Albert Heijn’s AI assistant: the answer needs an owner

A shopper asks whether the oat drink on the shelf contains an ingredient they avoid. The colleague has a phone, a queue and perhaps thirty seconds to help. An AI answer can be wonderfully quick. The catch is that a beautifully phrased answer from last month's product data is worse than an honest pause. In a store, speed only earns its place when the information survives a check against the pack in front of you.

This is no longer a purely hypothetical setting. In its 2025 annual report, Ahold Delhaize says Albert Heijn launched an AI Store Associate Assistant that gives colleagues access to key information when needed. In a July 2026 company interview, it says 80,000 colleagues use an AI assistant daily for operational and product information. Those are company-reported deployment and usage claims. They do not, by themselves, establish an independently measured improvement in customer outcomes.

Which version of the truth is on screen?

A product answer should carry its source and date. The official product specification may have changed; a recipe card might be older; the label on the current pack may say something more specific. These are different levels of evidence. The assistant should not flatten them into one confident sentence. Where ingredients, allergens or safety matter, it should point the colleague to the authoritative current source and make uncertainty visible.

Location matters just as much as freshness. A general catalogue can tell you that an item exists while a particular shop has sold out. A return policy can vary by item type or channel. A promotion may have ended at midnight. Before presenting an operational answer as actionable, the assistant needs the store, date and relevant policy version. If that context is missing, asking one extra question is quicker than correcting a confident mistake later.

Teach the handoff, not only the prompt

An employee learning exercise could start with an ordinary request: 'Where is this product?' A second round adds a substitution, a customer with a dietary constraint or a price that disagrees with the shelf label. The learner has to decide when to trust the tool, when to inspect the pack or system of record, and when to involve a supervisor. This is my proposed training design, not a description of Albert Heijn's internal curriculum.

The most valuable moment might be the sentence the colleague says when the tool cannot settle the issue: 'I can check that for you.' An assistant should be allowed to abstain. It can pass along the question, the source it checked and the precise point of uncertainty. A human handoff without that context makes the colleague start again; a handoff with an invented answer makes the problem worse.

Measure what happened after the answer

I would test a small set of high-volume questions first: location, price conditions, substitutions and product attributes. For each, compare time to a verified answer, correction rates, escalations, repeat customer questions and colleague confidence. Sample real answers against the current source of truth. Count an abstention that prevented an unsafe guess as a success, not a failure to generate text.

A usage count such as 80,000 daily colleagues is useful evidence of scale, but it is not a customer-service score. The harder question is whether shoppers leave with the right product and whether staff can explain why the answer is right. The best store assistant may sound less magical than a demo. It names the source, respects the limits of the data and leaves the colleague room to say, 'Let me check.'