How to tell which AI features are worth building
Not every problem is an AI problem. A simple test for deciding where the technology genuinely changes the economics, and where it's just noise.
The pressure to 'add AI' has produced a lot of features nobody asked for. The useful ones share a pattern, and once you see it, the decision gets a lot easier.
Ask two questions. First: is the task one where being roughly right, fast, is more valuable than being exactly right, slowly? Summarization, drafting, routing, and search all fit. Financial reconciliation does not.
Second: can you measure whether the output is good? If you can't evaluate quality, you can't improve it, and you certainly can't trust it in production. An AI feature without an evaluation harness is a liability wearing a demo's clothing.
When both answers are yes, AI can genuinely change the economics of a problem. When either is no, you're usually better served by ordinary software, and your users will thank you for the restraint.