding raw fields. They then connected Claude to the Power BI model using each leader's own credentials and ran a 60-day pilot with eight leaders.

Early on, the verification habit paid off. One dashboard showed a visit count that didn't match the monthly report; Claude had counted from a raw table instead of using the official measure. They hid that column, added a standing instruction, and the problem didn't recur. By the end of the pilot, ad-hoc requests to the analytics team had dropped by roughly half, two Claude prototypes had been rebuilt as official Power BI reports, and the analysts were spending their time on work that mattered.

A 30-day starter plan

For your IT team: setup checklist

The bottom line

Claude Dashboards and Power BI work best as partners with clear roles: clean data at the bottom, Power BI as the single source of truth in the middle, and Claude as the fast, conversational way to ask questions on top. Set it up in that order, give each tool its lane, and verify the numbers, and you get the speed of AI without giving up the trust your reporting depends on.

If you'd like help preparing your Power BI model for AI, setting up a secure Claude pilot, or designing the governance to keep it all consistent, I help companies put this playbook into practice.

Book a consultation →