The Mom Test for AI: Why Your Idea Will Fail
Most AI agent products die young because developers are solving imaginary problems that real users don't actually want to pay for.
Last week, an acquaintance pitched me an idea for an AI agent that automatically summarizes every Slack message thread. It sounds high-tech, but I knew the project was likely doomed to fail within just a few weeks.
The real problem with AI agent projects
As developers, we are very easily swept up in technology. When OpenAI releases GPT-5.2 or Anthropic updates Claude Sonnet 4.6, our first reflex is usually to find a way to cram them into a product. You come up with an idea, ask your friends if it’s good, and they nod in approval.
But the truth is, everyone is lying to you. The concept of “The Mom Test” from the book of the same name by Rob Fitzpatrick (momtestbook.com) points out that if you ask the wrong way, even your own mother will lie to make you happy. In the world of AI, this lie is amplified by curiosity about new technology. People praise your idea, but when the product launches, no one pulls out a credit card to pay.
Why your questions are killing your product
The mistake of asking about the future
If you’re planning to build an AI agent for appointment management and you ask a user: “Would you like an AI that automatically schedules appointments?” almost 100% will say yes. Who wouldn’t want to be more free?
However, according to Y Combinator’s startup library (ycombinator.com/library), asking about future behavior always yields misleading results. The right question should be: “How much time did you spend scheduling appointments last week? What tools did you use to solve it?”. If they answer that they don’t remember, or it only took them two minutes to type in Google Calendar, then your problem isn’t big enough to justify embedding a heavy AI model.
API costs and real value
Another blind spot is operating costs. According to the official pricing on docs.anthropic.com, Claude Sonnet 4.6 costs about $3 per million input tokens. If your agent has to constantly read long data streams, costs will skyrocket.
If the value the user receives is only saving 5 minutes a day, they will never pay a $20 monthly fee to cover your server costs. This is why many people fall into the Using too much AI leads to burnout: The productivity paradox loop. They spend too much effort maintaining bots that don’t provide core value.
Mom Test vs. Standard AI Interviews
| Criteria | AI-style Interview (Wrong) | Mom Test-style Interview (Right) | Notes |
|---|---|---|---|
| Opening question | What do you think about this AI agent? | How are you currently solving problem X? | Focus on the present. |
| Features | Can my AI automatically do Y? | When was the last time you had trouble with Y? | Look for real pain points. |
| Payment | Would you pay $10 for this app? | How much are you currently paying for existing tools? | Real money is the only proof of demand. |
The biggest difference between a surviving product and a junk project is whether you are willing to face the harsh reality.
How to test AI agent ideas without coding
- Understand the current workflow. Don’t mention a single word about AI or LLMs. Ask the user to open their screen and show you exactly how they did that task yesterday.
- Run a manual experiment. Before writing code to connect to the Gemini 3.1 Pro API, play the role of the AI yourself. Receive user data via email, process it by hand, and send it back. If they don’t value the results you produce manually, they won’t care about AI-generated results either.
- Assess willingness to pay. Ask them to place a deposit or commit to trying a paid version even if the product is not yet finished.
Frequently Asked Questions
Should I use AI to build a fast prototype?
Yes, but only once you have correctly identified the problem. Using Cursor or Windsurf to quickly code a demo is good, but don’t use it as an excuse to skip talking to users.
How do I know if users are lying?
They are lying when they use words like “usually,” “probably,” or “will.” They are telling the truth when they can recount specific details of events that happened yesterday or last week, accompanied by specific frustrations.
Does GPT-5.2 change these principles?
Not at all. No matter how smart a model is, it is useless if it solves a problem no one cares about. The better reasoning capabilities of GPT-5.2 only help the agent run more stably; they don’t create market demand out of thin air.
Conclusion
AI is currently a massive bubble of bad ideas wrapped in a shiny technological shell. I see too many engineers burning months writing prompts for a task that an actual user only needs three clicks to complete. Stop hallucinating along with the new language models. Get back to the most basic questions, look at the user’s wallet, and accept the truth that most initial ideas aren’t worth coding.