Peter Whitehead, a Solutions Engineer at Omni, shows you how to build an AI-driven dashboard using Omni. You’ll work through a real-world dataset, learn how to design and structure an effective dashboard, and explore ways to incorporate AI features that enhance analysis and user experience. This session is ideal for BI analysts who want to build more modern, intelligent reporting tools.
Summary
Peter Whitehead calls AI in BI "perhaps the biggest shift in business intelligence in recent memory" — and then spends an hour showing what that shift actually looks like in practice.
Whitehead, a solution engineer at Omni and a veteran of Google's Looker team, leads a hands-on code-along demonstrating how to build an AI-driven dashboard from scratch using the Omni BI platform. Working through an Amsterdam Airbnb dataset, he covers Omni's layered architecture — SQL warehouse, shared semantic model, workbook, dashboard — and shows how each layer connects. He queries data through conversational AI (via Omni's agent, Blobby), builds visualizations using the point-and-click field picker, adds a rolling average with an Excel-syntax formula, constructs a geospatial heat map, and embeds AI summary tiles with custom prompts directly into the finished dashboard. A one-minute JSON import themes the entire dashboard to match DataCamp's brand palette.
The architectural through-line is Omni's semantic model: a governed layer where data teams define metric logic, business context, and row-level security before any AI query runs. This is what separates AI business intelligence tools that produce reliable answers from those that guess wrong joins and hallucinate metric definitions.
A closing Q&A covers semantic layer governance, how users verify AI-generated answers, and how Omni customers are feeding user query logs back into the model to sharpen AI responses over time. Anyone learning how to build an AI dashboard for a business context will find a detailed, working example in this session.
Key Takeaways
- Omni's semantic model sits between the SQL warehouse and every dashboard query, encoding metric definitions, business context, and row-level security so the AI agent always works from the same governed logic.
- Conversational AI is one entry point, not the only one — users can move between natural language, point-and-click, spreadsheet formulas, and raw SQL within the same workbook window.
- AI summary tiles can be embedded directly in dashboards with custom prompts controlling tone, focus, and output format, including markdown and emoji.
- Every AI-generated result in the workbook is verifiable: users can inspect the SQL, check which fields were selected, and ask the agent to explain its own reasoning.
- Omni administrators can export a full log of user natural-language queries. Customers are feeding these logs back into the semantic model to close gaps in AI response quality.
- Custom workbook-level metrics can be built on the fly — duplicate a measure, add a filter (private rooms only, for instance), and Omni generates the underlying YAML automatically.
- Dashboard themes import as JSON in about a minute, applying a brand's color palette across all visualizations at once.
Deep Dives
Why the Semantic Layer Is the Foundation of Accurate AI Analytics
Most text-to-SQL approaches work by sending an AI model a raw database schema and letting it write queries. The problem is schemas don't encode business logic. What counts as a "customer"? Is "revenue" recognized at booking or at checkout? Without those answers baked in, the AI guesses — and wrong guesses make dashboards unreliable for real decisions.
Omni's approach, shaped by the founding team's Looker background, is a shared semantic layer sitting between the data warehouse and every query the platform issues. Metric definitions, descriptions, join paths, and security rules live there. So does business context: descriptions that tell the AI what a field actually represents and when to use it. The model governing AI analytics in Omni isn't a raw schema — it's a curated layer of human judgment.
Whitehead explains the practical risk of skipping this:
"We don't wanna risk AI hallucinating the wrong metric definition or joining in the wrong table so things are slightly off and who's making the right decision."
The stakes aren't just technical. At the organizational level, inconsistent metric definitions create conflict:
"If two people come to the meeting and each of them have defined sales revenue in slightly different ways, there's gonna be an argument."
The semantic model also opens up who can contribute context. Marketers who understand customer acquisition cost better than any data engineer can write their own definitions directly into the model. Those definitions become available to every user querying it — and to every AI query running against it. Business knowledge moves into the layer where the AI actually reads it.
Four Ways to Interact with Data, One Window
Omni's premise is that different analysts need different interfaces, and the same analyst needs different interfaces at different moments. The platform puts conversational AI, point-and-click exploration, spreadsheet formulas, and SQL into a single workbook window and lets users move between them mid-analysis.
Whitehead frames the design choice directly:
"It's not just a one way street with natural language. Users sometimes want to take over the wheel actually start interacting with the data, you know, in a traditional BI sense or even a traditional Excel sense."
In practice: a user asks Blobby a question, gets a visualization back, clicks into the workbook to adjust it through the field picker, adds a calculated column using standard Excel syntax, and finishes by writing a window function in SQL — all without switching tabs. Omni's Excel formula layer is a one-to-one syntax match, so anyone who can write an AVERAGE function can extend a query.
This multi-modal design also speeds up work for technical users who don't want to hand everything to AI. Whitehead on how he uses it himself:
"I have started using this more and more just to get me to a starting point. So I don't need to spend, you know, fifteen clicks to get to the analysis I want. I can just have it give me that starting point, and then I can start building on top of it." The AI provides a faster on-ramp. The user finishes the job.
Building an AI-Driven Dashboard: The Full Walkthrough
The session's core is a live build using Amsterdam Airbnb data. Whitehead starts from Omni's home screen, asks Blobby for a set of KPIs — average nightly price, unique hosts, average revenue, total listings — then moves the resulting query into a workbook to construct each visualization manually.
The workbook's field picker organizes columns into topics: curated, pre-joined table sets with predefined relationships. Selecting a dimension or measure updates the SQL in real time. For an availability trend chart, Whitehead adds a seven-day rolling average by writing an AVERAGE formula in a new spreadsheet column, then drags that column into the chart's y-axis alongside the raw availability line.
The map visualization demonstrates Omni's geospatial support. Selecting latitude and longitude fields and switching the chart type to a point map projection produces a heat map of listings across Amsterdam. Row limits are adjustable — Whitehead raises the cap to 20,000 rows mid-build. Tooltip fields (host name, rental type, nightly price, average review score) give hovering users per-listing context without cluttering the map itself.
Building a stacked bar chart — total listings by room type, broken down by rental type — takes two steps: ask the AI to construct it, then pivot on rental type through the field picker's kebab menu. Whitehead also demonstrates creating a custom metric on the fly: duplicating the average review score measure, adding a filter for "private room" only, and saving it as a new field available to any visualization in the workbook. The GUI builds the underlying YAML automatically.
Once each visualization is ready in a workbook tab, it promotes to the dashboard in one click. Dashboard-level filters — a neighborhood dropdown in this case — can target specific tiles or map generically across the board, routing to the appropriate date or dimension column in each chart.
AI Summary Tiles and the Transparency Problem
Omni includes a purpose-built AI summary visualization: a dashboard tile whose content is generated by a prompt running against live query results each time the dashboard loads. The tile pulls in whatever data the underlying workbook tab returns, sends it to a language model with the user-defined prompt, and renders the output as formatted text.
The prompts can be specific. Whitehead demonstrates asking for the top three hosts and the worst-reviewed host, formatted with markdown and emojis. The tile returns exactly that. The same mechanism works for executive summaries, segment highlights, or any narrative that's easier to describe in plain language than to encode in a chart.
On transparency, Whitehead is direct: "We wanna be very transparent about what is AI generating and what is not." AI summary tiles carry a default disclaimer "Generated by AI — verify accuracy". Dashboard editors can remove it, but the platform includes it because AI text is nondeterministic: the same query can produce different summaries on different runs. Users who need to verify an AI answer can open the workbook, inspect the SQL, review the fields and joins, and ask the agent to walk through its reasoning step by step.
The AI in Omni's system is explicitly positioned as a starting point, not a final authority. The workbook is always accessible, the SQL is always visible, and the override path is always one click away.
Using User Questions to Improve the Semantic Model Over Time
Late in the Q&A, host Richie raises a point that surfaces often in enterprise BI deployments: if fifty users are asking the same question through the AI chat, that pattern signals a gap in the dashboard — a filter, dimension, or metric that's needed but not yet exposed.
Omni logs every natural-language query, the user who sent it, and the AI response. Administrators can view this through the platform's analytics page and export the full dataset. Some of Omni's more advanced customers are feeding these logs into a model refinement workflow: identifying which questions the AI handles poorly, tracing failures back to missing context in the semantic layer, and updating metric descriptions or adding fields in response.
The feedback loop moves model improvement from a scheduled data engineering task to something that responds to actual user behavior. Business users — through the questions they ask daily — become an indirect signal for where the semantic model needs to grow. They don't need to write YAML or read schema documentation. They just ask questions, and someone is listening to the pattern.
As Whitehead puts it, building context into the model doesn't have to fall solely on the data team: "It could be the marketers who are sitting closest to the business who understand what customer acquisition cost means to them and why it's important to them so then everyone across the business can share in that same importance as they're asking questions of the data."
