Volume, variety, and value: triaging a first AI use case

A head of product at an analytics vendor came to us with a list. Twelve AI ideas, all plausible, all championed by someone internally, all fighting for the same first slot. She had spent two meetings trying to rank them and the meetings kept ending in a tie. Everyone had a favorite, and every favorite had a good story, so the list did not move.

The problem was not a shortage of ideas. It was the lack of a shared way to compare them. Twelve stories cannot be ranked against each other, because each one sounds best when its champion tells it. What she needed was three questions she could ask about every idea, the same way, so the comparison stopped being about who argued hardest.

We gave her three. How much volume. How varied are the inputs. How much value. She ran all twelve ideas through those questions in an afternoon, and the list that had been frozen for two weeks sorted itself.

The three questions

Volume is how often the task happens. A task that runs ten thousand times a month has room to pay back the work of building for it. A task that runs twice a quarter rarely does, no matter how annoying it is each time. AI has a fixed cost to build and a per-use payoff, so frequency is the first filter. High volume is what makes the build worth it.

Variety is how different the inputs are from each other. A task where every input looks roughly the same is a task a model can learn to handle reliably. A task where every input is a snowflake, wildly different structure, tone, and edge cases each time, is far harder and far more likely to produce confident wrong answers. Low variety is easier and safer. High variety is where projects get stuck.

Value is what a correct answer is worth, and what a wrong one costs. Some tasks save a few seconds. Some prevent an expensive mistake. Some are so low stakes that even a perfect model would not move anything that matters. High value is what makes the whole thing worth attention.

The best first use case sits in one corner: high volume, low variety, high value. Happens constantly, looks similar every time, matters when you get it right. That corner is where AI is both easiest to build and most worth building.

Why the triage beats the argument

Her twelve ideas spread out fast once she scored them. The champion’s favorite, an AI feature to write custom analyses for enterprise clients, turned out to be high value but also high variety and low volume. Every client wanted something different, and there were not that many of them. Interesting, and a terrible first project. It would have been the hardest possible thing to build and the slowest to pay back.

The idea that won was one nobody had been excited about. Automatically categorizing incoming support questions so they routed to the right specialist. High volume, it happened all day. Low variety, the questions clustered into a handful of types. Real value, misrouted questions wasted expensive people’s time. Boring, and exactly right for a first build.

A director of engineering at a media company ran the same triage on his own backlog and had the same experience. The flashy idea scored badly on variety. The dull idea scored well on all three. He built the dull one, it worked, and it earned the credibility to fund the flashy one later.

How we approach it at Density Labs

In the AI Opportunity Assessment, our fixed two-week engagement at $2,500, scoring candidate use cases on volume, variety, and value is one of the first things we do, because a first project that scores badly on any of the three tends to fail slowly and publicly. We would rather find the boring high-scoring use case in week one than watch a team spend a year on the exciting one that never had the shape to work. The first win should be the safe bet, and the safe bet is usually the one nobody was arguing for.

Pick the use case that happens often, looks the same each time, and matters when it is right. Save the snowflake for later.