Pairing Power BI with an Azure OpenAI assistant for leaders

Pairing Power BI with an Azure OpenAI assistant gives leaders two tools that read the same governed data: dashboards to see performance at a glance, and a private chat assistant to ask about the numbers and the documents behind them. The assistant runs in your own Azure tenant, searches only approved sources, and filters every answer by the user’s permissions. The work starts with agreed KPI definitions and one data model, not with the AI.

Why pair dashboards with an assistant?

Executive dashboards answer “where do we stand?” quickly. They are less good at “why did this move?” or “what does our policy say about it?”. Leaders end up emailing analysts or searching shared drives. An AI assistant fills that gap: it answers questions in plain language, summarises documents and points to the source. When both read from the same governed data, the dashboard and the assistant never disagree about a number.

What does the architecture look like?

Azure OpenAI runs the language model inside your Azure subscription, so prompts and documents stay within your agreed region and contracts. The assistant uses retrieval-augmented generation: it searches an index of approved documents, sends only the relevant passages to the model and answers with references. Our RAG explainer covers this pattern in more detail.

How do you keep it secure?

Tie access to identity: Use the organisation’s sign-in, so permissions follow each user automatically.
Filter before retrieval: Only search documents and data the user is allowed to see.
Approve sources: Index only documents that owners have approved for the assistant.
Log and review: Keep a record of questions, sources and answers for audit.
Review every release: Run security reviews with risk and IT teams in each sprint, not only at launch.

For a wider view of AI risks such as prompt injection and data leaks, see our AI security guide.

What design choices matter most?

One index up front: Lead with a single weighted score, colour-coded for below, near and on target.
Pillars beneath it: Show each performance area with its change against the previous period.
Target and gap side by side: Put score, target and gap in one table so nobody has to calculate.
Filters that match how leaders think: Country, city and branch, plus date and period.
Chat that fits work: Saved history by day, copy and regenerate buttons, voice and attachments.
Sources on every answer: Show which document an answer came from, so leaders can check it.

Dashboard layout follows the same rules as any good report; our guide to dashboard design principles covers hierarchy and colour.

What did this look like for Mashreq NEO?

Mashreq NEO is the digital bank of Mashreq in the UAE. Its leadership measured performance across six areas: financial results, customers, risk, operations, strategic initiatives and market position, each reported separately. We designed and built an executive analytics platform that brings all six into one weighted performance index in Power BI, with targets, gaps, factor trends and drill-down to a single branch. Alongside it, NEO GPT is a private assistant that answers from approved internal documents and only reads what each user is allowed to see.

Read the full Mashreq NEO case study for the screens and process.

How should you run the project?

KPI framework: Agree pillars, weightings and targets with leadership.
Data modelling: Build one governed model that every report reads from.
Dashboard design: Prototype the executive view and test it with real users.
Assistant build: Index approved documents, connect identity and test answers.
Rollout and review: Train users, collect feedback and review answers weekly.

Work in short sprints with the client’s data, risk and IT teams. The KPI framework takes longer than most people expect, but it is what makes everything after it trustworthy.

Could you use Copilot instead?

For questions about numbers in a semantic model, Copilot in Power BI may be enough, provided you have Fabric or Premium capacity. A custom assistant is worth it when leaders also need answers from documents, a branded experience or tighter control over sources. We compare the options in Power BI Q&A, Copilot or a custom assistant.

How UIUX Studio helps

We design and build executive Power BI dashboards and private AI assistants that share one governed data model. See our Agentic AI & Automation and Technology Advisory services.

Frequently asked questions

Does Azure OpenAI use our data to train models?

Microsoft states that Azure OpenAI does not use customer prompts or data to train its models; check your agreement.

Can the assistant read Power BI data?

Yes, through the semantic model or an API, filtered by the user’s permissions.

How long does a project take?

Usually several months, with the KPI framework and data model taking the most time.

Who should own the assistant?

A named business owner plus the data and IT teams.

Can leaders ask by voice?

Yes, voice input can be added to the chat.