Most "AI integration" is a wrapper around one API call.

Most of what I use AI for never reaches a user. It's the research pass before a design starts, a rough first draft of copy I then rewrite properly, a script that turns forty rows of client data into forty personalised assets instead of me doing it by hand at midnight. Repetitive task, small judgment call — good candidate. Big judgment call — I do it myself.

Sometimes the product has to reason about something: summarise it, classify it, answer a question in plain language. I design those the way I design anything else. What happens when the model is slow, or wrong, or just making something up with total confidence — that gets worked out before launch, not discovered by a user.

Wiring an LLM into a product is the easy part. What takes the time is a prompt structure that holds up outside the three examples you tested, a cost and latency budget, a fallback for when a call fails, and usually OpenRouter over a single vendor so the product isn't hostage to one company's pricing.

Honestly, this is where AI has changed my own work the most, not the client-facing stuff. Faster first passes in Figma, code scaffolding I go through by hand afterward, Apify pulling reference material I used to collect one tab at a time. It hasn't replaced any of the decisions. It's just fewer hours between having an idea and having something on screen to react to.