eCommerce customer support with AI agents: from ticket to sale
A well-designed AI support agent doesn't just cut costs: it turns tickets into orders, recovers abandoned carts, and feeds the product team with structured insight.
An eCommerce AI support agent is an LLM connected to order, catalog and shipping systems, able to resolve autonomously 60–80% of recurring tickets (tracking, returns, availability, sizes/colors) and escalate only ambiguous cases to humans. What separates a useless chatbot from an agent that actually pays back is real data access — not prompts.
Setup that works: LLM with tool calling into Magento/Shopify APIs (orders, tracking, catalog), carrier APIs (DHL, GLS, BRT), FAQ CMS, RMA system. Every tool has minimum permissions and a rate limit. No access to sensitive data outside scope.
Smart handoff: escalation to the human team is ACTIVE ('let me hand you to a colleague who's seen cases like this') not passive ('I didn't understand, try again'). The ticket lands on the operator already with context, history, mentioned products and estimated sentiment. Average resolution time drops even on human-handled cases.
Most underrated payoff: AI agents produce structured data on what customers actually ask. In 30 days you collect a taxonomy of problems, logistics bottlenecks, SKUs with recurring friction. The product team discovers in two weeks what used to surface at quarterly reviews.
Support-to-sale conversion: if the agent sees an open cart or a customer asking for an out-of-stock size, it proposes alternatives or a notify-when-available flow directly in chat. On a fashion client this added 4.2% support-attributed revenue — turning a cost center into a channel.
What doesn't work: AI agents trained only on static knowledge bases without transactional access. The customer wants to know WHERE their package is, not a generic shipping-time article. Without real tool calling, an AI chatbot is worse than a FAQ.
How we measure the agent's real quality
Automatic resolution rate isn't enough. We track four metrics in parallel: CSAT on agent-closed tickets (must be ≥ human CSAT, not just 'acceptable'), 7-day reopen rate (a 'resolved' ticket the customer rewrites 3 days later is a masked failure), average operator time on escalated tickets (should drop thanks to the context passed by the agent), incremental revenue attributed to the agent's proactive suggestions. If even one metric worsens, the model must be recalibrated before scope is expanded.
Tone of voice and brand alignment
A mistake we often see is leaving the LLM 'default' — neutral, formal, American-translated tone. It makes the brand feel cold and generic. We invest 2–3 days at project start defining a tone-of-voice document (200–300 good/bad reply examples) and inject it as system prompt plus few-shot examples. The outcome: customers writing 'thanks, you've been super kind' without realizing they talked to an AI agent. On luxury and fashion brands this piece is decisive — get the tone wrong and brand equity erodes.
Fallback and transparency to the end customer
We always disclose that the user is talking to an AI assistant (EU AI Act transparency requirement in force from August 2026) and we allow at any point an explicit request for a human operator ('talk to a person' → immediate handoff, no questions). This transparency doesn't hurt metrics — quite the opposite, it raises CSAT because the customer feels in control. Merchants who hide the AI get more complaints when something goes wrong, not fewer.
Every signal comes from a real project.