Short answer
Automate verifiable routine support, preserve human escalation and audit repeat contacts. Scope EU AI disclosure rules to the markets they cover.
The answer in detail
An AI agent can handle a defined share of repetitive e-commerce support: order status, standard returns and factual product questions. Keep exceptions, disputed refunds and angry customers within a human escalation process. Tell shoppers when they are interacting with AI and make the route to a person clear.
The benchmark
The source cites Gartner's March 2025 forecast that agentic AI would independently resolve 80% of common customer-service issues by 2029 while reducing operational costs by 30%. That is a forecast, not a result your team should promise.
A separate Gartner survey of 5,728 customers, reported in July 2024, found 64% would prefer companies not to use AI for customer service and 53% would consider switching to a competitor. The leading concern was losing access to a human. Those findings support testing both resolution and escalation, rather than celebrating automated closures alone.
| Request | Candidate for automation | Escalation condition |
|---|---|---|
| Order status | Reads live order and carrier data | Parcel blocked or lost |
| Standard return | Applies the published policy | Outside the window or damaged item |
| Size, composition, care | Uses verified product facts | Health, allergy or uncertain interpretation |
| Refund exception or goodwill gesture | Prepare context for the team | Human approval under the proposed policy |
| Complaint or distressed customer | Identify and transfer | Prompt human ownership |
This is a proposed division of responsibility, not a claim that software can never execute an approved refund workflow.
For EU-covered systems, AI Act Article 50 transparency obligations generally apply from August 2, 2026. The Commission's guidance distinguishes providers and deployers and includes disclosure for direct AI interaction unless the AI nature is obvious in context. Do not turn that EU framework into a worldwide law or assume every AI Act obligation began on the same date.
The rule
Automate answers that can be checked against authoritative data. Escalate judgment, exceptions and uncertain advice. A health-related brand described in the source declined live generated answers until their reliability could be established.
The tool vendor may have provider obligations, but the merchant should verify the actual customer experience. Display a clear disclosure in the shopper's language. Check what the agent can read and change, and make its permitted actions narrower than a general staff account.
The steps
- Export three months of tickets and classify their reasons. Count status requests, returns, product questions and disputes to estimate the addressable share.
- Connect real systems before promising live answers. Order tracking needs current order and carrier status. Audit the widget too: a field audit in the source found a helpdesk chat adding 47 KB on each page and exposing email and Shopify identifiers in page code. The reported correction used the official integration and loaded chat on demand; the exact fix depends on the implementation.
- Make the introduction explicit. Example: “I'm the brand's AI assistant. I can check an order or help start a standard return. Type ‘human’ whenever you want our support team.” Test that the instruction actually works.
- Define handoff conditions: refund requests outside permissions, complaint language, repeated unsuccessful answers or high-value cases. Send the transcript and a concise summary so the customer does not repeat everything. The source reports a vendor changing its handoff behavior after new topics created duplicate tickets.
- Maintain one approved policy source for shipping, returns and warranties. Use it in the agent, FAQ, customer account and order-confirmation page. Shopify Knowledge Base can supply FAQs to AI discovery services; its App Store listing checked October 29, 2026 lists English as the supported language. Check suitability for the customer languages your store serves. It does not guarantee visibility.
- Review 20 randomly selected conversations weekly. Track the share marked resolved whose customer contacts you again within 48 hours, across channels. Pair that with incorrect answers, escalation success and customer satisfaction; a recontact measure alone misses some failures.
The pitfall
An invented policy can become a documented customer dispute. In Moffatt v. Air Canada, 2024 BCCRT 149, a British Columbia tribunal awarded compensation after misleading chatbot information and rejected the idea that the chatbot was a separate responsible entity. It is a Canadian decision, not an automatic statement of every country's law. The Club's responsibility article discusses the European context separately.
A blocked handoff recreates the survey's main concern. A useful recovery, by contrast, can preserve trust and make a second purchase possible.
What to tell your team
“In this fictional audit, 58% of tickets concern tracking and returns. The agent will cover approved cases, identify itself and hand off exceptions with context. We will monitor incorrect resolutions and 48-hour recontacts, currently 11%, before expanding access.”
