“AI upselling” looks absurd when suggestions ignore allergens or recommend a dish that is 86'd. Good recommendations need real baskets, clear guardrails, and a connection to the menu the kitchen can fulfil. Useful signals combine history, menu structure, and context. Simple explainable models often win because teams can see why a suggestion disappears. New dishes need a starting plan, and mistakes need a kill switch for each dish.
A suggestion either feels “intelligent” or ridiculous. If a system recommends a dish that is 86'd, or one containing flour after a guest filters for gluten-free, it loses trust immediately. Good suggestions are ranked recommendations trained on how guests really order from you — and limited by what the kitchen can fulfil today. Done well, they lift the basket without turning the final step into chaos. Done badly, every strange suggestion chips away at trust.
The prerequisite is a clean basket state: customisations, allergens, availability, and prices all need to speak the same truth. A system cannot pair items sensibly when the website and POS disagree.
Signals that survive the rush
Useful suggestions combine three kinds of knowledge: history (what sells together and what regulars choose), menu structure (which add-ons are permitted and fit together), and context (time of day, channel, pickup, or delivery). You often do not need a complex model. A simple, explainable approach frequently wins in practice because people can see why a suggestion disappears when a dish is 86'd. When a team has to trust a system, transparency beats magic.
Have a starting plan for new dishes
A new dish has no history, and that is exactly where many “AI upsells” produce nonsense on day one. The answer is a starting plan: until orders arrive, base the suggestion on the dish’s attributes — cuisine style, heat level, drink category — until enough real data builds up. That gives a new dish a sensible place from the start rather than recommending it at random or not at all.
Rank for margin, not clicks
Optimise for contribution margin and acceptance, not simply the number of add-ons. A frequently accepted suggestion that causes forgotten items or remakes is ultimately a net loss. Include kitchen effort, the effect on fulfilment time, and refund risk. The system should favour suggestions the pass can still absorb in the same busy hour. Set upper and lower limits on bundle discounts too, so a promotion does not eat the basket gain it is meant to create.
A kill switch everyone trusts
Two things are non-negotiable. First, a hard allergen and dietary filter: it overrides every margin calculation, without exception. Second, a kill switch — the ability to turn off a suggestion for a dish or group immediately when it behaves badly. No one wants to wait out a weekend of poor suggestions while the system retrains. Before a broad launch, run a quiet test: review suggestions internally and compare them against the existing basket before guests see them.
The 7 most common mistakes
- Suggesting a dish that is 86'd.
- Failing to put allergen filters above everything else.
- Letting website and POS disagree about the basket.
- Having no starting plan for new dishes.
- Ranking by clicks instead of contribution margin and effort.
- Having no kill switch per dish for bad behaviour.
- Launching to everyone without a quiet test.
How to build a system everyone trusts
Frequently asked questions
Why do many ‘AI upsells’ feel so unsuitable?+
Do I need a complex AI model?+
What do I do with new dishes that have no history?+
How do I protect against bad suggestions?+
Growth without losing trust
A recommendation system is only as good as its guardrails. Suggestions from real orders, tied firmly to the menu the kitchen can fulfil, ranked by margin, and protected with a kill switch lift the basket without fighting the pass or confusing guests. That closes the upsell chain: from the right dish, to the right moment, to a sharper menu, to a system guests and kitchen can both trust.
