The smallest useful version fits on an index card

A product manager at an online marketplace told me about the test she started using after one too many AI projects ballooned past its budget. It was not a framework or a template. It was an index card.

Her rule was this. Before a project got approved, someone had to write the smallest version that would still deliver real value, and it had to fit on a single index card. If it did not fit, the scope was too big, and they were not done thinking.

She was not being cute. She had watched enough AI proposals sprawl to know that a scope you cannot state briefly is a scope you do not understand yet. The card forced the hard part, which is deciding what to leave out.

The card forces you to name the value

Most AI project descriptions are long because they are hedging. They list every feature someone might want, every input the model could someday handle, every edge case a stakeholder raised in a meeting. The length feels like thoroughness. It is usually the opposite. It is what happens when nobody has decided what the first useful version actually is.

The index card removes the hiding places. On a card you have room for the decision the feature changes, the single input it takes, the output it produces, and what counts as good enough. That is it. If you cannot fit those, it is because one of them is not clear yet, and the missing clarity is exactly what would have sunk the project three months in.

She gave me an example. Her team wanted an AI tool to help sellers write better listings. The first proposal was enormous. Rewrite titles, suggest pricing, recommend photos, flag policy issues, translate for international buyers. Every one of those was reasonable. Together they were a year of work with no clear first win.

The card version was one line. For sellers whose listing has an empty description field, suggest a description drafted from the title and category, which the seller edits before publishing. That was the smallest thing that delivered value. It shipped in weeks, sellers used it, and it taught the team what the next card should say.

Small is not the same as easy

The point of the card is not to make the project trivial. It is to make it honest. A tightly scoped feature can still be hard to build. What the card guarantees is that the hard work is aimed at one clear win instead of spread across ten fuzzy ones.

A VP of engineering at a fintech described the same discipline in different words. He made every AI proposal state, in one sentence, the first thing it would do that a person would notice and value. If the team could not write that sentence, he sent them back. The idea might be fine. His concern was that a project that cannot name its smallest win tends to spend its budget searching for one.

There is a second benefit that shows up later. A small, clearly scoped first version is testable. You can put it in front of real users and learn whether the whole premise holds before you have spent the budget. The sprawling version is unfalsifiable. It has so many parts that when it underdelivers, and MIT’s 2025 research found that around 95 percent of enterprise generative AI pilots deliver no measurable return, you cannot even tell which part failed. The card gives you something you can actually judge.

How we approach it at Density Labs

The output of an AI Opportunity Assessment, our two-week fixed engagement at $2,500, is close to that index card. We take a broad ambition and cut it down to the smallest version that would deliver value and could ship, written plainly enough that everyone in the room agrees on what it is. The cut is the deliverable. Deciding what not to build in the first version is most of the work, and it is the part that saves clients from funding a scope nobody could hold in their head.

If you cannot write the smallest useful version of your AI feature on an index card, do not build yet. Keep cutting until it fits.