What a complaint actually is
Underneath the words, a complaint is a request for two things: acknowledgement that something went wrong, and a commitment that somebody will look at it. Notice that neither is information. The complainant is not short of facts about your policy, they have usually read it, and reading it again is what made them angry.
This is why the automated answer fails so badly. Retrieval finds the policy page, because a complaint is full of vocabulary from the policy page, and the policy page is genuinely the closest match to the words in the message. The system then answers the question nobody asked, in a confident and slightly cheerful register, which reads as a refusal to engage.
The escalation from there is predictable. The next message is angrier, is often copied to a public channel, and now contains a screenshot of your business telling a person with a legitimate grievance to consult the terms. The automation did not cause the original problem, but it converted a solvable one into a story.
The three things an automatic reply can safely contain
Acknowledgement, first, and specific to what they said rather than generic. Not an apology for the inconvenience, which is a phrase that signals a form letter. Something closer to a plain confirmation that they have described a problem with a specific thing and that it is being passed on.
Routing, second, stated as a fact rather than a possibility. Somebody will look at this. Ideally a role rather than a mystery: the team that handles this, the manager, the named person if you have one. A complainant wants to know their message has left the automated layer, and saying who it is going to is what proves it.
An expectation, third, in the form of a time. The same discipline applies as anywhere else: a real time you can keep, not a reassurance. Complaints are the category where a missed expectation compounds fastest, because the complainant is already in a frame of mind where the delay confirms what they suspected.
Nothing else. No explanation of the policy, no reasoning about why the thing happened, no suggestion of what they could have done differently. Everything beyond the three items above is the assistant putting words in your business's mouth about a matter under dispute.
Liability is not the assistant's to accept or deny
There are two symmetrical failures here and both are expensive. An assistant that concedes fault has created a written admission on behalf of a business that has not yet looked at the facts. An assistant that denies fault has created a written rejection before anybody investigated, which is the sentence that turns a complaint into a formal one.
Both come from the same root: the system is composing on a topic where the correct position is not yet known. The safe behaviour is not careful wording, it is silence on the question of fault. Say a person is looking at it. Do not characterise what they will find.
This has an implication for the indexed material, not just for the escalation rule. If your published pages contain confident statements about where responsibility sits, retrieval will surface them into complaint conversations, and the assistant will effectively argue your legal position without being asked to. Review anything in the material that reads like an assignment of blame, and consider whether it needs to be phrased for a reader who is currently upset.
Naming the regulated route
In several industries, complaint handling is defined rather than optional. Financial services, insurance, utilities, telecoms, healthcare, letting and property, travel and legal services all have some form of prescribed process in many countries, often including a defined final stage where the complainant can go to an external body if the business does not resolve it.
Where that applies to you, the escalation route is not a design choice. It has a name, and the name belongs in your complaint handling, including in whatever the automated layer says. Getting this wrong is a compliance failure independent of whether the underlying complaint was justified.
Two practical notes. The correct external body differs by country and often by product line, so do not let a single generic sentence cover a business that operates in more than one jurisdiction. And the requirement usually attaches to a specific stage of the process rather than to first contact, so the right move for an assistant is normally to route to the person who owns the procedure rather than to recite the whole procedure itself.
Recognising a complaint before it is called one
Almost nobody opens with the word complaint. They open with a description of what happened, in the past tense, with a note of what they expected instead. The distinguishing feature is not anger, it is the gap between expectation and outcome stated as a fact.
Because the signal is structural rather than lexical, a keyword rule alone will miss a lot of it. That is an argument for pairing a rule on the obvious markers with a low threshold for handover generally: when the assistant is not confident, the correct behaviour on a message about something that already went wrong is to route rather than to reach for the nearest passage.
There is a useful diagnostic in your own logs. Look for conversations where the assistant answered confidently and the next message from the visitor is short and cold. That pattern is usually a complaint that was answered as a question, and each one is a rule you have not written yet.
Where the record has to go
Complaints have a record keeping dimension that ordinary questions do not. Even outside regulated industries, the ability to say when a complaint arrived, what it said and when it was answered is what protects you later, and it is the first thing anybody asks for when a dispute escalates.
That means an automated layer handling complaints must not be the only place the message exists. If the handover emails somebody and nothing else, a mail failure erases a complaint, and you will not know it happened. Storage that survives the email is not a nice extra here, it is the difference between having a record and believing you have one.
It also means the timestamp matters. In processes with defined response windows, the clock usually starts when the complaint reached the business, not when a person read it. If your assistant took the message at ten on Saturday night, that is when it arrived, and your internal handling should reflect that rather than starting from Monday.