Will an agent that writes summaries replace the dashboard? Brilliant Earth is halfway there

Visualization usually sits at the very end of the data pipeline. Data lands in the warehouse, metrics get computed, and a chart is drawn for people to look at. Charts exist so humans can read numbers faster. But an LLM can take in far more numbers than any person. So will that last step turn from a dashboard into an agent that reads everything for you and tells you what it means?

The Brilliant Earth case Snowflake published on October 22 gets halfway to an answer.

Four data products, and only one is a fixed dashboard

Brilliant Earth runs Posit’s analytics platform as a native app inside Snowflake, with access following Snowflake roles. On it the team built four kinds of things: a scenario planner for its marketing mix model; store dashboards that replaced a legacy BI tool for more than 100 showroom users, wrapped in querychat so a manager can type “the last 14 days of bookings for my showroom”; weekly narrative reports written by an LLM; and throwaway Shiny dashboards built for one question.¹

Matthias Mueller, who leads analytics, says he doesn’t want to rebuild an ever-growing number of dashboards and would rather build custom solutions. The throwaway ones, he says, usually explain the problem better.¹ Fixed dashboards are still there, but they are no longer the only way out of the pipeline.

Let the LLM write; let code compute

The weekly report is the part to copy. Python and R compute the metrics straight from Snowflake data, and only after the numbers are validated do they go to an LLM to be written up. In Mueller’s words, the LLM summarizes in prose, but “mathematical functions” own the computation.¹

That is what makes a summarizing agent workable at all. An LLM reads a lot but does not reliably calculate. Let it narrate, let reproducible code supply the numbers, and every sentence can be traced back to a figure. It matches what I argued in the fast decision model post: numbers that must be right stay with SQL and code.

Summaries alone miss the shape

Could the agent skip charts altogether? Statistics has an old answer. In 1973 Anscombe built four sets of 11 points with the same means, variances, correlation and regression line.²

Hand an agent only those statistics and it writes one paragraph for all four. Plot them and you see a line, a curve, an outlier, and a correlation held up by one point. In 2017 Autodesk researchers took the idea further, starting from a dataset that plots as a dinosaur and generating a dozen more with the same statistics and different shapes. Their advice: never trust summary statistics alone; always visualize your data.³

So my answer: the fixed dashboard will fade, but visualization will not.

Charts stay, as evidence the agent draws on demand

The tools are already heading there. In June 2026 Posit gave querychat a visualize tool: when a question is better answered with a chart, the LLM writes a graphics query, the chart appears in the chat, and the same chart goes back to the LLM to interpret. Every tool only runs SQL, so each query is visible and reproducible.⁴

Don’t expect it soon, though. In 2021 a Gartner analyst predicted that by 2025 data stories would be the most common way to consume analytics, 75% of them generated automatically.⁵ In 2026 Brilliant Earth still keeps dashboards for 100+ users, and Mueller notes that “we’re all creatures of habit”: a new tool has to feel better than the old one.¹

For data teams: get the numbers right before deciding who tells them

• Keep metric definitions and calculations in code, and validate before any agent sees them. As Snowflake’s Sem Sergunin puts it, agents on unprepared data only “scale your mess.”¹
• Attach the query to every conclusion an agent writes, so it traces back to the data.
• When a conclusion depends on a distribution, a trend or an outlier, have the agent draw the chart instead of writing one sentence.
• For the few metrics people check every day, a fixed dashboard is still the fastest; don’t rush to retire it.