When the volume is too low for AI to ever pay off

The founder of a niche e-commerce brand runs a store with a devoted, small audience. She sells a specialized product to people who care a lot about it. She came to us with a feature idea she was excited about. An AI assistant that would help each shopper pick the right variant based on how they described what they needed. Her customers ask thoughtful questions, and she wanted to answer them at scale.

We liked the idea. Then we did the math with her, and the math said don’t build it.

The cost of the feature does not care how small you are

Here is the part that catches people. The cost to build and run an AI feature is close to fixed. It does not scale down with your volume. You still have to scope it, wire it to your catalog, write the prompt and the guardrails, test it against real questions, watch it for bad answers, and keep it working every time your product line or your platform changes. A store doing a few hundred orders a month pays roughly the same build and upkeep as a store doing a few hundred thousand, because the work is in the feature, not in the traffic.

So the payback is a division problem. Fixed cost on top, number of uses on the bottom. When the bottom number is small, the cost per use is enormous, and it stays enormous no matter how good the feature is. A brilliant AI assistant that helps forty shoppers a month is still an assistant whose build and maintenance you are amortizing over forty shoppers a month. The quality of the feature cannot fix the size of the denominator.

The founder’s instinct was that being small was a reason to want the help more. Each customer matters, so serve each one better. That instinct is kind and it is exactly backwards on cost. Being small is the reason the feature struggles to pay off, because there are not enough uses to spread the fixed cost across.

“It never pays back” is a real answer

Teams treat “we should not build this” as a failure of imagination. It is usually a correct reading of the numbers. There is a whole class of features that are technically buildable, genuinely useful per use, and still net negative because the volume will never get there. Naming that out loud is one of the more valuable things scoping does.

MIT’s 2025 research found that around 95% of enterprise generative AI pilots delivered no measurable return. Not because the technology failed in some dramatic way, but because a lot of them were pointed at work where the payoff was never going to cover the cost of building and running the thing. Volume is one of the quiet reasons. A pilot aimed at a low-frequency, low-volume slice of the business starts underwater and never surfaces.

What to do instead of building it

The founder did not walk away empty-handed. Low volume closes one door and opens cheaper ones.

For her, the right answer was not an AI feature but a very good static sizing-and-fit guide, written once, that captured the same expertise she wanted the assistant to give. It answers most of the questions, it costs almost nothing to maintain, and it does not need to earn back a build. Where a shopper’s question is genuinely unusual, she answers it herself, because at her volume she can, and that personal answer is worth more to a small devoted audience than a machine’s.

The general rule: when volume is too low, spend on things whose cost is also low. A template. A guide. A simple rule. A human doing the rare case by hand. Reserve the fixed-cost machine for the work that happens enough to feed it.

How we approach it at Density Labs

Part of what the AI Opportunity Assessment does, our fixed two-week engagement at $2,500, is run the payback math before you commit to a build. We estimate what the feature costs to make and keep, we estimate how often it will actually get used, and we tell you plainly when the second number cannot cover the first. Sometimes the most useful sentence we say in two weeks is “at your volume, this one will not pay back, and here is the cheaper thing that will.”

Some AI features are worth building. Some are only worth building if the volume shows up. Do the division before you write the code.