Adding an AI Agent to Embedded Power BI in Three Days

How Lumio turned embedded reporting into a guided, governed AI experience

Recently, our io.Intelligence team set out to answer a practical question: could we help a SaaS vendor turn the Power BI reporting embedded in its product into an AI experience its users could trust—and get to a working prototype in days rather than months?

The vendor was Lumio, a financial-services-focused provider whose Power BI-based reporting platform serves C-suite leaders and finance teams at community banks. We arrived with one discovery meeting behind us and no fixed plan. Three days later, Lumio’s stakeholders watched an agent work through live Power BI reports.

The problem wasn’t the reports

Lumio’s customers already relied on its Power BI reporting every day. The data was there, and so were the dashboards. The friction was navigation.

Lumio’s users are domain experts, not software experts. A finance leader might want to know what is behind a drop in deposits since month-end. They know the question, but they do not always know which report, filter, or drill-down will answer it.

The goal was not to replace the reports. It was to let users ask for what they need and have an agent navigate the reports for them, the way a trained analyst would.

What if your customers could chat directly with their Power BI reports? That is the experience we set out to build with Lumio.

Why this matters to SaaS vendors

For a SaaS vendor, the opportunity is bigger than a better reporting interface. An agent can make embedded analytics easier to use, helping customers get more value from capabilities the vendor has already built and licensed. That can increase the day-to-day usefulness of the product without requiring a redesign of the reporting estate.

• Improve adoption: Help users reach answers without extensive report training or dashboard navigation.
• Differentiate the product: Turn embedded reporting into a more guided, conversational experience.
• Support commercial packaging: Create a feature that can be positioned as part of a premium offering, where appropriate.
• Protect trust: Work within the permissions users already have, rather than creating a separate route to sensitive data.

Deciding what to build before building it

The most important work happened on the first morning, before anyone wrote code.

“The hackathon changed how we think about AI projects. With our team and interop.io’s in one room, we knew what we were building before lunch on day one. Three days later, we had something we could put in front of clients.”

— Keith Taylor, Chief Product Officer, Lumio

Lumio’s technical leads, subject-matter experts, and stakeholders sat in one room with our engineers and worked through the decisions that usually stall AI projects for weeks over email. By noon on day one, everyone had a clear build plan.

Breaking the project down came down to five questions.

1. Where does the agent live?

Inside the product experience, not in a separate general-purpose assistant. The agent sits alongside Lumio in an io.Connect workspace. io.Connect is our desktop integration platform, enabling applications to share relevant context and coordinate actions in the user’s workflow. Users can see what the agent is doing and check its answers against the reports they are already viewing.

2. How does it get access?

Through the Lumio application itself. The agent inherits each user’s existing permissions, so it can only see and do what that user already can. This helps the vendor keep the AI experience aligned with its established access model.

3. What context does the model receive?

Only what it needs. Tools filter and summarise information before it reaches the model, and the agent looks up the right workflow for each question rather than carrying all of Lumio’s business knowledge in its instructions. This keeps answers focused and reduces unnecessary model usage.

4. How do we protect unit economics?

The design keeps the model from receiving large volumes of raw reporting data. Power BI filters and application tools do the data retrieval, aggregation, and filtering first; the model receives the smaller, relevant result needed to interpret or explain the answer. Vendors can then apply usage limits at both application and user level, helping make AI costs observable and manageable as adoption grows.

5. What is it allowed to do?

The agent navigates reports, pulls relevant figures, and explains what it found. When it makes sense, it can draft a follow-up email in Outlook for the user to review. The user remains in control of whether to send it.

Starting with one valuable workflow kept the scope realistic and gave stakeholders something concrete.

Days two and three: build, test, repeat

With the plan agreed, we started building, putting early versions in front of the team to test and iterating quickly on their feedback.

On day three, the team demonstrated the prototype to Lumio’s stakeholders on live Power BI reports. An application that had not been enabled to share context and coordinate with other desktop applications three days earlier was now interop- and AI-enabled.

Watch Ivan Pidov, Lead Consultant for io.Intelligence, walk through the demo and the lessons behind it in Notes from the Field.

Why it moved this fast

Projects like this usually run asynchronously. There is an exciting discovery meeting, then weeks of email while decisions wait on the right people. A proof of value can take two or three months, and by the end, momentum may have faded.

Putting the decision-makers in one room changed that. It removed the bottleneck around the crucial decision of where to start and made it possible to validate the approach against real reports, real users, and real constraints immediately.

What we’d tell any vendor starting out

Every implementation teaches something. A few lessons from Lumio and other io.Intelligence projects apply to almost any vendor adding AI to embedded Power BI.

Product experience

• Keep the AI in the user’s visible workflow. Users are more likely to trust an experience they can see, verify, and relate to the reports in front of them.
• Start with one recurring, high-value question where users currently hunt through reports for an answer.

Security and governance

• Use the application’s existing permissions so the agent operates within the access model users already understand.
• Define clearly what the agent can retrieve, explain, draft, and act on—and retain user review where an action leaves the product.

Unit economics

• Let tools shape the data. Filtering and aggregating before information reaches the model is faster and more cost-conscious than asking the model to work through everything.
• Keep instructions lean. Store domain knowledge where the agent can retrieve it when needed, rather than including it in every request.
• Set usage and cost limits per application and per user, not just company-wide. This gives product and operations teams visibility into where usage is growing.
• Prototype with a capable model, then evaluate whether less expensive models can deliver the required quality for routine tasks. Optimise only after the experience is working reliably.

Lessons like these are often learned the expensive way, one project at a time. Bringing them in from the start is a big part of what our engineers contribute to every engagement.

From prototype to production

Lumio is now demonstrating the agent to its clients, and it has become part of Lumio’s offering.

Our build sprints end with a choice. We can hand over an implementation that the vendor’s team maintains and extends, or we can stay involved as a deployment partner through rollout.

This example is just the start. The Outlook example in the demo showed how other applications can follow suit. Because the application is interop-enabled—able to share context and coordinate actions with other desktop applications—the agent can support workflows beyond Power BI without forcing users to move between disconnected tools.

A pattern other Power BI vendors can use

You do not need to redesign your product to find out whether an AI agent can improve your customers’ Power BI experience. The Lumio approach is repeatable:

Your team keeps ownership of the product. We bring implementation experience that helps you avoid research and rework, so you can reach customers sooner.

See the full story in Ivan Pidov’s talk, Notes from the Field, or talk to our team about the Power BI workflow you would start with.