Get Power BI AI Ready — Complete 2026 Guide

Q&A Retirement — December 2026

Microsoft announced in January 2026 that Power BI Q&A will be fully retired in December 2026. Every Q&A visual, dashboard question box, and embedded Q&A experience will stop working at that date. This is not a cosmetic change — Q&A and Copilot use fundamentally different technology. Q&A used a rules-based linguistic schema with synonyms and marked keywords. Copilot uses generative AI backed by Azure OpenAI. The metadata it reads is different. The preparation steps are different. The failure modes are different.

If your semantic models are not prepared for Copilot before December 2026, Copilot will produce inaccurate or misleading answers for your users — and those users will not be able to tell the difference from a correct answer. This guide covers exactly what preparation involves and how to complete it for every semantic model that serves Power BI reports.

Prerequisites — What You Need Before Starting

AI Data Schema — Help Copilot Understand Your Model

The AI data schema tells Copilot which tables and fields matter most, how to interpret them, and what vocabulary users will use. Without this, Copilot has to infer context from raw field names — which works poorly when names are technical abbreviations, multi-word compound names, or legacy database column names like DimCli or ID_VTA.

What to Configure in AI Data Schema

1. Define entities and their attributesMark the tables and columns that represent core business objects. An entity is a business concept — Customer, Product, Order, Region. Attributes are the descriptive properties of those entities — Customer Segment, Product Category, Order Date. Copilot uses entity classification to prioritise which tables to query when users ask ambiguous questions.
2. Rename cryptic field namesFields like DimCli, ID_VTA, fct_rev_ytd mean nothing to Copilot or to users. Rename them to human-readable labels — Customer, Sales ID, Year-to-Date Revenue. If renaming in the model breaks existing reports, use the Display Name property rather than the underlying column name. Copilot reads display names first.
3. Add descriptions to measures and key columnsCopilot reads the first 200 characters of a description when generating DAX or deciding which measure to use. Add descriptions to every measure and every key column used in business questions. A well-described measure is the single highest-value action you can take to improve Copilot answer quality. Format: what the measure calculates, which business concept it represents, any filters it always applies.
4. Add synonyms for business vocabularyUsers say “revenue”, “sales”, “bookings” — your model might use “Net Revenue Amount”. Add all the business terms users actually say as synonyms on the measure or field. Unlike Q&A synonyms (which retire with Q&A), these synonyms feed Copilot’s NL-to-DAX interpretation and are part of the AI data schema that persists after Q&A retires.
5. Mark analytical focus fieldsFlag the fields with the greatest analytical weight — date dimensions, revenue measures, customer segments, KPIs. Copilot uses this to prioritise which fields to include in responses when a question is broad. Focus on the 20% of fields that drive 80% of business questions — not the entire model.

AI Instructions — Business Context for Copilot

AI instructions are plain-language text you write to tell Copilot things it cannot infer from the model schema alone: business terminology, seasonal definitions, which table to use when multiple tables represent similar concepts, how to interpret ambiguous field values, and which filters to apply by default. They are written in the “Add AI instructions” tab of the Prep data for AI dialog and applied immediately on click of Apply.

What AI Instructions Should Cover

Example AI Instructions (Verbatim from Microsoft Learn)

# Customer Identification
- accountid refers to customers in Revenue table
- earningsid refers to Customers in Partners table
- Define "top customers" by revenue table, highest order values

# Product Metrics
- Filter by State=Washington OR State=California unless specified
- For Total Active Partners, use measure Monthly Active Partner Count_ID
- Food products always show store from Store table via store_id

# Temporal Context
- Busy season is October to February
- Fiscal year starts July 1
- Always analyze sales quarterly unless user specifies otherwise

Verified Answers — Pin Specific Visuals to Questions

A verified answer pins a specific report visual to a set of trigger phrases. When a user asks a question matching one of your trigger phrases, Copilot returns that exact visual rather than generating a new one. This guarantees consistency for business-critical questions where the answer must be exactly what your team certified — not a Copilot-generated approximation.

Requirements for Verified Answers

All five conditions must be true before you can create a verified answer on a visual:

Setting a Verified Answer

1. Select the visual you want to pinChoose a visual that represents a key business insight — a top-10 ranking, a KPI summary, a quarterly trend. Verified answers work best for questions that are asked frequently and where the answer should always come from the same certified visual.
2. Open the visual menu and select Set verified answerIn Power BI Desktop: click the … menu on the visual → Set verified answer. In Power BI Service: click the … menu → Set up a verified answer. Both paths open the same trigger phrase configuration dialog.
3. Write trigger phrases covering how users actually ask the questionInclude multiple phrasing variants: “What are our top sales regions?”, “Show top regions by revenue”, “Which regions performed best this quarter?”. Copilot matches trigger phrases against user questions — broader coverage means fewer misses. All verified answers are saved to the semantic model, not the report.
4. Save and testChanges save to the semantic model and are available across all reports built on that model. Test immediately in the Copilot pane with the exact trigger phrases you wrote, then test with paraphrased variations to confirm the match breadth.

Approved for Copilot — The Final Step

After completing your Prep data for AI configuration — AI data schema, AI instructions, verified answers — mark the semantic model as Approved for Copilot. This setting was previously called “Prepped for AI” and was renamed in 2025 for clarity. The function is identical: when a model is marked Approved for Copilot, the standalone Copilot agent does not apply friction treatment to answers from that model. Friction treatment is the extra caution Copilot applies to unverified models — hedging answers, adding more caveats, reducing confidence in generated visuals.

How to Mark a Model Approved for Copilot

1. Go to Power BI Service and find the semantic modelNavigate to the workspace containing the model. Select it to open the model detail page.
2. Open SettingsSelect the Settings icon in the upper-left corner of the model detail page.
3. Expand the Approved for Copilot sectionScroll to the Approved for Copilot section in the Settings panel. It was labelled “Prepped for AI” before the rename — if you see that label on an older documentation source, it refers to the same setting.
4. Select the Approved for Copilot checkbox and click ApplyOnce saved, the model is marked. The standalone Copilot agent will use your model’s AI instructions, verified answers, and AI data schema without applying friction treatment. Propagation may take a few minutes to an hour to fully take effect.

Copilot Features — Analyse Report vs Analyse Data

Power BI Copilot has two distinct operating modes. Understanding which one applies to a given scenario determines whether you should troubleshoot the report visual, the semantic model, or the AI instructions. Both modes are available in the Copilot pane in Desktop and Service.

June 2026 Updates — What Changed

Testing and Validation

Changes to Prep data for AI features do not take effect instantly. Test systematically after each change, not after making several changes at once — otherwise you cannot isolate which change produced which result.

Testing Workflow

1. Apply one change at a timeMake one update to AI instructions, AI data schema, or verified answers. Click Apply. Wait 5–15 minutes. Test that specific change before making another. Multiple simultaneous changes make it impossible to attribute response improvements to specific configurations.
2. Use the skill picker in DesktopOpen the Copilot pane in Power BI Desktop. Use the skill picker to select “Answers questions about the data” — this is the mode that reads AI instructions, AI data schema, and verified answers. Do not test with “Summarize this page” if you are testing model-level preparation — that mode reads visuals, not the model.
3. Test with real business questions, not technical queriesAsk questions the way your actual users will ask them: “What were our top 5 customers by revenue last quarter?”, “Show me margin by product category for busy season.” If your sales team uses “bookings” not “revenue”, test with “bookings”. Your AI instructions and synonyms should handle the vocabulary your users actually use.
4. Use the HCAAT diagnostic for opaque answers“How Copilot Arrived At This” (HCAAT) is included in Copilot answers from your semantic model. Click it when an answer is wrong or incomplete — it shows which tables Copilot queried, what measures it used, and how it interpreted the question. This is the primary debugging tool for AI instructions and AI data schema issues.
5. Download diagnostics for persistent issuesAccess the diagnostic download via the … menu on the report Copilot pane. The downloaded file shows the full context Copilot received — model metadata, AI instructions, and the query it generated. Use this when HCAAT alone does not identify the root cause.