Skip to content

AI product questions · Medium

How do you decide whether a problem should be solved with AI at all?

Short answer

Start from the user problem, not the technology. AI suits tasks with fuzzy inputs, tolerance for occasional errors and enough data or examples to learn from. Prefer simpler rules or better design when the logic is clear, errors are costly or explainability is required. Compare options on value, accuracy needed, cost, latency, risk and maintenance.

Detailed answer

A practical checklist

Ask these questions about the problem:

  • Is the input unstructured or variable? Free text, images and speech favour AI. Structured fields often do not need it.
  • Can the rule be written down? If a few clear rules solve it, use rules.
  • What does an error cost? Low-cost errors (a slightly off recommendation) are acceptable; high-cost errors (wrong medical or financial advice) need strong safeguards or a different approach.
  • Is explainability required? Regulated decisions may need clear reasons.
  • Is there data to evaluate it? Without examples of good outcomes, you cannot measure quality.
  • Do cost and latency fit? Model calls cost money and time at scale.

Worked example

A lending app wants to "use AI" in its application flow. Three possible uses:

  1. Auto-filling fields from uploaded documents: variable input, errors can be corrected by the user, saves time. Good fit, with user confirmation.
  2. Answering applicant questions about the process: good fit if grounded in official policy and with a human fallback.
  3. Making the final credit decision with a language model: poor fit. Decisions need consistency, explainability and regulatory compliance; established, auditable methods are more appropriate.

Recommend the first two, define success for each and explain why the third is out of scope.

Common pitfalls

  • Adding AI because leadership wants an "AI story".
  • Ignoring ongoing costs of evaluation and monitoring.
  • Treating a promising demo as proof of production quality.