Building small AI products: start from the job, not the model

Every week brings a new model and a new leaderboard. It is tempting to start an AI product by asking “what can this model do?”. I think the better question is: what work are people doing by hand, over and over, and hating every time?

The model is the easiest part to replace

Today’s best model will be beaten by a better, cheaper one within months. If your product’s value is “it uses model X”, that value disappears the moment everyone can use model X. What is harder to copy is a precise understanding of one job, your users’ context and data, and what happens when the AI gets it wrong.

Pick a job that is narrow, painful and measurable

• Narrow: you can describe it in one sentence. “Rewrite product descriptions for marketplace listings” beats “writing assistant”.
• Painful: people spend time or money on it every week.
• Measurable: you can look at a result and tell whether it is good.

Collect examples before writing prompts

Before you tune a prompt, gather 20–30 real examples: real inputs, and the output a skilled person would produce. That is your test. Every time you change the prompt, swap the model or add a step, run the whole set and compare. Without it, every “improvement” is just a feeling.

Design for when the AI is wrong

The AI will be wrong; the question is what happens then. Let people review before sending, fix things quickly, or see where the AI is unsure. A product that is right 90% of the time and lets you fix the rest in seconds usually beats one that is right 95% of the time but fails silently.

Count the cost from day one

Every model call costs money and waiting time. Estimate early: how many calls per user per day, how long each one is, how long people wait. Often a simple step, like filtering with plain code, reusing results you already have, or a small model for easy cases, removes most of the cost without users noticing any difference.

Launch small, iterate fast

The first version only needs to do one job well for one group of people. Put it in front of a few real users, watch how they use it, fix, repeat. AI makes building faster than ever; the advantage now is choosing the right thing to build.