Your dashboard already told you what happened. It still can't tell you why.
The chart was right. That was the problem.
A regional revenue line had dropped nine percent month on month. The dashboard rendered it accurately, on time, in the correct colour. Someone screenshotted it into a chat thread with a question mark. Then four working days disappeared — not into fixing anything, into finding out what had happened. By the time the answer came back (two clients, one of whom had quietly moved a facility elsewhere in March), the meeting that needed it had been held without it.
I build agentic AI systems for a living, and I’ve spent a good part of this year sitting with this exact pattern. The least interesting thing about it is also the most important: the BI stack didn’t fail. It did precisely what it was built to do.
A dashboard is a frozen answer
Every dashboard is a saved response to a question somebody asked once. Usually a sharp question, asked by someone who understood the business. Then it got built, and scheduled, and it started answering that same question every morning forever.
Which is valuable, right up until the number does something unexpected. Because the interesting question is never the one the dashboard was built for. It’s the one that comes after — is that everywhere or just there, was it like this last year — and about that, the dashboard has nothing to say.
So you go and find a person.
The loop nobody measures
Someone senior asks why. An analyst goes and looks. The answer arrives two days later and immediately raises two more questions, because good answers do. Somewhere around the third round trip, either the question gets answered or the moment passes and everyone moves on.
What’s being consumed here isn’t compute. It’s calendar. The delay is made of handoffs — waiting for someone to be free, waiting for the next meeting, waiting for a follow-up nobody got round to. I’d call it organisational latency if that didn’t sound like a consultant made it up.
And the real loss isn’t the analyst’s afternoon. It’s the decisions taken in the meanwhile, on the assumption that things were roughly fine, because the analysis wasn’t back yet. Those never appear as a line item. That is what you’re paying for the gap.
Why a chatbot on the dashboard doesn't close it
The obvious move in 2026 is to put natural language on the front of the BI tool. Ask in English, get a chart. I’ve built versions of this. It demos beautifully, and it mostly speeds up the part that was already fast.
Those four days weren’t spent retrieving numbers. They were spent deciding which numbers were worth retrieving, then working out what the ones that came back actually meant. That’s investigation, not query. “Why did revenue fall” isn’t one question — done properly it’s thirty or forty. By product, by client, this year against last, actual against plan, price against volume, with and without the one large deal that distorts everything it touches.
Most come back boring. That’s the point. The skill an experienced analyst has is knowing which handful of cuts to try first, recognising a dead end quickly, and knowing when to stop looking. None of that lives in the dashboard, or in a chat box sitting on top of the dashboard.
Answering “why” properly takes three unglamorous things. Decomposition — turning one vague question into forty checkable ones and throwing most of the results away. Real knowledge of the data, not the schema but the semantics: which measures can be summed across months and which absolutely cannot. And evidence. “Two clients drove the drop” is a claim, not an answer. It becomes an answer when you can see which two and how much each. Anything less and the person receiving it either trusts you or redoes your work, and in my experience they redo the work.
The distinction worth drawing
Business intelligence is descriptive. It tells you what happened, accurately, and stops there — by design, not by failure. Decision intelligence is the attempt to carry on: why it happened, what’s likely next, what to do about it.
The difference isn’t the technology. It’s what comes out the other end. BI’s output is a view — here is the data, you draw the conclusion. Decision intelligence’s output is supposed to be the conclusion itself, with the evidence attached so you can check it.
Supposed to be, anyway. Most of what’s sold under that heading right now is a language model with database access and a great deal of confidence. I tried building it that way first. Fantastic demo, unusable system, for reasons I’ll get into next week.
Three questions about your own organisation
- How long is it really, in working days, from "that number looks wrong" to "here's why"? Take the last five occasions and count.
- Who does that work — and what are they not doing while they do it? In every team I've seen it's the same two people, and they're the two you'd rather have on something else.
- And when the answer came back, could anyone see how it was reached? Or did it arrive as a sentence in an email that everyone accepted because of who sent it?
