The one metric that decides whether the feature is worth building

A head of analytics at a national retailer told me about the question that quietly killed half the AI proposals that came across her desk, and saved the other half from themselves. She asked it early, before scope, before budget. If this works, what one number moves?

Not a list of benefits. Not “efficiency” or “customer experience.” One number. The single metric that would be different if the feature succeeded, and would stay the same if it failed. She wanted the team to name it out loud and commit to it.

It sounds like a reporting formality. It is actually the sharpest scoping tool she had.

The metric forces you to know what the feature is for

When a team can name the one metric, a lot follows. They know who cares about the feature, because someone owns that number. They know how big the opportunity is, because they can look at where the metric sits today and imagine where it could go. And they know when to stop, because they can watch the metric and see whether the feature actually moved it.

When a team cannot name the metric, that failure is the finding. It usually means one of two things. Either the feature does not connect to anything the business measures, in which case why are you building it. Or the value is real but so diffuse that no single number captures it, which means you have not thought hard enough about what specifically improves. Both are worth discovering before you spend the budget, not after.

She gave me an example. A team proposed an AI feature to summarize product reviews on each item page. It sounded nice. When she asked for the one number, the room went quiet, then offered “engagement.” She pushed. Engagement how? Eventually they landed on something real. The metric was the rate at which a shopper who read reviews went on to add the item to their cart. That was a number they could see today and a number the summary could plausibly move. Now the project had a purpose and a test. Before that question, it was a nice idea with no way to tell if it worked.

One metric, not a dashboard of them

The discipline is picking one, and that is harder than it sounds. Teams want to list five metrics because each one is partly true and no one wants to leave a benefit off the slide. But five metrics is the same as zero. When the feature ships and two go up while two go down and one does not move, you have no way to say whether it worked. You will argue. The one metric is the tiebreaker you agree on in advance, when nobody is defending a project they have already sunk months into.

Picking one also disciplines the scope. Once the retailer’s team committed to the add-to-cart rate for review readers, the scope organized itself. The summary had to be genuinely useful to someone deciding whether to buy, which meant surfacing the tradeoffs in the reviews, not just the average star rating. Features that would not move that number, like making the summary longer or prettier, dropped away on their own. The metric did the cutting.

A finance lead at a subscription company told me she uses the same test in a harsher form. She asks what number moves, and then by roughly how much it would have to move for the project to be worth its cost. If the honest answer is that the metric would move too little to pay for the work, the project dies right there, cheaply, which is the best time for a project to die. Given that MIT’s 2025 research found around 95 percent of enterprise generative AI pilots show no measurable return, a hard number gate at the start is not pessimism. It is how you end up in the surviving five percent.

How we approach it at Density Labs

In the AI Opportunity Assessment, our fixed two-week engagement at $2,500, we do not let a project through without its one metric. We make the client name the single number that would move, find out where it sits today, and get honest about whether the feature could move it enough to matter. Sometimes that conversation ends a project on day two, and the client thanks us, because we just saved them a quarter spent building something with no way to prove it worked.

Before you scope an AI feature, name the one number it would move. If you cannot name it, you are not ready to build it, and the missing number is telling you something.