How should an AI-assisted service handle what it thinks a customer wants?
Suppose someone asks for a product within a budget, available locally, with a specific ingredient preference. Those are stated constraints. Deciding they must also like a certain color or brand is an inference, and it should be recorded as one.
The anticipatory-experiences panel described detailed shopping requests that combine several requirements. The operational lesson is to keep the person's own words distinct from the interpretation a system adds.
Put that distinction in the brief the recommendation runs on. If a requirement is unclear, ask a focused question or show the uncertainty, and leave the requirement as stated.
Make the comparison explainable: which options meet the stated requirements and which involve a tradeoff. A nearly suitable item shouldn't look like it meets every condition just because the system wants to finish the task.
Corrections stay narrow
Give customers an easy way to correct an assumption, and carry that correction into the current result. It doesn't give you permission to infer unrelated preferences.
Personalization should help someone express and act on what matters to them. A hidden profile shouldn't outrank the request they just made.
What to try this week
Take a recent customer request and mark each requirement as stated or inferred. Anything inferred gets its own label in the brief.