Your BI investment is fine. The queue in front of it is not.

Most executive frustration with data is not about the data. The warehouse is populated, the dashboards exist, the analysts are capable. The frustration is the queue: a question becomes a ticket, the ticket becomes a dashboard, and by the time the dashboard arrives the decision has usually been made without it.

The cost of that queue never appears on a budget line, because it is the questions that go unasked. When an answer takes two weeks, people stop asking anything that is not worth two weeks — which excludes most questions, including the early, half-formed ones that lead somewhere. Analyst teams feel it from the other side: hired to find insight, they spend their weeks servicing a reporting backlog, and the reporting backlog is what talented analysts eventually leave.

What AI actually changes

A business leader asks a question in plain language and gets the answer now — then sees it was not quite the right question, and asks a better one. Ask, refine, act. That loop is how understanding is actually built, and it is the loop the reporting queue has never been able to support, however good the dashboards were. A dashboard answers the question someone anticipated months ago. The loop answers the question you have now, and then the one after it, which is usually the one that matters.

None of this requires replacing what exists. Years of work went into modelling the data, building the pipelines and getting a single version of the truth into the warehouse — and that investment is precisely what AI needs underneath it. The pattern that works is additive: AI sits on top of the platform you have, as a new way in, while everything that already works keeps working. The dashboards that earn their place stay too, with their role upgraded — they stop being the end of the process and become the start of an investigation. See something in the number, ask about it, keep asking until you understand it.

The one thing that decides whether it works

There is a hard technical fact under this, and it is worth a CEO knowing because it separates the deployments that build trust from the ones that destroy it. Pointed at raw database tables with no business context, AI answers questions correctly less than half the time — public benchmarks on realistic enterprise databases have landed frontier models well under 50%, and one 2026 analysis put real-world enterprise query coverage near a fifth. Grounded in a semantic layer — the defined business terms your company actually runs on — current benchmarks reach 98% and above.

The difference is not a smarter model. It is teaching the model what your numbers mean before letting it answer: which of the four revenue figures is the one the board uses, what counts as an active customer, when the month closes, how returns are treated. Every company has this vocabulary; almost no company has written it down anywhere a machine can read. That work — defining the terms once, centrally, so every answer uses the same definitions — is unglamorous, finite and entirely decisive. Skip it and you get a confident chatbot that is wrong often enough that nobody trusts it twice. Do it and the loop becomes trustworthy enough to put in front of the board.

The organizations that win this are not the ones with the most dashboards. They are the ones where asking became cheap.

What changes for the people

The common fear is that this displaces the analysts. The observed pattern is closer to a promotion. When routine questions answer themselves, the analyst function moves up the stack — from producing reports to owning the semantic layer, pressure-testing the AI's answers, and doing the deep investigative work that was never possible from inside a ticket queue. The scarce skill stops being "can operate the BI tool" and becomes "understands what the business is actually asking."

Governance moves up the agenda for the same reason. When anyone can ask, the definitions have to be right once and centrally — a wrong number delivered instantly to fifty people is worse than a slow one delivered to five. The companies doing this well treat the semantic layer like they treat the chart of accounts: owned, versioned, and changed deliberately.

Questions worth asking your own team

Three, in order. How long does it take a commercial question to become an answer today — honestly measured, ticket to insight? If we put AI over our data tomorrow, what would it believe "revenue" means, and who decided? And which of our dashboards would we keep if every leader could simply ask?

The first measures the queue. The second measures readiness. The third usually shortens the list considerably — and the answers to all three cost nothing to find out.