7 Questions to Ask Before Buying Retail AI (A Checklist for Indie Operators)

The retail AI category is loud. Vendors are pitching different products, calling them similar things, and making promises that range from honest to embarrassing. If you are an indie operator evaluating an AI assistant for your store, these are the seven questions I would put on the table in every demo before signing anything. Get clear answers in writing for all seven and the decision becomes a lot less risky.

This is a buyer's checklist, not a vendor pitch. I run Remi and I am happy for you to ask me these questions too. The goal is for you to walk into your evaluation knowing what good answers look like.

1. Who owns the data?

The first and most important question. The kiosk is going to generate conversation logs, shopper interaction data, and recommendations history. Three things to nail down:

Conversation data ownership. Are the conversations between your shoppers and the kiosk yours, the vendor's, or jointly held? The right answer for an indie operator is "yours." If the vendor claims joint ownership or sole ownership, that is a problem.

Model training rights. Is the vendor training their AI models on your store's data? If yes, can you opt out and still use the product? "Training on your data" sometimes means improving the system you specifically use (acceptable) and sometimes means feeding a generic model the vendor sells to your competitors (not acceptable). Ask which one.

Data export. If you want to leave, can you export everything — conversations, configurations, your custom prompts, your product attributes? In what format? How long does it take?

A clean answer looks like: "You own your conversation data. We train operator-specific improvements on it but never feed your data into a model used by other operators. You can export everything in JSON within seven days of request, no charge."

A bad answer looks like: vague language about "joint stewardship," no clear export policy, or "we don't share data with competitors" without a contractual commitment to that effect.

2. What happens when the internet goes down?

Indie stores lose internet. It happens. Storms, ISP issues, the squirrel that chewed through the line behind the building. The right question is not "does it work offline" — most don't, fully — but "how does it degrade?"

The honest expectations for a kiosk-style AI assistant:

Should still work without internet: Product catalog lookup, store-map navigation, basic price information, fallback messaging that tells the customer the system is degraded.

Will not work without internet: Conversational AI (the model lives in the cloud), real-time inventory sync, voice transcription if the speech engine is cloud-hosted.

What "graceful degradation" looks like: The kiosk does not freeze, does not show error screens, and does not crash. It tells the shopper "I am running in offline mode — I can help with locations and stocked items but cannot have a conversation right now," and the operator gets a notification.

A vendor that cannot describe their offline behavior in concrete terms either has not thought about it or is hiding a bad answer. Demand specifics.

For Remi, the offline mode is described above. Catalog and store-map work; conversation degrades to a static message.

3. What languages does it actually support — well?

"Multilingual" is one of the most overstated features in AI products. Almost every model technically speaks multiple languages, but the quality varies dramatically. The questions:

What languages are productized — meaning officially supported, tested, and tuned? Be skeptical of "supports all languages." That usually means "the underlying model can produce text in any language but we have not validated quality."

For each productized language, what is the quality on domain-specific terms? A general-purpose Spanish translation is easy. A Spanish that handles "vino tinto suave" correctly in a wine context — see the bilingual liquor store post — is a different bar.

Voice as well as text? A kiosk that supports a language in text but only English in voice is a half-solution.

Switch behavior? Does the kiosk switch languages automatically when the customer speaks in a different language, or is there a button press? Auto-switch is the right answer for the customers who need it most.

For Remi, English and Spanish are the productized languages. Other languages are technically supported by the underlying model but not officially tuned. We will tell you that directly.

4. What integrations actually exist — and how deep are they?

Vendor websites list every logo they have ever touched. The real question is which integrations are production-grade and which are brochureware.

For each integration that matters to you (POS, ecommerce, loyalty, inventory tools), ask:

Is this integration live in production with real customers? How many? Reference customers should exist if the integration is real.

What data flows in which direction? Read-only? Read-write? See the Square POS integration post for what a clean integration looks like.

Who maintains the integration when the underlying API changes? POS vendors update their APIs. Someone has to keep up. If the answer is "we'll get to it" the integration will rot.

What happens during integration outages? If your POS API is down for two hours, does the kiosk degrade gracefully or break?

A vendor that has a small list of deeply-integrated POS systems is more honest than a vendor that claims integration with twenty platforms. Depth beats breadth.

5. How long is the training and ramp time?

Setting up an AI kiosk is not turnkey, no matter what the sales deck says. The honest ramp has three phases.

Phase 1 — catalog ingestion (1-2 weeks). Pulling your product catalog, cleaning attribute data, configuring categories. The vendor should do most of this; the operator should expect to spend two to four hours validating.

Phase 2 — voice and persona configuration (1 week). Recording or selecting the voice, setting greeting and closing messages, configuring promotional messaging, training the assistant on your specific recommendations and store voice.

Phase 3 — pilot and refinement (4-8 weeks). The kiosk is live. The operator and vendor review conversation logs together, identify gaps, and refine the assistant's behavior based on real shopper questions.

Total ramp from "signed contract" to "kiosk performing well in your store" is typically 6 to 10 weeks. Anyone telling you 48 hours is either selling a generic product (no real customization) or not telling you the truth.

For Remi pilots, single-store goes live in week 2 with first refinements in week 4. Full bake-in is typically week 8.

6. What does exit and portability look like?

This is the question vendors hate but operators should always ask. If you sign up and decide in twelve months that the product is not working, what happens?

Contract length. Month-to-month after an initial pilot is operator-friendly. Annual lock-ins are vendor-friendly. Negotiate.

Data export. Covered in question 1, repeated here because it is critical at exit specifically.

Hardware ownership. If the vendor provided hardware, who owns it after termination? If you bought hardware, can you keep using it with a different vendor?

Catalog and configuration portability. Your custom product attributes, categories, and prompts are work product. Can you take them?

Wind-down support. If you cancel, does the system go dark immediately or do you get a 30-60-90 day transition window?

A clean exit policy is a sign of a confident vendor. A vendor that fights you on these terms is signaling that they expect to retain customers through lock-in rather than through quality.

7. Has it been tested in real stores — like mine?

Demos are rehearsed. Real stores are not. The single most important question I would ask any vendor is: how many stores like mine have you deployed in, and can I talk to one of them?

"Like mine" matters. A kiosk that has been tested in five large urban liquor stores may not work the same way in a rural convenience store with intermittent internet. A kiosk tested in English-only markets may not handle Spanish well. Ask for references that match your store profile.

Reference calls. Real reference calls — not a curated list of paid testimonials. Operators talking to other operators. If the vendor cannot produce a single store owner who will talk to you on the phone for fifteen minutes, that is a signal.

Pilot data. Ask for a redacted version of conversation logs from a real pilot. You should be able to see what shoppers are actually asking. The pattern of questions tells you whether the kiosk is being used like a useful tool or being ignored after week one.

For Remi, we are pre-pilot phase by some measures and have early production stores by others. We will tell you exactly which stores we have, what their volume is, and put you on the phone with their owners. You should expect that level of transparency from any vendor you are evaluating.

How to use this checklist

A practical suggestion: print these questions, take them into every vendor meeting, and write down the answers verbatim. Compare answers across vendors after the calls. The differences will be more revealing than anything in the sales deck.

For each question, the right answer is specific and concrete. Vague answers — "we take privacy seriously," "our integrations are robust," "you'll be live quickly" — are non-answers. Push for the specifics.

If you want to walk through these questions for Remi specifically, book a demo and we can answer all seven on the call. We try to get to the answers without being pushed; if we miss any, hold us to it. The pricing page covers the commercial side; this checklist covers the operational side.

A final note

The retail AI category is going to get noisier before it gets quieter. Vendors will continue to make claims that range from honest to misleading. Operators who use a structured checklist like this one — these seven questions or your own version — make better decisions than operators who get caught up in the demo polish.

The decision matters. The wrong AI kiosk in your store is not just a wasted budget; it is a daily customer-experience risk for as long as it is plugged in. The right one is a real lift to basket size, customer experience, and staff efficiency. The seven questions above are how you separate the two.

Frequently asked

What if a vendor refuses to answer one of these questions?

The refusal is the answer. Move on.

Are these questions only for kiosk vendors, or also for chatbot, recommendation, and inventory AI?

Most apply across the category. Question 4 (integrations) and question 7 (real-store testing) are universal. Question 3 (languages) matters more for customer-facing tools and less for back-office AI. Question 6 (exit) applies to any vendor relationship.

Should I run a paid pilot or a free pilot?

Paid pilots get more vendor attention. Free pilots feel cheaper but often get deprioritized. A short paid pilot — 60 to 90 days at a discounted rate — usually gets you the best vendor focus. Negotiate a clear go/no-go decision point at the end.

How many vendors should I evaluate?

Three is the sweet spot. One vendor is a coin flip; five is too many to manage. Pick three that look credible from public materials, run the seven questions on each, then pick.

What if my store is too small for a vendor to take seriously?

Some vendors target only multi-unit operators. That is a legitimate business choice, just not the right vendor for you. Look for vendors who explicitly serve indie single-store operators. We do at Remi, and so do a few others.