Ask your data: Power BI Q&A, Copilot or a custom AI assistant?
Power BI Q&A, the older natural-language box, stops working in February 2027. For questions about numbers in a semantic model, Copilot in Power BI or a Fabric data agent is the natural replacement. For questions that also need policies, procedures or other documents, many organisations add a custom AI assistant that searches approved documents with retrieval-augmented generation (RAG) and respects each user’s permissions.
What is changing with Power BI Q&A?
Q&A let report readers type questions such as “sales by region last month” and get a chart. Microsoft announced its retirement in December 2025 and, in a September 2026 update, moved the date to February 2027. From then, Q&A no longer works anywhere in Power BI and existing Q&A visuals show an error. Q&A Setup, including synonyms and “teach Q&A”, is replaced by Prep data for AI.
Before the date, inventory every report that uses a Q&A visual. You can select the convert icon on a Q&A answer to turn it into a standard visual that keeps working. Embedded and sovereign-cloud scenarios need extra planning, because Microsoft says there is currently no Copilot equivalent for some of them.
What are the options for asking questions?
| Option | Best for | Reads from |
|---|---|---|
| Copilot in Power BI | Questions about a report or model | Semantic model |
| Standalone Copilot | Finding and querying any report you can access | Semantic models across workspaces |
| Fabric data agent | A focused question-answering agent for one domain | Chosen tables, lakehouses and models |
| Custom AI assistant | Policies, procedures and mixed questions | Approved documents, sometimes plus data |
Copilot needs paid Fabric capacity at F2 or above, or Premium P1 or above. We cover its features and limits in Copilot in Power BI.
When does a custom AI assistant make sense?
Copilot is strong when the answer is a number in a well-built model. Many leadership questions are not like that. “What does our credit policy say about this?” or “Which procedure applies to a VIP outage?” live in documents, not tables. A custom assistant uses retrieval-augmented generation: it searches approved documents, passes the relevant passages to a language model and answers with references. Our explainer on RAG covers how retrieval works, and vector databases covers where the documents are indexed.
How do you keep answers within permissions?
- Tie sign-in to your identity system: Use the same accounts and groups as the rest of your tools.
- Filter before the model sees anything: Only retrieve documents and rows the user is allowed to see.
- Use row-level security in models: So Copilot and data agents return each user’s slice of data.
- Log questions and sources: Keep an audit trail of what was asked and which sources were used.
- Test for prompt injection: Documents can contain hidden instructions, so test the assistant with hostile content before launch.
Security reviews should be part of every release, not a final step. Our guide to prompt injection and AI security explains the common attacks and defences. In regulated sectors such as banking, involve risk and compliance teams from the first sprint, because they will need to sign off on what data the assistant can reach and how answers are logged.
What does a combined setup look like?
For Mashreq NEO, we designed two products that work together. Executive dashboards in Power BI show one weighted performance index, six pillars and drill-down to branch level. NEO GPT, a private assistant, answers in plain language from approved internal documents, and only reads what each user is allowed to see, because access is tied to the bank’s identity system. Both read from the same governed data, so leaders see a number in the dashboard and can ask about the policy behind it.
How do you measure answer quality?
Whichever option you choose, decide up front how you will judge whether answers are good enough. Build a test set of real questions with the correct answer and the source it should come from. Run it after every model change, every new document batch and every software update. Track three things: whether the answer is right, whether it cites the right source, and whether it refused when it should have. Readers lose trust quickly after one confident wrong number, so a short list of well-tested questions is better than a wide launch that nobody checks. Add a simple feedback button so users can flag wrong answers, and review those flags weekly with the data owner.
How do you choose?
- List real questions: Collect 30 questions leaders and staff ask each week.
- Sort them: Mark each as numbers, documents or both.
- Check your models: If most questions are numbers and models are clean, start with Copilot.
- Plan for documents: If many questions need policies or procedures, scope an assistant.
- Pilot with one team: Measure answer accuracy and time saved before rolling out.
How UIUX Studio helps
We design and build Power BI dashboards and secure AI assistants that read from the same governed data. See our Agentic AI & Automation service, or read about pairing Power BI with an Azure OpenAI assistant.
Frequently asked questions
When does Power BI Q&A stop working?
From February 2027, after Microsoft moved the date from December 2026.
Can I keep a Q&A answer?
Yes. Convert it to a standard visual before retirement.
Does Copilot replace Q&A?
Yes, along with standalone Copilot, Copilot in mobile and Prep data for AI.
Can an assistant answer from PDFs?
Yes, a RAG assistant can search approved documents and cite them.
Do I need both?
Many teams use dashboards for numbers and an assistant for documents.