There are two kinds of AI features shipping right now. The first kind removes work your users already resent: triaging the queue, drafting the reply, reconciling the ledger, naming the meeting. The second kind is a chat box bolted to the sidebar because the board asked what the AI strategy was.
Users can tell them apart in seconds. So can churn data.
Start from the resentment
The best discovery question we know: "What part of your job do you do while sighing?" The answers are almost never "I wish I could chat with my documents." They are specific, repetitive, judgement-shaped tasks: exactly the shape that current models are good at, provided a human stays on the approve button.
Automate the sighing. Keep the judgement.
Make it boring, make it auditable
The AI features that survive procurement have unfashionable properties: they log their reasoning, they degrade gracefully to the manual path, they never block the workflow when the model is down, and their accuracy is measured against a golden set that someone owns. This is not the fun part. It is the part that separates a demo from a deployment.
Our rule of thumb: if removing the word "AI" from the feature announcement makes it sound useless, the feature is useless. "AI-powered insights" survives the test poorly. "Month-end close in a day instead of eleven" survives it fine, and happens to have a model inside.
Written by the codemyx studio, the people who build the things these opinions come from.