AI triage for customer messages: the urgent ones first, the rest to the right desk
Let an AI router read every WhatsApp, chat or email message, name what it is, and send it to the person, team or AI agent that should handle it, with a reason.
Part of the guide AI employees
AI triage for customer messages is the job a good receptionist does with a pile of notes: read each one, decide what kind of message it is and how urgent, and put it in front of whoever should deal with it. In xMap Flows the AI router does that for every message that arrives on WhatsApp, a chat page or email: it picks one of the outputs you named, such as Emergency, Booking, Billing or Complaint, writes a few words on why, and each output leads to its own next step.
This is for customer service leads, operations managers and owners of service businesses who share one inbox between several people and find that urgent messages wait behind routine ones. It explains how the router reads a message, how to design the outputs, a made-up day at a trades company, and the traps worth avoiding.
What triage changes
Without it, messages are answered in the order they arrive, by whoever happens to be looking. A burst pipe reported at 07:58 sits under twelve questions about invoices. A complaint goes to someone who cannot issue a refund and is forwarded twice. A simple question about opening hours takes a person five minutes to answer.
With triage, every message gets a label the moment it lands, and the label decides the path: some go straight to a person, some to an AI agent that can answer them, some to a shared list for later. People spend their time on the messages that need a person.
How the router reads a message
The AI router takes each incoming message as one item. You tell it what to read (the message text, the sender's name, the subject of an email) or leave that empty and it reads the short text fields. It compares the message with the description you wrote for each output and picks exactly one. Each message gets two fields: route, the output it went down, and route_reason, such as water coming through the ceiling now.
It can only answer with one of your outputs or None of these, so a message never goes somewhere you did not plan. Between two and six outputs fit on one router; each is a dot under the block that you link to what should happen next.
Designing the outputs
The description of each output does most of the work, so write them the way you would brief a new colleague on their first day:
- Name the kind, then give examples. Emergency: a leak, no power, no heating in winter, a gas smell, anything unsafe happening now.
- Say what does not belong. Booking: asks for a visit or to move one. Not a complaint about a past visit.
- Decide what wins when a message is two things. Use the extra instructions box: If a message is urgent and also a complaint, choose Emergency.
- Keep a catch-all. None of these is always there. Link it to a person rather than letting it fall away.
- Stay under six. If you need more, route broadly first and put a second router after one output.
A day at a trades company (made up)
Picture a plumbing and electrical company in Melbourne with two office staff, a manager and eleven technicians. Customers write on WhatsApp, and the website has a chat page. The flow starts from both, and one router sorts every message into five outputs.
- Emergency posts the message to the on-call channel in Slack and asks the on-call technician, through Ask a person, whether they can be there within the hour; the answer comes back into the run.
- Booking goes to an AI agent that offers the next free slots from a table the office keeps in My data, then Reply answers in the same WhatsApp chat or chat page.
- Billing gets a reply with the link to pay or download an invoice, and a copy goes to the office's billing sheet.
- Complaint is emailed to the manager with the reason the router gave, so they read it before calling back.
- Spam and sales pitches are only counted.
On a wet Monday morning forty messages come in before nine. Three are leaks; the on-call technician sees them in Slack within a minute of each arriving. Twenty-two are booking requests, and the agent answers most of them on its own; the office picks up the few it could not place. The manager reads two complaints with their reasons before ten. Nobody scrolled through forty messages to find the three that mattered.
Where an AI agent fits after triage
The router decides where a message goes; it does not answer it. Answering is a separate step, and splitting the two keeps each one simple. An agent behind the Booking output only needs booking tools and booking instructions. An agent behind a Questions output can be given the company's own documents through its Look things up in setting. Neither has to know how to handle an emergency, because emergencies never reach it.
This also keeps sending safe. The agent that writes back to customers can sit behind Check the work, so every reply meets your rules before it leaves, while urgent messages skip straight to a person.
Traps to avoid
- Mixing urgency and topic. Outputs like Urgent, Billing and Normal overlap: an urgent billing message fits two. Pick one way of dividing, and let the extra instructions settle ties.
- Trusting the test preview. In a test run the router sorts by the words in each message, as a preview. Live runs ask xMap AI. Send a handful of real-looking messages through a live run and read the reasons.
- Leaving an output unlinked. Messages that go down a dot with nothing after it are only counted. Check every dot leads somewhere.
- Forgetting the sender. Read the reason, not only the label. A complaint from a long-standing customer and one from a first-time caller deserve different speeds, and the reason often says which it is.
Triage is one of the first jobs worth handing to an AI employee because it is frequent, rule-like and easy to check. The guide to AI employees covers how to give one a job, watch its work and decide how much it does on its own.
Questions
Which channels can the router sort?
Anything that starts the flow with a message: a chat page, a form, an API call, or app events such as a new WhatsApp message or a new email in Outlook.
Can the router answer the customer itself?
No, it only chooses the path. Put an AI agent and a Reply after an output to answer, or an email to a person to handle it.
What happens to a message that fits nowhere?
It leaves by None of these. Link that dot to a person or a shared list so it is still read.
How do we know it chose well?
Every message keeps the output it went down and a short reason. Read a week of them, and sharpen the description of any output that collects the wrong messages.
Templates that do this
Read next
More on ai employees and automation · Every article · Templates