Complaints are where an online store's reputation is won or lost. They are also where small teams lose the most time. A message arrives by email, a contact form, or a chat widget. Someone has to find the order, read the policy, check the courier status, decide what is fair, and write a reply that sounds like a person wrote it. Multiply that by every busy week, and the same few people are doing the same detective work over and over.
This post describes a proposed workflow for an AI assistant that handles that routine work, with a human reviewing every exception. It is a design, not a finished product. The details depend on your shop platform, your courier, and your policies, so any integration and acceptance criteria need to be verified for your specific environment before anything goes live.
The core idea: the assistant prepares, people decide the hard cases
The assistant does not replace your customer service team. It takes over the repetitive steps that come before a decision, and it hands anything unusual to a person with the context already assembled.
How the workflow runs, step by step
Intake. A complaint arrives through whichever channel you already use. The assistant reads the message and extracts what matters: the order reference, the product, what the customer says went wrong, and what they are asking for, such as a replacement, a refund, or an explanation.
Data lookup. Using connections to your shop platform and, where available, your courier's tracking, the workflow pulls the real order record: what was ordered, when, how it was shipped, and what the delivery status says. The assistant works from your data, not from guesses.
Classification. The complaint is sorted into a category, for example damaged in transit, wrong item, late delivery, product not as described, or a warranty question. If the message is unclear, the assistant drafts a clarifying question instead of assuming.
Policy check. The case is compared against your written return and complaint policy. The assistant does not invent rules. If the policy clearly covers the situation, it proposes the matching resolution. If the policy is silent or ambiguous, the case is marked as an exception.
Draft reply. For clear-cut cases, the assistant writes a response in your brand's tone, in the customer's language, referencing the actual order details. Depending on how you configure it, drafts can wait for a quick approval, or routine categories can be sent automatically while everything else waits.
Escalation with context. Exceptions go to a named person with a short summary: what the customer wrote, what the order data shows, which policy section applies or why none does, and a suggested reply. The reviewer decides, edits if needed, and approves. Nothing sensitive is resolved without a human.
Logging. Every case is recorded with the decision and the reason behind it. Over time, your team can see which complaint types repeat, which products or carriers cause friction, and where the policy needs clarification.
Where human review is non-negotiable
Some cases should never be fully automated: a customer who is angry or distressed, a possible legal or consumer-rights dispute, a high-value order, a suspected fraud pattern, or anything involving safety. The workflow should treat these as mandatory escalations. The goal is not to remove judgment but to reserve it for the situations that need it.
Why a connected workflow beats a chat widget
A generic chatbot can answer questions about your return policy. It cannot see that a parcel was marked delivered, that the customer already received a replacement, or that the same buyer filed a similar claim last month. Value comes from connecting the assistant to your actual systems and giving it clear boundaries. That is plumbing work: tools like n8n for orchestration, a backend service for business logic, and a language model such as Claude for reading and drafting. The model is only one part. The integrations, the policy rules, and the review step are what make it dependable.
What you would need to decide first
Before building anything, a few questions shape the design. Which channels do complaints come through? Where does your order and shipping data live, and can it be accessed safely? Is your complaint policy written down clearly enough to follow? Who reviews exceptions, and how quickly? Answering these honestly often improves your process even before any automation exists.
Next step
If you run an online store and complaint handling is eating your team's time, we would be glad to talk through whether this proposed workflow fits your setup. We will look at your channels, your tools, and your policies, and tell you plainly what is realistic. Contact Pragma AI to start the conversation.