Which AI features are worth shipping?
Three patterns earn their keep every time: turning messy text into structured data (parsing, extraction, categorisation), summarising long content into a glance (digests, briefs, exec summaries), and natural-language search over the user's own data. Everything else is decoration.
Why those three?
They all replace a job a human currently does badly or slowly. Send Feedback uses pattern one — a thousand support messages clustered into a ranked roadmap. Real time saved, immediately measurable.
Which AI features quietly burn cash?
Chatbots glued onto a marketing site (nobody talks to them). 'AI-powered' recommendations on tiny datasets (a sorted list wins). Open-ended generation features with no constraints (users can't tell good output from bad).
How much should AI cost to run?
For most SaaS, AI inference should be under 10% of revenue. If it's higher, you're either pricing wrong or using the wrong model. We default to small/cheap models and only upgrade when the use case demands it.
The rule we use
If a sorted list, a SQL query, or a regex would do the job, ship that instead. AI is the right tool for genuinely fuzzy problems — not for problems we already know how to solve.