Not every problem benefits from AI. Sometimes an explicit rule, a guided flow or a human operator is fairer, clearer and cheaper. The designer should actively participate in this assessment, not wait until technical choices are already settled.
Use AI when it adds measurable value and risks are manageable: non-manipulative personalisation, assistance that reduces cognitive load, synthesis of large feedback volumes with human review, accessibility (e.g. automatic descriptions with quality control). In these cases, design transparency and user control from the start.
Avoid or reject AI when the decision affects fundamental rights without possibility of human appeal: healthcare, justice, welfare, credit, immigration. Avoid it when you lack representative data or cannot explain the output to users. Avoid it when the only reason is "because competitors do it".
The concept of "human-in-the-loop" must be designed, not bolted on afterwards. It means: visible escalation points, ability to correct AI output, defined human response times, fallback when the model fails. A chatbot that never admits mistakes and offers no human contact is a design problem, not just a technology one.
Always ask: what happens in the worst case? If the model discriminates, hallucinates or is used unexpectedly, does the interface make it visible? Can the user report it? Does the team have a path to disable the feature quickly?
Saying no to AI can be the most mature design proposal. Present it as an alternative: "We can solve the same problem with a transparent step-by-step flow, test it in two weeks, and it does not require AI Act compliance". Often the team accepts — and the user wins.