Start with the decision the AI is supposed to change

A logistics operations lead told me about a stretch where every AI idea that reached her desk opened the same way. Someone would say the company was sitting on a mountain of shipment data, and surely there was a model in there somewhere. She would nod, ask a few questions, and watch the idea dissolve. Nobody could tell her what anyone would do differently once the model existed.

She started running the conversation backward. Before anyone talked about data or models, she asked one question. What decision is this supposed to change, and who makes that decision today?

It sounds obvious until you try to answer it out loud. Most AI proposals cannot. They point at a capability, not a decision. “Predict delays” is a capability. The decision is a dispatcher at four in the afternoon choosing whether to hold a truck for a late inbound load or send it out half full. That person has a name. They make that call dozens of times a shift, on instinct and a whiteboard, and they are usually a little wrong in a direction that costs money.

Once she had the decision, the whole shape of the project changed.

The decision tells you almost everything

When you write the decision down, you get the scope for free. You know who the user is, because it is the person making the call. You know when the feature has to produce an answer, because the decision happens at a specific moment. You know what “good” means, because you can compare the AI’s suggestion against what the dispatcher would have done anyway.

You also find out fast whether AI is even the right tool. A friend who runs support at a software company had the same experience from the other side. Her team wanted a model to route tickets. When she asked which decision that changed, the honest answer was that a rule based on the customer’s plan tier already made the call correctly nine times out of ten. There was no decision worth a model. There was a rule that needed cleaning up. Naming the decision saved her a project she did not need.

The data-first approach hides this. You can spend a quarter cleaning and joining tables and never once say out loud what anyone will do with the result. The decision-first approach forces the question early, when it is cheap to walk away.

Here is the practical version of what the logistics lead now asks before any AI project moves forward:

What is the specific decision this feature changes?
Who makes that decision right now, and how?
How often do they make it, and how wrong are they today?
What would they do differently if the AI were right?
What happens the times the AI is wrong?

If a team cannot answer those, they do not have a project yet. They have a wish about their data.

The decision sets the accuracy bar too

There is a second reason to start here. The decision tells you how good the model has to be, and that number is rarely the one people assume. AI features usually need to clear about 85 percent accuracy before their mistakes stop compounding into something worse than the manual process. But the real bar comes from the decision, not from a benchmark. If a dispatcher is right 70 percent of the time on hold-or-send calls, a model at 80 percent is a genuine improvement. If a fraud reviewer is right 98 percent of the time, a model at 80 percent is a step backward that will quietly cost you.

You cannot know which situation you are in until you have named the decision and looked at how the human does it today. The data does not tell you that. The person does.

How we approach it at Density Labs

When a company brings us an AI idea, the first thing we do in the AI Opportunity Assessment, our two-week fixed engagement at $2,500, is refuse to talk about the data. We ask them to walk us to the decision. We find the person who makes it, watch how they make it now, and write down what would have to change for the AI to matter. Half the value shows up in that room, before a single line of code. Some of the time we tell a client the honest thing, that the decision does not need a model, and that conversation saves them far more than the engagement costs.

Do not start from the data you have. Start from the one decision a real person makes, and work back to whether AI belongs anywhere near it.