Most AI products die for the same reason: they were built to show off a model, not to finish a job.

The winners look boring from the outside. They're not the most impressive demos. They're the tools people quietly refuse to give up.

Here are the five criteria they all share.

1. It attacks a job people already pay for

Nobody wakes up wanting "AI." They want the invoice sorted, the video cut, the lead answered, the contract checked.

If you can't name the line item your product replaces — an agency retainer, a freelancer, three hours of someone's Tuesday — you don't have a product. You have a feature waiting to be absorbed by whoever owns the workflow.

The test: can you finish this sentence without using the word AI? "Before this, they paid ___ to do ___."

2. Being wrong is cheap

This is the criterion almost everyone underestimates.

Models are probabilistic. That's not a bug you'll patch — it's the material you're building with. So the question isn't "how accurate is it?" but "what happens on the bad outputs?"

Draft an ad script badly → the user deletes it and regenerates. Cost: 4 seconds. File a tax return badly → the user gets a letter from the government.

The best AI products live where errors are cheap, obvious, and instantly reversible. Where they aren't, the product has to make verification effortless: show sources, show diffs, show its work, keep a human on the last click.

Trust isn't a marketing problem. It's an architecture decision.

3. Value arrives before commitment

Traditional software earns its onboarding. You configure it, you import your data, you learn it, and eventually it pays off.

AI products don't get that patience. The expectation is magic, and magic doesn't ask you to connect a data source first.

Paste something in. Get something out. Under a minute, ideally under ten seconds. Everything else — accounts, integrations, settings, pricing — comes after the person has felt it work once.

If your product needs a demo call to be understood, the demo call is now part of your product. Price accordingly, or redesign it.

4. The model is the commodity — the wrapper is the product

"It's just a wrapper around GPT" stopped being an insult about two years ago.

Everyone has access to the same weights. What no one else has:

  • the workflow you designed around the output
  • the context and data you feed in that the raw model can't see
  • the taste encoded in your prompts, your defaults, your rejections
  • the place in the user's day where you sit

A generic model plus deep opinions about one specific job beats a frontier model with no opinions every time. Your moat isn't the intelligence. It's the judgment wrapped around it.

Corollary: never build something a foundation lab would ship as a checkbox next quarter.

5. It compounds with use

The first version is a tool. The version that survives is a system that gets better the longer someone stays.

Compounding can come from a lot of places — accumulated context about the user's brand and preferences, corrections that feed back into defaults, a library of past outputs that becomes an asset, a team's shared history inside the product.

The mechanism matters less than the effect: after three months, switching costs something. Not because you locked anyone in, but because leaving means starting over.

Products without compounding get churned the moment a slightly cheaper clone appears. And a clone always appears.

The uncomfortable summary

Four of these five criteria have nothing to do with AI.

Real problem. Manageable failure. Fast value. Sharp opinions. Compounding returns. That's just product work — the kind that existed long before transformers and will outlive whatever comes next.

The model is the easy part now. It's available to everyone, priced like electricity, improving on someone else's roadmap.

Everything that's actually hard is still yours to build.

Which of the five do you think kills the most AI startups? My bet is #2 — and it's usually discovered after launch.