--- title: Customer Support Agent description: An agent that answers "where is my order," tracking, and return status from live order data — and hands off the moment it is out of its depth. Not a refund machine. Not a chatbot on your help center. url: https://www.joinhorizon.co/ai-agents/customer-support/ publishDate: 2026-09-22 updateDate: 2026-09-22 --- # Customer Support Agent > The same "where is my order" tickets every morning. An agent looks them up in the systems you already run, and stops when the case is not a lookup. Monday, the queue is the same question in different clothes. Where is it. Did it ship. Can I return it. The answers are already in the order system, the carrier feed, and the return policy. The work is copying them. **This week** the agent does that copy: look up, answer from the record, hand off when the record is wrong or the ask moves money. It does not refund. **A person still signs** refunds, cancellations, address changes, and gift cards. **Skip it** if an order-status email or a carrier template already closes most of those tickets. That is a fine outcome. You do not need an agent to restate a rule. This page is the commercial summary. The long build — which lookups, the week-one list, and how you test it — lives in the guides linked below. Those guides keep their own URLs. ## Returns and fraud, on this page "Was it received?" and "did the refund go out?" are the same job as "where is my order": look up a record, answer from it, hand off when the record is wrong. Abuse and refund investigation stay a **section of this family**. They are not a separate page and not a separate product. If the pain is "people keep asking where the label is," start with where-is-my-order. If the pain is "we are eating fraudulent returns," read the fraud and refund-investigation guides first. Those are different lookups and a different approval path. ## What we will not do in week one - Give the agent a refund, a cancel, an address change, or a gift card before a named reviewer owns the queue. - Let it browse the open web "to be helpful" on a tracking question. - Call a help-center chatbot an agent. ## How we start First question: can the agent join an order to a shipment without guessing? If it cannot, that is a data problem, and we will say so. If it can, the engagement is one scoped agent at [eCommerce AI Agents](https://www.joinhorizon.co/services/ecommerce-ai-agents/). AWS is how it stays up. It is not the first question we ask you. ## Problem The queue fills with tracking, delivery, and return-status questions that already have answers in the order system, the carrier feed, and the policy. People spend the morning copying those answers. Then someone asks for "AI that just refunds." That is a helpful model with a refund button and no stop. ## Agent A support agent reads the request, looks up only what it is allowed to see, answers from the live record, and hands off chargebacks, regulators, "it says delivered and I don't have it," and anything that moves money. Week one is lookups plus a person in the queue. The agent does not issue the refund. ## Workflow ### Read the ticket Work out what the shopper is asking — tracking, a delay, the return window, an address — before any lookup runs. ### Look up the order Check the order, the shipment, and the policy by name. If those systems disagree, say so. Do not invent a status. ### Answer or hand off Reply from the record, or open the human queue with the evidence already pulled together. ### Stop on money Refunds, cancellations, address changes, and gift cards wait for a person. The agent drafts. It does not send. ## Systems ### Order and shipment Your order system plus carrier events. One order number the agent can trust — not three aliases it has to guess between. ### Policy and help content Return windows, restocking rules, and the help articles the agent is allowed to quote. ### Ticket system Where the human queue lives. The agent writes a brief into that queue. It does not open a second inbox nobody owns. ## Approval Anything that moves money or changes a fulfilled order is a proposal. A named reviewer sees the evidence and takes the action in the store. A sentence in the prompt is not a permission. ## How the customer support agent stays bounded Look up where-is-my-order, tracking, and return status, then hand off when the case is not a lookup. - **Maturity:** Level 2 — reference architecture - **Risk:** high - **Oversight:** approval - **Starts when:** A shopper ticket is one of the repeat status questions. - **Tools:** A write goes through a router, a permission check, a policy check, and a budget check. The model does not commit it. An approval token lives in tool context, not in the user message. - **Stops for a person:** Anything that moves money or changes a fulfilled order is a proposal. A named reviewer acts in the store. - **Checked by:** The answer matches the order and shipment tools. Disagreement or a missing record is a handoff. - **Untrusted data:** The shopper message, including text that tries to authorize a refund. - **If it fails:** If the customer support agent stops, name the reason: completed, budget_exceeded, timed_out, cancelled, guardrail_blocked, approval_required, tool_failure, verification_failed, or partial_completion. Retry a timeout at most twice. Back off on rate limits. After repeated verification failure, escalate. Stop when the budget is exhausted or a permission is denied. No unbounded loop. Anything that moves money or changes a fulfilled order is a proposal. A named reviewer acts in the store. - **Lifecycle:** Explore, then act, then confirm, then verify. Explore gathers the context for the customer support agent. Act stays reversible. Confirm stops before an irreversible step. Verify is a separate check of the outcome. Reference architecture, not a published production deployment. The shared contract is the [AWS store-agent architecture](https://www.joinhorizon.co/blog/aws-ai-agents-for-ecommerce-factualminds-2026/#operating-contract). ## Outcomes ### A first pass on the repeat tickets Where-is-my-order and tracking leave the queue when the record is clean. Ambiguous cases still reach a person, faster, with the context attached. ### A written stop list Chargebacks, attorneys, regulators, corrections to personal data, and off-policy discounts are named handoffs. They are not leftover "be helpful" cases. ### A record of what it saw Every lookup is logged. You can show what the agent read, what it said, and who approved a change. ## Field-guide posts - /blog/ai-customer-support-agent-ecommerce-2026/ - /blog/wismo-automation-ai-agents-ecommerce-2026/ - /blog/ai-return-fraud-detection-ecommerce-2026/ - /blog/ai-refund-investigation-agent-ecommerce-2026/ - /blog/ai-agent-vs-workflow-automation-ecommerce-2026/ - /blog/human-in-the-loop-ai-agents-ecommerce-2026/ - /blog/ecommerce-ai-agent-readiness-assessment-2026/ ## FAQ ### Is this a chatbot on the help center? No. A chatbot that only searches help articles, and cannot look up the order, is search. Ship it as search. A support agent uses named lookups, a handoff rule, and a stop on anything that moves money. ### Will it auto-refund? Not in week one, and not by default. Refunds, cancellations, and gift cards wait for a person. The field-guide support post says the same thing. This page is not a looser promise. ### Do you have case studies of support agents in production? Not yet. What we can show is the published support and where-is-my-order guides, with copyable lookup lists and test notes. AWS case studies on this site are not stories about agents we delivered. ### Where do returns and fraud sit? On this page, as a section, not as a separate product. "Was the return received?" and "did the refund go out?" belong here. Fraud scoring and refund investigation have their own long guides, linked below. --- *Source: https://www.joinhorizon.co/ai-agents/customer-support/*