Generative BI or a dashboard: which do you need?
Start with management questions and data quality. Learn when a conversation with your data is useful and when a clear report is enough.
Start with the question, not the interface
If managers review revenue, overdue payments and the sales plan in the same format each week, a dashboard is convenient. Its structure is stable and the figures are comparable. Chat does not have to replace it. Consider Generative BI when each answer leads to another question: why margin changed, which products contributed, or what happened in another channel.
One metric needs one rule
Imagine a retailer. CRM revenue follows signed deals, accounting follows payments, while management expects a figure after returns. AI does not automatically resolve the disagreement. Define the metric, period, currency, inclusions and exclusions first. Then connect sources without counting the same order twice. This is an illustrative scenario, not a client case.
Run a limited pilot
Choose one process and several management questions. Find the correct answers manually and record the sources. Test a simple fact, period comparison, filter and more complex rule. Include questions with missing data and questions from users without permission. The expected response is to state the limitation. A polished answer is not proof of a correct calculation.
What to inspect in an answer
The user should understand the reporting period, filters and data source. Whether the query and calculation are visible depends on the platform and configuration; check this when selecting the solution. An incomplete source should be disclosed. A “why” question often still requires human investigation: correlated changes do not establish a cause.
When to postpone implementation
If data cannot be accessed, metrics are undefined or nobody can verify answers, start with preparation. Conventional BI may cover recurring reports. For varied new questions over prepared data, a Generative BI pilot can be worthwhile. Evaluate useful answers, time to obtain them and verification cost rather than the presence of an AI label.