AI features in SaaS: where to start (and what to avoid)

AI features have a 90% demo-to-ship gap. A weekend prototype looks like magic. The same feature in production has to be reliable, fast, cheap, and trustworthy — and that's where most attempts fall apart.

Pick use cases that survive being wrong

The first question we ask: what happens when the AI is wrong? If the answer is "a user makes a bad decision", that's an unsafe use case. If the answer is "the user reviews the draft and edits it", that's a great use case.

Internal first, customer-facing second

Internal users tolerate beta-quality features and give you feedback. Customers do not. Start where the cost of a mistake is low and the feedback loop is fast.

Evaluate against real KPIs

"Cool" is not a KPI. Time saved per task, conversion lift, ticket reduction — those are KPIs. If you can't define what the feature is supposed to move, you don't know what to optimize.

Guard against cost blowups

An AI feature that's loved and costs €0.10 per use can quietly turn into a €50k/month line item. Per-tenant cost caps, caching, smaller models for cheap calls — these belong in the design from day one.

AI is a great accelerant in the right product. It's a great way to burn a year in the wrong one. The difference is almost always whether the team picked a use case that survives reality.