All posts

Cool Things You Can Do with Glean: Query Power BI Datasets with NLP

Cool Things You Can Do with Glean: Query Power BI Datasets with NLP

For most analytics teams, the workflow looks like this: a stakeholder has a question, the team builds (or finds) a report, the stakeholder opens Power BI, applies a filter or two, exports to Excel, and asks a follow-up question the dashboard wasn't built to answer. Then the analyst gets pulled back in.

Glean's Power BI DAX Query Action collapses that loop. It lets anyone, an analytics leader, an AE, a GM, a marketer, ask a question in plain English from Glean chat, get a live answer from a Power BI dataset, and keep asking follow-ups until they're done. No DAX, no exports, no analyst sitting in the middle.

Here's what that actually looks like.

In action: from "what are sales?" to a marketing campaign in six turns

We pointed Glean at the Superstore dataset (a Tableau sample most analytics teams know) and asked one starting question. What followed was a six-turn conversation that ended with a specific marketing strategy and a draft outreach email all from live data, with no manual report-building.

Turn 1: Top-line. "What are sales for 2025?" Glean ran a DAX query against the Orders table using the [Sales $] measure, filtered to 2025, and returned $539,772.18.

Turn 2: Drill into categories. "What categories are performing the best?" Three rows came back: Technology ($189.9k), Office Supplies ($182.1k), Furniture ($167.8k).

Turn 3: Slice by region and margin. Without rebuilding anything, we asked for a Category × Region breakdown with profit margin. The result surfaced Technology–East and Office Supplies–West as the strongest combinations, while flagging Furniture in Central and South as carrying negative margins.

Turn 4: Best and worst products. Glean ranked products by 2025 sales and pulled margin alongside, putting the Canon imageCLASS 2200 Copier ($21.7k, 41.3% margin) at the top and a tail of unprofitable appliance accessories at the bottom.

Turn 5: Year-over-year underperformers. "What products underperformed compared to 2024?" Glean computed sales for both years, took the delta, and sorted ascending. The biggest drop: the GBC Ibimaster 500 Manual ProClick Binding System went from $13.6k in 2024 to $0 in 2025.

Turn 6: Turn the insight into action. "Who should I market this to? Create me a strategy and an outreach email." Glean used the underlying customer-level data to identify the top 5 prior Ibimaster buyers (Adrian Barton at $9.9k, plus four others) and the top 5 cross-sell targets (heavy binder buyers who didn't purchase the Ibimaster in 2024). It then drafted a "refresh and optimize" outreach email that referenced each customer's actual 2024 purchase.

Six turns took a vague top-line question to a named-account play with messaging attached. In a typical workflow, that's days of back-and-forth between an AE, an analyst, and a marketer.

Why this matters for analytics teams

The most visible value is for the business user, who gets answers without bothering an analyst. The more interesting story is what it does for the analytics team itself.

Analysts spend a lot of time on questions the data already answers: what were sales last month, which region grew the most, who are our top 10 customers. Those get intercepted at the chat layer. What lands on the analyst's desk is the harder stuff: questions that require new data or new models. That's a better use of analyst time and a more accurate signal of where the data platform needs to evolve.

Natural-language access also puts healthy pressure on data quality. When a dataset is queryable in plain English, every measure name and every column becomes a user-facing surface. Bad measure names get caught. Inconsistent category labels get flagged. A dataset is now a product, no longer just a backend.

There's a benefit for leaders, too. Most BI portfolios suffer from a heavy long tail of dashboards that were built once, used twice, and never decommissioned. With NLP access, leaders ask the question they have against the dataset they need, without first having to find the right report. What gets consumed is the dataset itself.

What's under the hood

The action is a Glean Custom Action that takes the user's natural-language question, generates a DAX query, executes it against the target Power BI dataset, and returns the result inline in chat. Permissions follow the user's existing Power BI access; if a user can't see a dataset in Power BI, they can't query it through Glean.

For analytics teams, setup is straightforward:

  • Connect the action to your Power BI tenant
  • Specify which datasets are queryable, starting with the well-modeled, well-named ones
  • Make the action available to relevant Glean agents or expose it directly in chat

Once it's live, any authorized user can query in natural language, and the action handles the DAX generation and execution. Combined with Glean's Power BI connector which indexes report metadata so users can find the right dataset to query in the first place, the two together close the gap between locating a report and getting an answer from it.

The bigger pattern

For a lot of analytics functions, the default question has been "can someone build me a dashboard for this?" When the answer to "can I just ask the dataset?" is yes most of the time, analysts get their time back for the work that actually needs them, and the business gets answers in seconds.

If you want to see this running against your own Power BI environment, get in touch. We'll walk through what a phased rollout looks like for your most-used datasets.