Home / Insights / When not to add AI

When not to add AI to your product

Saying no is a product skill. These are the situations where we advise clients to wait—or never ship the feature.

The problem is already solved without a model

If search, rules, or a form workflow already gives correct answers with auditable logic, AI adds variance and cost. Fix UX and data quality first.

You cannot measure success

Without a labeled set of good and bad outputs, you cannot know if releases help. If stakeholders only want a demo for investors, postpone until you have evaluation criteria tied to business metrics.

Humans cannot review failures safely

Medical, legal, or financial advice at scale needs human oversight and rollback. If you lack that capacity, do not expose generative answers to end customers.

Data you need is missing or messy

RAG and agents fail when source documents are outdated, duplicated, or permissioned incorrectly. Cleaning the knowledge base often delivers more value than a larger model.

Latency or cost breaks the UX

Some flows need sub-second responses on mobile networks. If only a 3-second LLM call fits, users will abandon the feature—no matter how smart the answer reads.

What to do instead

Invest in observability, automation of deterministic steps, and better onboarding. When the use case is real, you will feel pull from users—not pressure from a trend deck.