Feasibility is three questions you can ask before kickoff

A CTO at a healthtech startup told me he used to treat feasibility as a project of its own. An AI idea would come up, and he would commission a study, spin up a small team, spend a few weeks and a chunk of budget to find out whether the thing was even possible. Sometimes the answer came back no, and the weeks were gone.

He replaced most of that with three questions he can ask in a single meeting, before anyone is assigned, before any budget moves. They do not tell him the idea will work. They tell him fast whether it is dead on arrival, which is most of what an early feasibility check needs to do. The expensive study, when it is still warranted, is reserved for ideas that pass all three.

Question one: does the data exist

The first question is whether the data the model would need actually exists, and exists in a form you could get to. Not whether it could exist in principle. Whether it is there now.

Teams skip this because they assume the data is somewhere. Often it is not. The signal the model needs was never recorded, or it lived in a field people stopped filling in years ago, or it exists but is locked in a system you cannot legally pull from for this purpose, which in healthtech is a real and frequent wall. If the data does not exist, no model works, and every downstream conversation about accuracy and design is wasted. Ask it first because a no here ends things cheapest.

Question two: can a human do it from that data

The second question is whether a person, handed exactly the data the model would have and nothing more, could make the decision reliably. Not an expert with extra context in their head. A person with only the inputs the model would see.

This one catches a subtle and expensive failure. Sometimes the data exists but does not contain enough signal to support the decision, and the only reason humans manage today is that they bring outside knowledge the model will never have. If a nurse can only triage a case because she also remembers a conversation with the patient last week, and that conversation is nowhere in the data, then the model is being asked to do something the data cannot support. A human proxy test surfaces this in an afternoon. Hand a smart person the raw inputs, ask them to make the call, and watch whether they can. If they cannot, the model cannot either, and you just saved a project.

A data lead at an insurance carrier told me this test alone changed how her team scoped. They would take the exact fields a proposed model would use, give them to an experienced underwriter with everything else stripped away, and see if the underwriter could still decide. When the answer was no, they knew the problem was missing signal, not a weak model, and no amount of model work would have fixed it.

Question three: is a wrong answer survivable

The third question is what happens when the model is wrong, because it will be wrong sometimes. AI features generally need to reach about 85 percent accuracy before their errors stop compounding, which means you are always designing for a stream of mistakes. The question is whether your situation can absorb them.

In some places a wrong answer is cheap. A mediocre product recommendation costs a shrug. In others a wrong answer is a disaster you cannot walk back, and in healthtech that line is bright. If a wrong answer is survivable, you can ship earlier and improve in the open. If it is not, you either need a human checking every output, which changes the whole economics, or you should not build the feature at this accuracy at all. Answering this early tells you what kind of project you are in before you have committed to it.

The three together are a filter, not a guarantee

Passing all three does not mean the idea will succeed. It means it is worth the more expensive look. Failing any one means stop, and stopping in a meeting costs nothing compared to stopping after a quarter. The CTO’s point was that most bad AI projects die on one of these three, and all three are answerable cheaply, so there is no excuse for finding out the hard way.

How we approach it at Density Labs

These three questions are the opening moves of every AI Opportunity Assessment, our two-week fixed engagement at $2,500. Before we design anything, we check that the data exists, that a person could make the decision from that data alone, and that a wrong answer is something the business can survive. Plenty of ideas do not clear all three, and telling a client that in week one, with the reason, is one of the most valuable things we do. It is far cheaper than the study they were about to fund.

Before you kick off an AI project, ask whether the data exists, whether a human could decide from it, and whether a wrong answer is survivable. Three cheap questions will spare you most of the expensive mistakes.