Within a few months, generative AI has gone from a curiosity to an everyday tool, and executives have noticed what it does well. They type a question in plain language and get a fluent answer in seconds. The next request to the BI team follows naturally: why search through several dashboards to find out why margin fell in the West, when you could simply ask?
BI tools have offered natural-language questions for years, with mixed results. What is changing is the quality of the language models behind them, and the expectations of the people asking. For BI teams, this is less a feature to switch on than a change in what they build and what they are accountable for.
Dashboards answer the questions someone anticipated
A dashboard is a set of answers to questions that someone predicted months ago: how sales compare with plan, where margin is under pressure, which customers are paying late. It does this well, and it should keep doing it. Where dashboards struggle is the next question: why did the number move, which products drove it, how does it compare with the same period last year?
Today those follow-ups usually travel by email to an analyst, who rebuilds the context, reruns the numbers and replies a day or two later. By then the meeting is over, and the decision was made without the answer. This gap between seeing a number and understanding it is where natural-language questions are most valuable.
Anticipate the questions, not just the pages
Before choosing a tool, collect the questions. Read the follow-up emails sent to the analytics team, sit in on a few business reviews, and ask executives what they looked for last month and could not find. A few dozen real questions, written in the words people actually use, are worth more than any feature list.
Group them by type and check that the data model can support each one. Questions about change need comparisons over time. Questions that break a number down need clean hierarchies for products, regions and customers. Questions about plan need the plan in the same model as the actuals. The question bank then becomes a test set: each question gets an expected answer, and every change to the model is checked against it. Keep the bank alive after launch: the questions people actually ask will differ from the ones they predicted, and the log of real questions is the best guide to what the model needs next.
Make every definition unambiguous
A dashboard hides ambiguity behind its layout. The page title, the filters and the analyst who built it tell the reader that “revenue” means net revenue after returns and discounts. A typed question carries no such context. Ask for revenue last quarter, and the system has to choose between gross sales, orders booked, invoiced sales and net revenue. It will not always choose the one finance reports.
The fix is a discipline that good BI teams already apply, made explicit for every business term:
- One certified definition. A single measure per term, with a named owner, and the alternatives hidden or renamed so that nobody can mistake them for it.
- A plain-language description. What the measure includes and excludes, written for the executive rather than the developer.
- The words people use. Sales, revenue and top line mapped to the same measure; bookings mapped to orders, not to revenue.
- Sensible defaults. The period, currency and scope assumed when a question does not specify them.
Where a term remains genuinely ambiguous, a well-designed system asks the user to clarify rather than guessing.
Show where every answer comes from
An answer without a source is an opinion. Every answer should show the measure it used, the filters and period applied and when the data was last refreshed, with a link to the governed report where the number can be checked. Executives test a new system by asking questions whose answers they already know. When the source is visible, a mismatch can be explained in a minute. When it is not, one wrong number can undo months of trust.
Answers must also respect who is asking. If a regional manager sees only their region in a dashboard, a question must not reveal the other regions. Access rules belong in the data model, where every report and every question inherits them, not in individual pages.
Just as important is knowing when not to answer. If a question falls outside what the model covers, such as a month that has not yet closed or a measure that does not exist, the right response is to say so and point to the person who can help. A system that admits its limits is trusted with the questions it can answer.
Keep dashboards for monitoring
None of this makes dashboards obsolete. Monitoring depends on consistency: the same measures, in the same place, compared the same way every week, so that patterns and exceptions stand out at a glance. A question cannot do that job, because nobody asks about a problem they have not yet noticed.
The two work best as a pair. The dashboard shows that something moved; the question explains why. Design them together: each dashboard page should anticipate its most common follow-ups, and a question that people ask again and again is a sign that a page is missing or unclear. A good test of the design is whether an executive can go from a red number on a dashboard to a trustworthy explanation without writing an email.
What changes for the BI team
The work moves upstream. Fewer one-off pages are needed, because a follow-up no longer requires a new report. More effort goes into definitions, descriptions, synonyms, test questions and a regular review of the questions that failed. The skills that matter most are a command of the business vocabulary, clean data modelling and the habit of testing answers the way finance tests a month-end close. None of this is glamorous, but it is the difference between an impressive demonstration and a tool executives use every week.
Start small: one domain, one executive team, one well-governed dataset and the questions that team actually asks. Prove the answers there before widening the scope. Teams that work this way find that their dashboards improve as well, because the clarity that natural-language questions demand serves every report.