I’m a Shopify developer with 13+ years across many stores, and I’ve been working on a longer piece about AI customer support, specifically about where the current generation of tools is falling short for merchants.
I’ve noticed a consistent pattern: AI tools handle the routine queries (order status, FAQs, basic policy questions) reasonably well. But at the moments that matter most—damaged items, lost packages, wrong items shipped, frustrated returning customers—they tend to fall back on “we’ve escalated your case” language that leaves customers feeling dismissed rather than helped.
But I’d love to hear directly from merchants in this community:
1. If you’re using an AI support tool, what’s working? What’s not?
2. When AI has failed a customer interaction, what type of situation was it?
3. If you haven’t adopted AI support yet — what would it need to be able to do before you would?
This is genuine research — I want to understand the real merchant experience, not the product demo version. All responses are helpful.
Good questions. From what I’ve seen working on support automation, here’s the honest version.
What works: anything the AI can ground in the store’s own data. Order status, shipping windows, what the return policy literally says, sizing pulled from the product page. When it’s reading your facts back, it’s reliable and it kills most of the volume. What doesn’t work: the moment it has to make a decision or it can’t find the answer, it either invents something or falls back on the “we’ve escalated” line you mentioned.
Where it fails: the money and judgment cases. Refunds, reships, goodwill credits, “tracking says delivered but I never got it”, and the angry repeat buyer. The deeper failure is ungrounded guessing, the bot stating a policy or a delivery promise that isn’t actually true, because that one reply costs more trust than the 50 it deflected.
What it would take before I’d trust it on a store: three things. It only answers from the store’s own data and cites where the answer came from, so it can’t make things up. Anything that costs money sits behind a human approval step, the bot drafts the reply, a person clicks send. And escalation hands the human the full context and a suggested response, not a dead “case #1234 has been escalated”. That last one is the gap almost nobody closes. “We’ve escalated” is where the customer feels dismissed, so the handoff itself has to be designed, not bolted on.