Tesco: substituting a product should not substitute the whole dinner

Ordering ingredients for dinner and receiving ingredients for a different dinner is a particular kind of service. The bags arrived. The item count is nearly right. Yet the customer intended to cook something specific. Grocery fulfilment discussions can become so interested in completing the order that they forget to ask whether the intended evening survived.

Understand a constraint before suggesting an alternative

A free-text note needs to become a small number of rules the customer can confirm. Keep the quantity, preserve the selected product type, stay below a price limit. The model can interpret language, but it cannot decide that one restriction matters less than another. That distinction becomes especially important for ingredients.

Only then does it make sense to offer alternatives, with the differences explained. Smaller pack, different unit price, same stated format: these are more useful than a generic claim that the product is similar. Where verified information is insufficient, leave the item unfilled. Generative confidence is not equally welcome in every aisle.

A rejected suggestion deserves a question about why it was rejected. Was it flavour, quantity, price or the fact that no substitute was wanted? These are different signals. Saving preferences should be optional and category-specific. Accepting another variety of apple does not give the algorithm permanent authority over the contents of somebody's fridge.

Quality matters after the bags arrive

One straightforward category allows a useful comparison with ordinary substitution handling. Pickers must retain a quick way to reject a suggestion. Time spent navigating an assistant should not consume the time saved picking. A narrow pilot helps distinguish a catalogue problem from an interface or policy problem.

Evidence will appear in substitution acceptance, post-delivery contacts about unsuitable replacements and picker time spent resolving an unavailable item. Increasing acceptance alone is insufficient. Someone may approve a suggestion without understanding what changed. Checking the experience after delivery is therefore part of assessing the decision, not an optional survey for the presentation.

Tesco lets customers leave notes for personal shoppers, including substitution preferences. That is the documented starting point. This is my independent AI proposal for handling these preferences, not a description of Tesco's current technology or work I delivered for the company.

Dietary requirements should never be inferred from shopping history or softened through approximate language. Use explicit constraints, verified ingredient information and a safe refusal when uncertain. The AI I want does not fill a basket at any cost. It understands when leaving a space is more helpful than shipping an entirely different idea of dinner.