Choose
Choose a problem worth the cost
AI makes prototypes cheap. It does not make customer problems, distribution, or operational change cheap. Start with work that is painful enough to deserve a better way.
Primary question
Which costly, recurring problem is specific enough to pursue before you decide how AI should solve it?
Why it matters
The decision behind the output
A convincing prototype can make a weak problem look temporarily important. The interface moves, the model responds, and the team begins discussing features before anyone has shown that the underlying work creates enough cost to change behavior. Choosing well means resisting that momentum until the workflow, the person carrying it, and the consequence of leaving it unchanged are visible.
The strongest starting point is rarely a broad aspiration such as helping teams work faster. It is a repeated situation in which delay, expert labor, errors, risk, or coordination already consume something valuable. That existing cost gives discovery a concrete object: recent cases, current workarounds, accountable people, and a reason a better outcome might earn adoption.
Watch for
Common mistakes
- Starting from a model capability instead of a costly workflow.
- Treating broad interest in AI as evidence that a specific problem matters.
- Choosing a problem without identifying who pays the cost of leaving it unsolved.
Use this sequence
A working framework
- Name the workflow where time, money, risk, or attention is repeatedly lost.
- Find the person who feels that cost often enough to change their behavior.
- Describe the outcome they need without assuming an AI feature is the answer.
- Choose the narrowest problem whose cost makes a better approach worth trying.
From the field
A field example
Consider a compliance team that manually compares policy revisions across jurisdictions. The promising opportunity is not document summarization in general. It is the repeated handoff where an analyst must locate changed obligations, trace them to source language, and explain the operational consequence before a deadline. That narrower problem exposes who feels the cost, what a trustworthy outcome requires, and where AI uncertainty may or may not be acceptable.
Use this threshold
The decision rule
Keep investigating a problem when you can name a recurring workflow, identify the person accountable for its outcome, observe a costly workaround, and describe the desired result without relying on AI as the reason it matters. Narrow or leave the idea when the pain disappears outside the demo, the affected person cannot be reached, or the consequence is too small to support a meaningful change in behavior.
Field Notes
Related essays
- How to Find AI Startup Ideas From Expensive Workflows
Which repeated workflows already cost enough in delay, expertise, errors, or coordination to support a startup?
- How to Tell Whether a Problem Actually Needs AI
Does this workflow need probabilistic AI, or would rules and conventional software solve it more reliably?
- How to Choose Between Multiple AI Startup Ideas
Which AI startup idea has the strongest combination of painful demand, access, testability, and a first transaction?