Draft — awaiting owner approval · noindex until released
When an AI feature earns its keep
Chimbuk studio4 min read
An AI feature earns its keep when it removes minutes that real users were already spending — not when it adds a sparkle to a screenshot. That sounds obvious; it is remarkably rare in pitches. Our discovery filter is three questions, asked in this order.
1 · Where is the bottleneck, in minutes?
Support tickets answered from the same ten documents. Reports typed from the same three dashboards. Drafts that start from the same template. If you can't name the minute-sink, the feature is decoration. We ask for a week of workflow truth before a model is chosen.
2 · What does wrong look like — and who cleans it up?
LLMs fail confidently, so product design starts from the failure, not the demo. Every AI feature we build ships with guardrails, a visible fallback (the deterministic path), and a feedback loop that makes the second version measurably better. The evaluation set is part of the deliverable, not a slide about one.
3 · What is the monthly rent?
Tokens, embeddings, storage, the human reviewing the edge cases — AI features have a per-user cost curve. We model it at launch traffic and at ten times that, because a feature that only works at pilot scale is a prototype with extra steps.
Pass all three, we build it with an eval set and a cost budget. Fail one, it goes on the skip-list — and the skip-list has saved more projects than any model choice ever will.
Begin the ascent
We build the things we write about.
If the field notes read useful, the work should too — tell us about your project.