AI features worth shipping

A client asked us last month to add "an AI assistant" to their operations app. When we asked what it should do, they said their competitor had one, and their board had asked. That's not a spec — that's a checkbox. We use a three-question filter before we agree to build any AI feature. First: is there a task the human is doing right now that they resent? Not one they enjoy, not one that showcases their judgment, but one they actively want to hand off. Reformatting a spreadsheet, drafting a routine follow-up email, extracting the same three fields from every incoming PDF. If nobody on the team is complaining about the task, an AI won't be welcome — it will be a solution looking for a problem. Second: can we describe what "right" looks like in one sentence? If the answer to "what should this feature output?" takes more than a sentence, the feature isn't ready to build. It's ready to be discussed. Vague specs become vague features, and vague features become the ones users stop trusting. Third: what happens when the AI is wrong? Every model is wrong sometimes. If the answer is "the user reviews the output before it goes anywhere," you have a legitimate assistant. If the answer is "it just goes out," you have a liability. The best AI features we ship are the ones where the model does 90% of the work and a human owns the final send. The features that pass all three: contact-form pre-classification, meeting-note summarization with the human editing before send, image-tag suggestions, first-draft product descriptions. The features that fail: "AI-generated dashboards" that nobody trusts, "AI decisions" that route real money, chatbots that answer questions the docs already answer better. The interesting thing is that the passing features are almost never the ones a board asks about. Nobody puts "cut the ops manager's Monday reconciliation by 40 minutes" in a press release. But that's the one the team will thank you for six months later.