Journal / Kirill Naumov
A practical guide to AI translation for retail teams: choose the right use cases, protect product and policy meaning, and keep human support within reach.
By Kirill Naumov ·
Retail has a very human problem: people do not always share a language, but the question still needs an answer. ‘Where can I find this?’ is one thing. Explaining a return condition, a product warning or a delivery problem is another. A translation tool can help a conversation get started; it should not quietly turn a complicated policy into a confident guess.
AI translation is becoming a practical frontline use case. In June 2025, Walmart described a real-time translation feature in its associate app, supporting text and speech conversations across 44 languages. The company says it incorporates Walmart-specific knowledge so that house-brand names such as Great Value are handled in context. Retail language is full of product names, store terms and local shorthand that a general translator may not recognise. This is Walmart’s reported example, not a claim about a deployment by Kirill Naumov or his clients.
‘Translate everything’ sounds ambitious, but a useful pilot starts smaller. Pick one repeated interaction: helping a customer locate a product, clarifying a click-and-collect step, or helping two colleagues hand over a task. Write down what a good outcome looks like before choosing a tool.
For simple wayfinding, a short translated phrase may be enough. For questions about ingredients, safety, returns or payment, the system should use approved, current information and show where that information came from. If the source is missing or the answer is uncertain, the right next step is a colleague—not a more fluent-sounding sentence.
Translation quality is more than grammar. A shop has brand names, product variants, abbreviations, promotions, measurements and policies. A phrase can be linguistically smooth and operationally wrong. ‘Exchange’ and ‘refund’ may not be interchangeable. A product name may need to stay unchanged. A promotion may apply only to selected sizes or dates.
Before a pilot, build a short approved glossary for the products, service terms and policy language in scope. Test it with real examples, including misspellings, noisy speech, accents and incomplete questions. Ask bilingual colleagues or qualified language reviewers to check the output. If the tool cannot preserve a key term, correction should be easy and uncertainty should be visible.
Do not treat a translation model as the source of truth. Keep approved product and policy information in the systems responsible for maintaining it. The translation layer can help express that information in another language; it should not invent a new return rule because the customer is frustrated or a phrase sounds plausible.
Customers and colleagues need to know when a human can step in. Make the handoff visible, especially for complaints, safety questions, financial issues, accessibility needs and situations where the translated message does not make sense. A colleague should be able to see the original and translated text, correct a mistake and explain the next step in language the customer understands.
Be clear about what the tool does with conversation data. Use only the information needed for the interaction, follow company policy and applicable requirements, and tell users when a conversation is being translated if that is appropriate for the setting. Avoid sending sensitive details into an unapproved consumer app simply because it is the fastest option on a phone.
Language count makes a tidy headline; it does not tell you whether people understood each other. In a limited pilot, review a sample of conversations with permission and appropriate safeguards. Track whether the customer reached the right service outcome, how often staff had to correct the translation, whether the same misunderstanding returned, and whether colleagues felt confident using the tool. Balance any time saved against errors, escalations and customer feedback.
Start with one store, one service situation and a clear fallback. Compare the translated workflow with the current one, then decide whether to adjust, expand or stop. The point is not to make every conversation faster at any cost. It is to make a useful conversation possible when language might otherwise get in the way.
AI translation can be a strong addition to retail service when it supports the people in the conversation, preserves the retailer’s actual meaning and makes uncertainty visible. For practical applications across retail, sales, staff learning and loyalty, explore the AI for Retail workshop or discuss a retail AI consulting brief.
Source: Walmart’s 2025 announcement on real-time translation.