The discovery question that kills half of AI ideas: what if it’s wrong

An IT director at an industrial distributor told me he got tired of watching AI proposals die in production instead of in a meeting. The pattern was always the same. Someone would bring an idea, everyone would get excited about what it could do, the team would build a version, and then it would fall apart the first time it gave a confident wrong answer to a customer or a warehouse crew. By then the money was spent.

So he added one question to the front of every discussion. Before anyone described what the model would do, he made them answer this. What happens when it is wrong?

Not if. When. He treated a wrong answer as a certainty, because it is, and asked the room to describe the specific bad day it would cause.

A wrong answer is not one thing

The reason this question works is that it forces people off the happy path and onto the failure they have been avoiding. Some wrong answers are cheap. If a model suggests a slightly worse route for a delivery, a dispatcher glances at it, ignores it, and moves on. The cost of that mistake is a shrug.

Other wrong answers are expensive in ways nobody wants to say out loud in a kickoff meeting. He gave me an example. A team wanted AI to tell field techs which replacement part a machine needed based on a photo and a description. Sounds great. Then he asked his question. What happens when it names the wrong part? The tech drives ninety minutes to a site, opens the box, and has the wrong component in his hand. Now the customer’s line is down longer than if nobody had helped at all. The wrong answer was not a shrug. It was a worse outcome than doing nothing.

That project did not die because the technology could not work. It died because the cost of a wrong answer was too high for the accuracy they could realistically reach, and naming that in week one saved a quarter of building.

The question sorts your ideas for you

He described what happened once the whole team internalized the question. AI ideas started sorting themselves into two piles without much argument.

Ideas where a wrong answer is caught cheaply, by a human who was going to look anyway, before anything acts on it.
Ideas where a wrong answer flows straight into something costly, physical, or hard to reverse, with no one in the middle.

The first pile is where AI belongs early. A wrong suggestion in a draft, a wrong tag on a document, a wrong first guess that a person confirms. The mistakes are visible and cheap, and the tool is a genuine help even when it misses.

The second pile is not off limits forever. It just needs a different design, one that assumes error and builds a check around it, and it needs a much higher bar before it goes anywhere near a customer. A friend who runs operations at a parts manufacturer put it well. She said her rule is that AI can suggest, but if being wrong costs more than a phone call to fix, a person confirms before it acts.

The question also protects you from a quieter trap. AI features usually need to clear about 85 percent accuracy before their mistakes stop compounding into something worse than the manual process. If your idea sits in the expensive pile and you cannot get near that bar, you do not have a project. You have a liability with a nice demo.

How we approach it at Density Labs

In the AI Opportunity Assessment, our two-week fixed engagement at $2,500, one of the first things we do is put a price on a wrong answer for the specific idea in front of us. We walk through what the model produces, where it flows next, and who or what acts on it before a person could catch a mistake. Sometimes that conversation ends the project, and we tell the client so. That is not a failed engagement. That is the cheapest possible version of finding out.

Before you ask what an AI feature can do, ask what its wrong answer costs. Half of your ideas will not survive the question, and those are the half you did not want to build.