AI Agent for Complaints and Product Returns: What It May Decide and What It May Not

A complaint is the one situation in which a customer contacts you only after they have already paid and something has gone wrong. Expectations are high, patience is low, and every answer carries legal and financial weight. Most e-shops handle this with an inbox, a shared spreadsheet, and one person who remembers how it is done. It works reliably — until volume grows, until that person goes on holiday, or until someone who happens to be right starts complaining.

That is why complaints and returns almost always end up on the list of automation candidates. And they are almost always approached from one of two extremes: either nothing is done because "it is sensitive", or a chatbot is deployed that cheerfully confirms to the customer that they will get their money back — even though nobody approved it. The sensible path runs down the middle and starts with a single question: what exactly may the agent decide, and what should it merely prepare for a human?

The question "where is my parcel" has one correct answer and zero consequences if the agent repeats it three times. A complaint differs from it in three ways at once.

It is legally bound. Consumer rights in complaints and withdrawal from a contract are set by law, and the scope of those obligations differs by jurisdiction, by product type, and by whether the buyer is a consumer or a business. Have the exact wording confirmed by a lawyer — you cannot write "roughly" into a system. Automation here does not mean the agent will interpret the law. It means the rules your lawyer confirms will be followed the same way on Monday morning and on Friday at ten in the evening.

The customer is emotional. They arrive disappointed and often already convinced that the company will try to wriggle out of it. The tone of the answer matters more here than its speed. An agent that replies in three seconds but snaps back with a template does more damage than a human who writes back an hour later with an explanation.

A wrong answer costs money. If the agent incorrectly confirms a claim, the company either pays something it did not have to, or it retracts its promise — the most reliable way to turn a routine complaint into a review and a regulatory filing.

A chatbot that gets your opening hours wrong causes confusion. A chatbot that gets a complaint wrong creates a liability.

The dividing line is simpler than it looks. The agent may do everything that is collecting, verifying, looking up, and informing. It may not do what is assessing and deciding.

Intake and completeness checks

Most complaints are not delayed by the assessment but by the fact that the first submission is incomplete. There is no order number, no photo of the defect, no description of when and under what circumstances the problem appeared, or no account for the refund. What follows is a two- or three-email carousel that takes days. The agent runs a structured dialogue, asks for the missing details right at submission, and passes the case on only once it is complete. It is the dullest benefit on the list and usually the largest.

Order, proof of purchase, and warranty lookup

The agent can pull the order history, delivery note, serial number, date of receipt, and whether the same item has been claimed before out of the e-shop, the ERP, and the warehouse system. The precondition is that those systems can talk to each other — without connecting the ERP, CRM, and e-shop the agent is just another chat window that knows nothing.

Status updates

A large share of the communication around complaints is the question "what is the status". The agent can answer it at any time and, more importantly, proactively: when the goods are received, when they are sent to a service partner, when the status changes. A proactive message removes the question before it arises.

Document generation

Complaint protocols, receipt confirmations, return labels, pre-filled withdrawal forms. What matters is that documents are produced from approved templates by filling in data, not by freely generating text. Legal documents are not the place for a language model to be creative.

Routing to the right department

The difference between a mechanical defect, transport damage, a wrongly delivered item, and a return without giving a reason determines who handles the case and by what procedure. Classifying a submission and placing it in the right queue is exactly the kind of task where a language model helps and where a mistake means only a re-route, not a liability. The same logic applies to the broader picture of automating customer support with tickets and escalations.

The list is short and should not get any shorter:

• Judging whether a claim is justified. Whether it is a defect, wear and tear, or misuse is an expert judgement — often requiring a physical inspection of the goods.
• Exceptions and goodwill. Deciding to accommodate a customer beyond your obligations is a commercial decision with a price tag. It has to be made by someone who bears that price tag.
• Disputed and escalated cases. If the customer disagrees with the outcome or mentions a regulator, the case moves to a human immediately.
• High-value cases. Above an agreed amount a human always decides, no matter how clear-cut the case looks.

This is the one part that is not worth solving with an instruction in a prompt. It has to be written into the system.

Promise the process, never the outcome. The agent may say "we have received your complaint and a technician will assess it"; it may not say "we will refund you". The difference between those two sentences is the entire point of the project.

Rules decide, the model does not. The model translates between the customer and the system. Claims are evaluated by deterministic logic over the order data. If something does not fit the rules, it is not a case for a creative answer but for escalation.

An uncertainty threshold and mandatory escalation. Define when the agent stops answering and hands over: ambiguous classification, an emotionally charged conversation, a mention of legal action, a repeat submission on the same order. This consideration is the core of the choice between autonomous agents and agents with human oversight.

Auditability. Every submission needs a record of what data the agent saw, what it wrote, and exactly where a human took over. Without it, a disputed case cannot be reconstructed six months later.

Data handling. Complaints contain names, addresses, account numbers, and sometimes photographs from inside the customer's home. Limit what the agent sees at all to the necessary minimum and set retention periods — so when designing the solution, account for what processing personal data with AI agents under GDPR requires.

The complaints process runs on deadlines — for resolution, for a reply, for a refund. Their exact length depends on the jurisdiction and the type of case, so it belongs in a configuration confirmed by a lawyer, not in code. What does belong in the system is tracking them.

At intake the agent can calculate the checkpoints, warn the responsible person before a deadline expires, escalate a case stuck at a service partner, and flag cases nobody has opened. That is work done today by a calendar and someone's memory. If a company automates just one single thing about complaints, it should be this — it protects both the customer and the company, and it requires no decision-making authority at all.

Move in phases and let each one run long enough to expose its mistakes.

1. Read-only. The agent answers status questions and explains the procedure. It changes nothing.
2. Intake and completeness. The agent collects submissions, checks completeness, and opens cases. At this stage a human can still approve the replies.
3. Documents and routing. The agent generates protocols from templates and places cases into queues.
4. Narrowly defined automatic resolutions. For example, a return without giving a reason for unopened goods up to an agreed amount — always by a rule, never by the model's judgement.

In most e-shops it pays to stop after phase three. Phase four only makes sense once you have several months of data on how often the agent would have got it wrong. At INTERFASE we build these solutions in exactly this order, because the first two phases deliver most of the benefit at a fraction of the risk — you can read more about the approach in our AI agents and automation section.

An AI agent for complaints is not a decision machine. It is a layer that makes sure the submission arrives complete, the case reaches the right person, the customer knows where things stand, and no deadline slips. Judging whether a claim is justified, granting exceptions, and offering goodwill stay where they belong — with the people who are accountable for them.

If your error rate in complaints is low and the volume small, automation can wait. But if cases are piling up, deadlines are watched by hand, and half the communication is chasing missing details, this is a clear candidate. If you want to walk through which steps of your complaints process can safely be handed to an agent and which definitely cannot, get in touch and we will go through it on your actual numbers.