AI customer service for online stores: the honest buyer’s guide

Every support tool now says “AI”. Underneath that one word sit four very different machines, and buying the wrong one is how merchants end up concluding “we tried AI, it didn’t work.” We build one of these tools — so read this as an opinionated guide, but the framework holds whatever you end up choosing. Before any feature list, ask three architecture questions.

Question 1: Does it look things up, or just phrase things?

A language model on its own knows nothing about order #4521. If the “AI” answers a shipping question without pulling the live order and tracking data, it can only produce a polite shell — or worse, a confident invention. Hallucination in support is almost never exotic: it’s an ungrounded model guessing at a delivery date.

The test is simple: in the demo, ask about a specific order and check whether the answer contains the real tracking status. If the vendor demos with generic questions only, you’ve learned what you need to know.

Question 2: Does it act, or does it stop at text?

Sending a reply is the last step of a support case — the work is the lookup, the decision, and the backend actions: the refund in the shop, the return label, the replacement order. A tool that drafts beautiful replies but leaves the actions to a human has automated the cheap part; your cost per ticket barely moves, because typing was never the expensive part (the full cost math).

This is the line between the two big categories: copilots (draft, human executes) and resolution systems (execute, human supervises). Copilots feel safer but plateau fast. Resolution systems are where the case economics actually change — if question 3 is answered well.

Question 3: Who is in control, and how do you find out you can trust it?

Full autonomy on day one is a red flag, not a feature. The mechanisms that make automation trustworthy are boring and non-negotiable:

• Draft/approval mode — every reply and every action proposed first, with the reasoning and a confidence score visible, until you switch a category to automatic.
• Per-category autonomy — WISMO going automatic while warranty judgment calls stay human is a policy, not a compromise (how we sequenced it on our own brands).
• Escalation by uncertainty — anything the system isn’t sure about lands with a human, context attached, instead of being guessed at.
• An audit trail — every action, traceable, always.

The GDPR trap (and the AI Act footnote)

The improvised version of AI support — pasting customer emails into a consumer ChatGPT window — is, in almost all cases, a GDPR violation: personal customer data flows to US servers under a consumer contract with no data-processing agreement, and potentially into training data. “It’s just one email” doesn’t change the legal analysis.

The compliant checklist for any AI tool that touches customer data:

• A data processing agreement (DPA/AVV) with the provider
• Third-country transfers covered by EU Standard Contractual Clauses, sub-processors named
• EU-first processing where offered — data residency matters to your own privacy policy
• Transparency: the EU AI Act’s transparency obligations mean customers shouldn’t be left guessing whether they’re talking to a machine — and your privacy policy needs to name the processing

One more note that surprises people: with approval mode and your own tone guidelines, customers generally can’t tell drafts apart from human-written mail — the transparency duty is about chatbot-style direct interaction and honest documentation, not about banning AI from the inbox.

The 10-question vendor checklist

1. Does it read live order and tracking data per case?
2. Can it execute refunds, labels and replacement orders — in my shop?
3. Is there a draft/approval mode, per category?
4. What exactly happens when it’s uncertain?
5. Is there a full audit trail?
6. Where is data processed, is there an AVV/DPA, who are the sub-processors?
7. Which channels (email, WhatsApp, social) and shops (Shopify, WooCommerce) are native?
8. What does pricing scale with — seats, tickets, resolutions, or flat? Per-seat pricing punishes team growth; per-ticket pricing punishes your own growth.
9. What automation rate do comparable stores reach — and how is it counted? (End-to-end without human touch is the honest metric; “AI assisted” is marketing.)
10. How long to go live, and what happens when I leave — data export, no migration lock-in?

Realistic expectations

Whatever a landing page promises: automation is earned category by category. A realistic curve is draft mode in week one, first categories automatic within a few weeks, and a majority of cases end-to-end after the trust has been built — we run at 68% on our own store, counted strictly, with the remaining third being exactly the judgment calls that should reach a human. Anyone promising 95% in week one is measuring something else.

The cheapest way to ground all of this: run a real case from your own store through a live system and watch what it actually does — that’s what the demo is for, no signup required.