Category: English

English | Tiếng Việt

Every few months the claim comes back: dashboards are dead. Why would anyone build reports when you can simply ask an AI about your data? Some go further: “Power BI is dead.” It is a good headline and bad advice. The strongest evidence against it comes from the vendors themselves — and from your car.

What the vendors are actually doing

If chat were replacing BI, you would expect Microsoft and Databricks to be winding their BI stacks down. The opposite is happening.

Microsoft is building Copilot into Power BI: “Chat with Your Data” is a chat experience that answers questions about a report or semantic model — it sits on top of the BI stack, not beside it. The old Q&A feature is being retired in favor of this, and Microsoft’s own guidance is blunt about what makes the chat good: a well-prepared semantic model, with clean structure, clear names, and descriptions. Databricks follows the same pattern with Genie: natural-language questions answered over curated, governed data with business context attached.

In other words: the industry is not replacing the dashboard with chat. It is building chat on top of the same foundation the dashboard uses.

The car dashboard test

Think about how you drive.

Your car has a dashboard: speed, fuel, engine temperature, tire pressure. You never ask for these. You glance. The dashboard exists precisely so that the questions you have all the time are answered before you ask them. That is what a dashboard is for: it removes the cognitive load of asking the same standard questions again and again. The KPI is just there.

Now imagine the opposite: a car where you must ask, out loud, “how fast am I going?” every time you want to know. Nobody would buy that car. That is what “replace all reports with chat” actually proposes.

But the dashboard has a limit, and you know it from driving too. “With the fuel I have left, can I still make it to Đà Lạt?” No dashboard has a gauge for that. It is an ad-hoc question: it depends on context, and you ask it rarely. This is exactly where a conversational layer shines — you ask, in your own words, and get an answer computed for your situation.

Two interfaces, one set of sensors:

• The dashboard answers the standing questions: the twenty KPIs you check every day, identical for everyone, zero effort to consume. Revenue this month, unpaid invoices, stock level, production output.
• Chat with your data answers the long tail: the one-off, the drill-down, the “why is this number different from last March?” You would never pre-build a report page for each of those; there are too many.

Using chat for your daily KPIs is asking the car for your speed. Cramming two hundred visuals into a report to pre-answer every possible question is printing the whole owner’s manual on the dashboard. Each tool has its job.

The real headline: the semantic model matters more than ever

Here is the part the “dashboards are dead” crowd misses. Both surfaces — the report and the chat — read from the same semantic model: the layer that defines what “revenue” means, how the tables relate, which measure is the official one.

In a car, the speedometer and the range computer read the same sensors. If a sensor is broken, both lie to you. It is the same with data: if the model is messy — duplicated logic, unclear names, no descriptions — the dashboard misleads quietly, and the chat misleads confidently. An AI answering questions over a bad model does not say “your model is unclear”; it gives you a fluent, plausible, wrong answer.

That is why every serious guide to AI-in-BI says the same thing: prepare the model. Clean star schema. One measure per business definition. Descriptions and metadata on tables, columns, and measures, so both a human and an AI know what things mean. Certified, governed models rather than one export per department.

So the stack does not shrink; it re-orders. The semantic model moves from a technical detail to the most valuable data asset a company owns. Build it once, build it well, and it feeds both surfaces — every report today, and every AI feature tomorrow.

What this means in practice

For a small or medium business, the practical order is:

1. Start with the reports. Get your whole business — finance, sales, operations — onto one clear, trusted set of reports, built on a clean semantic model. This is where the immediate value is, and it forces the foundation to be right.
2. Add chat as the next level. Once the model is trustworthy, a natural-language layer on top is a modest step, and the answers inherit the model’s quality — including in Vietnamese.
3. Be suspicious of chat-first offers. If someone promises “just ask your data anything” without talking about the model underneath, ask them what happens when two tables disagree about revenue. The chat interface is the visible 10 percent; the model is the 90 percent that decides whether the answers are true.

The dashboard is not dead. It got a colleague. And they both depend on the same thing: a semantic model that deserves the trust you put in it.