From "What Happened?" to "Why Did It Happen?" - Resolving the Dashboard Fatigue

We live in the age of data abundance. Most modern companies have dashboards showing revenue trends, funnel conversion rates, weekly active users, churn, and dozens of other metrics. Yet despite all this visibility, a common frustration persists across teams:

"We can see that orders dropped this week but no one can tell us why."

This is dashboard fatigue. It's not a lack of data. It's a lack of answers. Analysts spend a lot of time slicing, dicing, and cross-referencing charts to explain movements that a business stakeholder needs to act on. The data is available but the insight isn’t.

Traditional BI tools are excellent at answering descriptive questions: What happened? When did it happen? How much? But they stop short of the diagnostic: Why did it happen? And what should we do about it?

This gap has real consequences. Delayed decisions. Misallocated budgets. Teams chasing the wrong problem. The analyst becomes a human search engine and that's a poor use of anyone's time.

Metabase is one of the most widely used open-source business intelligence platforms. It lets teams connect to their databases and data warehouses and build dashboards, run queries, and visualize data.Metabot is Metabase's AI layer which can be used with Metabase interface and integrated with Slack. Metabot allows users to ask data questions in natural language. Instead of opening a dashboard and hunting for a chart, one can simply type: "How many orders did we receive last week compared to the week before?" Metabot interprets the question, queries the underlying data, and returns charts, tables, or trend summaries.

But as powerful as Metabot is at retrieval and description, we discovered something during the #MetabaseAI Hackathon: it answers the "what" extremely well but struggles with the "why”.

To understand why this matters, it helps to think about the hierarchy of analytical questions. Most BI tools today including Metabase out of the box live firmly in the descriptive tier. Our hackathon experiment was about pushing into diagnostic territory.

Descriptive: What happened? "Purchases dropped 18% this week."

Diagnostic: Why did it happen? "Purchases dropped due to a holiday weekend and a paused ad campaign in Germany."

Predictive: What will happen? "If the campaign stays paused, we'll miss our monthly target by 12%."

Prescriptive: What should we do? "Reactivate the campaign and add a promotional offer to compensate."

Most BI tools stop at the first tier. Getting beyond is what is aimed to achieve.

Rather than hoping Metabot would magically infer the root cause of a metric movement, we designed a step-by-step interaction pattern that guides it toward diagnosis.

Step 1: Start with the symptom

A business user asks in Slack: "Why did orders drop this week?" Metabot responds with an initial overview : total order volume, week-over-week change, and a few surface-level hypotheses. This is a pure description - the "what."

Step 2: Break it down by category

We prompt Metabot to segment the data: "Break down the order drop by product category. Which category changed the most?" Metabot returns: Electronics dropped 34% week-over-week, while other categories held relatively flat.

Step 3: Isolate the geography

We narrow further: "For the Electronics category, which regions account for the drop?" Metabot identifies where the decline is concentrated.

Step 4: Check the mechanism

We ask Metabot to compare pricing and volume: "In those regions, did average order value change, or is this a volume issue?" Average order value is stable - this is a pure volume drop. Fewer transactions, not smaller ones. Pricing is ruled out as the cause.

Step 5: Generate the decision-ready summary

We prompt for a synthesis. The final output:

We did not build a new AI model, fine-tune anything, or create a custom data pipeline. The architecture is deliberately simple.

Slack: The interface where business users ask questions in natural language, without needing to open any dashboards.

Metabase: The BI layer. Metabot handles data access, querying, and initial visualisation. The semantic layer defines what the data means. If the semantic layer is poorly defined, ambiguous, or inconsistent, the AI will produce unreliable or outright wrong answers. This is one of the primary causes of AI "hallucinations" in data contexts: not a failure of the model, but a failure of the underlying data structure to communicate what things actually mean.

Anthropic AI: It powers the structured reasoning layer interpreting Metabot's outputs, generating follow-up questions, and synthesising the final insight.

Structured Prompt Flow: A sequence of pre-designed prompts that guide the AI from symptom to segment to root cause to recommendation. This is the "playbook" for diagnosis.