Your dashboard isn’t wrong. It’s finished.

That’s from a VB Pulse survey of 101 companies in July. Over the six months prior, 68% traced a confidently wrong AI output back to missing business context.

VentureBeat asked the same question in June and got 57% and 31%. The rate is going up, not down.

I read that and thought about a lunch I sat through a few weeks earlier.

Small conference room, sandwiches, the new lead quality dashboard up on the big screen.

The ops guy who built it came in to present it himself. Young, sharp, and genuinely excited. He’d put real work into this and you could tell. It was good work, too. Clean build, fast, everything rendering the way it should.

The CEO was impressed and said so. Then he started digging into the actual numbers, and about two minutes in, he stopped.

A set of data points had never been pulled in. Nothing hidden, nothing hard to get, just never included. Their absence didn’t make the dashboard look wrong. It made it look finished.

Then somebody at the table asked the harder version of the same question:

Nobody answered. People looked at each other. And you could feel the whole room arrive at the same conclusion at the same second: we have to start this over.

The kid’s face was the worst part.

Nobody in that room failed

He built exactly what was asked for. The request was clear, the tools worked, and every person in the chain did their job well. The result was still useless.

The problem happened months earlier, in whatever conversation kicked the project off. What the CEO said then was some version of here’s the information I want, here’s the output I want to see.

That’s an answer. Nobody asked the question underneath it: what will I do differently depending on what this tells me?

And honestly, why would they? A confident request doesn’t sound like something you’re supposed to poke at. It sounds like clarity, like a leader who knows exactly what he wants. Everyone in that kickoff walked out feeling good about it. The assumption buried in the ask was that he already knew which numbers mattered, and he didn’t, and he couldn’t have. That’s the whole reason he wanted a dashboard in the first place. He was asking someone to answer a question he hadn’t worked out yet, and it sailed straight through, because that always looks like good direction.

Why this keeps happening now

The failure itself is old. What’s changed is how fast it compounds.

There used to be friction between asking for something and getting it. Two weeks of build time. A junior analyst asking a dumb-sounding question that turned out not to be dumb at all. A meeting where somebody said “wait, what’s this actually for?” three days in, before anyone had wasted a month.

That friction was annoying, and it was also quietly checking whether the question was any good. Nobody put it on the org chart. So nobody noticed when it disappeared.

It won’t get bored, won’t push back, won’t wonder what you plan to do with any of this. Hand it a half-formed request and it hands back something polished and complete-looking, fast.

A survey out this week put a number on how thin the ground underneath all this really is. More than 540 data leaders across 66 countries, and only 8.4% said the data feeding their AI systems was trustworthy enough for production use. Meanwhile 57% of them are already running or piloting agents on it.

An agent moves faster than a person. So when the information is incomplete, you don’t get to see the wrong decision coming. You just get there sooner.

One unasked question used to produce one bad quarterly report. Now it produces twelve clean-looking outputs a week.

That’s why this feeling is so hard to name. There isn’t one broken thing to point at. It’s a single unasked question showing up in a dozen places at once.

If you feel it, you’re ahead

There’s a detail in the June wave of that survey almost nobody quoted.

The July wave found the same inversion on repeat failures: 50% for the companies building the fix, 21% for the ones who aren’t.

Read quickly, that looks like the fix isn’t working. It’s the opposite. The companies who’ve felt this are the ones doing something about it, and the ones who haven’t felt it yet simply haven’t been burned yet.

That nagging feeling isn’t evidence you’re behind. It’s an early warning, and it’s showing up right on schedule.

So before the next dashboard, the next agent, the next tool, ask this:

If you can’t answer the second half, you’re not ready to make the request. The capability is there. The order is wrong.

I should be upfront about something. I see this more than most people do, and not because I’m smarter about it. It’s what companies call me about. That’s a skewed sample, and every CEO I sit down with already suspects something is off before I walk in.

But the skew cuts both ways. Thirty years in, I have almost never found the problem where the CEO expected it. It’s rarely the tool and it’s rarely the team. It’s the order things got built in.

That’s why we called it Foundation First.

What’s yours? What did you build, correctly, that answered the wrong question?

Sources: VB Pulse survey, July 2026 — 101 enterprises with 100+ employees; the 68% and 37% figures are from the July wave, the 57%/31% comparison and the 78%-vs-20% split from the June wave, published by VentureBeat July 10, 2026. Both waves asked the same question. Trustworthy-data figures from The Modern Data Company’s third annual Modern Data Survey, interim findings released August 20, 2026 — 540+ responses across 66 countries. “The rate is going up, not down” and “you just get there sooner” are the author’s framing, not survey claims.