Automating customer complaint management means having an agent read every complaint that comes in, rate its severity and hand the case, already prepared, to the person who has to decide. Once the decision is made, the platform replies to the customer through the channel the complaint came in on. The team still decides what happens in each case, but no complaint is left without an owner or an answer.
In a small or medium-sized company, complaints rarely arrive by a single route. They land in the general inbox, in a salesperson's email, on the reception phone or in the official complaints book, on paper or online. This article is about organising that work internally, not about the specific rules that apply to official complaints books.
What is a complaints agent?
A complaints agent is an artificial intelligence agent that interprets complaints received in writing, identifies the customer and the subject, rates the severity according to the company's rules and hands the case to the person who decides. Steps that affect other systems, such as recording the case or sending the reply, are carried out by the platform, and only within what the team has configured.
The difference from a plain ticket queue lies in the work done before anyone opens the case. With an agent, the case reaches the person responsible with the customer identified, the severity rated and the sentence of the email that each piece of information was taken from.
Unlike orders or bookings, a complaint does not resolve itself. The team steps in on every case to decide what to do. What the agent saves is the reading, sorting and forwarding.
Why do some complaints go unanswered?
When a forgotten complaint is traced back, the cause is rarely a lack of goodwill. Almost always, one of these gaps in the process is to blame.
- Scattering. Complaints are spread across several mailboxes and several people, and nobody sees the whole picture.
- No clear owner. The complaint reaches someone who cannot resolve it and waits for someone else to forward it.
- First come, first served. A serious complaint waits behind a simple question only because it arrived later.
- Missing context. Whoever decides has to look up the order, the invoice or the customer's history before they can reply.
- Criteria in people's heads. What is urgent to one person is not to another, and two identical complaints are handled differently.
How should complaints be triaged by severity?
Severity triage is where automation helps most, because it is the step that currently forces someone to read everything. For it to work, the criteria have to be written down. In a smaller organisation, the most common ones are these.
- Risk to people or property, such as a piece of equipment that fails while in use.
- Impact on the customer's operations, such as a missing part that halts another company's production.
- Amount at stake, such as an invoicing error above a threshold the company has set.
- Repetition, when the same customer has already complained about the same issue.
- Commercial risk, when the customer says they will stop buying or cancel the contract.
With these criteria, the company defines the levels and who handles each one. The table below is only an example.
| Level | Example situation | Who decides |
|---|---|---|
| High | Risk to people, customer's operations at a standstill, amount above the set threshold | Management or the head of the area |
| Medium | Late delivery, invoicing error, defect with no risk | The head of the team that provided the service |
| Low | Dissatisfaction with no tangible harm, a question phrased as a complaint | The customer service contact |
In EngiAIFlow, rules are written in plain language, for instance “a complaint that mentions a risk to people is high severity” or “before replying to the customer, ask for approval”. The company's rules are managed in versioned sets, and each case's history shows which rule was applied. As a result, two identical complaints receive the same rating, and the reason can be traced.
How does automated complaint handling work, step by step?
With EngiAIFlow, Engibots' intelligent process automation platform, the complaints process follows these steps.
- Intake. The complaint arrives by email or from another system through the API, and each complaint becomes a case.
- Reading. The agent interprets the message with artificial intelligence and extracts the customer, the subject and the order or service concerned. Each value keeps the sentence of the email it was read from, and the platform confirms that sentence exists.
- Rating. The agent rates the severity according to the company's rules.
- Handover to the decision-maker. The case reaches the person responsible with the complaint already rated, what has already been done and the decision still needed.
- Reply. Once the decision is made, the platform replies to the customer through the channel the complaint came in on. The case history records what was done at every step and every human decision, with the change and the reason.
Imagine, as an example, a company that supplies and installs commercial kitchen equipment for restaurants. A customer writes to the general inbox to say that the oven installed the previous week has stopped heating and the restaurant opens for dinner that evening. The agent identifies the customer and the equipment and rates the complaint as high severity, because the company's rule gives priority to equipment that has stopped working at a customer who is open for business. The case reaches the head of technical support together with the sentence from the email that justifies the rating, and it is that person who decides to send a technician. The reply goes back to the customer by the same email thread.
What always stays in the team's hands?
A complaint involves two different tasks, understanding what the customer is saying and deciding what the company will do. The first benefits from artificial intelligence. The second stays with a person.
| Task | Who does it |
|---|---|
| Reading the complaint and extracting the customer, the subject and the order concerned | Artificial intelligence, with the source sentence for each value |
| Rating the severity | The agent, with the rules written by the team |
| Deciding what to do about the complaint | A person on the team, in every case |
| Sending the reply through the original channel | The platform |
- Execution limited to what is configured. AI executes nothing by itself. Only the actions declared for each system are executed, in each process's published version.
- Approval before the step. Limits such as “above this amount, ask for approval” suspend the case before the step until a person approves.
- Corrections that teach. When the team links a sender to a customer, later messages from that sender arrive already identified.
- Stop and pause. Whoever administers the installation can stop, with a single command, all external actions and every incoming channel, and the reason is recorded in the audit log. Nothing is lost, and when activity resumes the cases proceed in the order in which they stopped. Pausing the complaints process also suspends its cases that were in progress.
Where should you start?
Complaints are a good process to automate when they arrive by several routes and the organisation can already say who handles each type. Before starting, it helps to have the following in place.
- A mailbox for complaints, or forwarding from the existing mailboxes, so that they all arrive in the same place.
- The most frequent types of complaint, such as delays, defects, invoicing errors and service.
- The severity criteria in writing, with the examples the team already uses to decide what is urgent.
- An owner for each type and each level, named individually rather than by department.
- The most common reply texts, so that replies to customers go out in the company's own tone.
In an installation running in demonstration mode, the process can be tried out with sample emails, with no effect on the real systems.
Automation does not take complaints out of the team's hands. It takes away the searching, reading and forwarding, and leaves the team the part only it can do, which is deciding. How to separate what goes through on its own from what goes to a person is covered in the article on what automation does alone and what people validate.
Frequently asked questions
What is a complaints agent?
It is an artificial intelligence agent that reads complaints received in writing, identifies the customer and the subject and rates the severity according to the company's rules. The case is handed to whoever decides, and the platform replies to the customer through the channel the complaint came in on.
Does artificial intelligence decide how a complaint is resolved?
In EngiAIFlow, no. The team steps in on every case to decide what to do, with the complaint already rated. AI executes nothing by itself, and only the actions declared for each system are executed.
How is the severity of a complaint defined?
With criteria the company writes down, such as risk to people, impact on the customer's operations, the amount at stake, repetition of the problem or commercial risk. In EngiAIFlow, rules are written in plain language, and each case's history shows which rule was applied.
How do complaints reach EngiAIFlow?
Through IMAP and Microsoft 365 mailboxes or from other systems through the API. Each complaint becomes a case, handed to the person responsible.
What is recorded for each complaint?
Each case's history records what was done at every step and every human decision, with the change and the reason, and shows which rule was applied.
Can the complaints process be stopped at any time?
Yes. Pausing the complaints process also suspends its cases in progress. Whoever administers the installation can also stop, with a single command, all external actions and every incoming channel, with the reason recorded in the audit log. Nothing is lost, and cases proceed in the order in which they stopped.