Customer Support Triage with AI: Routing, Drafting and When to Escalate to a Person

A shared support inbox mixes everything: order questions, complaints, invoices sent to the wrong address, password resets, cancellations and the occasional legal threat. Someone has to read each message, work out what it is and pass it on before anyone can answer it. That sorting step is usually where AI helps first, and a few clear rules decide whether it helps or hurts.

This article treats routing, drafting and escalation separately, because each carries a different level of risk.

Routing: let the model read, let rules decide

Routing is the safest place to start, because a wrong label costs a transfer between queues, not a wrong promise to a customer. A language model reads the message and returns a few structured fields: the topic, the language, the product, how urgent it sounds and whether the customer is asking for a person. Facts the model cannot see in the text, such as the customer's contract or open orders, come from your CRM or order system, not from a guess.

Keep the list of topics short and tied to who handles them. If two categories always go to the same team, they are one category. Build the first version from a few months of real tickets, and ask the people who answer them today to label a sample by hand. That sample becomes your test set: routing works when the model agrees with your team on it, not when it looks convincing in a demo. Messages the model is unsure about go to a person, as they do today.

Drafting: answers grounded in your own sources

Drafting is where most of the time goes, and where most of the risk sits. A model can write a fluent reply to almost anything, which is exactly the problem: a confident answer about a refund policy you do not have is worse than no answer. Build drafts only from approved material, such as your help centre, current policies and the customer's order data, and show the agent which source each draft used so it can be checked in seconds. Fine-tuning versus RAG explains how this grounding works.

Drafts that a person approves are also where the evidence is strongest. A study of 5,179 customer support agents found that access to a generative AI assistant increased the number of issues resolved per hour by 14% on average, with a 34% improvement for novice and low-skilled agents and minimal impact on the most experienced and skilled ones. The authors suggest the assistant spread the practices of the best agents to newer colleagues.

Letting replies go out without review is a separate decision. In 2024, a tribunal in British Columbia, Canada, ruled against Air Canada after the chatbot on its website suggested that a bereavement fare could be claimed retroactively, which the airline did not allow. The airline argued it could not be held liable for what the chatbot said. The tribunal called this a remarkable submission and held that the company is responsible for all the information on its website, whether it comes from a static page or a chatbot. Whatever your system tells a customer, your company has said.

When to escalate to a person

Write the escalation rules down before launch and enforce them in code, not in the model's judgement. These are the triggers we start from in our projects.

Money is the clearest line. Under the GDPR, people have the right not to be subject to a decision based solely on automated processing that has legal or similarly significant effects on them. Not every refund decision reaches that bar, but it is rarely worth arguing which ones do. Where the rule applies, the exceptions come with safeguards, including the right to obtain human intervention and to contest the decision. A model can prepare a refund or a rejection. A person makes the call.

Tell customers, and hand over cleanly

Since 2 August 2026, the EU AI Act has required AI systems that interact directly with people to be designed so that those people are informed they are dealing with an AI system, unless that is obvious from the context. In support, the simplest way to do this is a clear line at the start of a chat or in an automatic reply.

The handover is the moment customers notice most. When a ticket moves to a person, it should carry a short summary: what the customer asked, what the system found, what it has already told them and why it escalated. Nobody should have to explain the problem twice.

How to start

1. Measure today. Count tickets per week by channel and topic, the time to first reply and how often a ticket is passed between teams.
2. Run routing in shadow mode. The model labels every ticket while people route as before. Compare both for a few weeks before the model routes on its own.
3. Add drafts for the most common topics. An agent approves or edits every draft. Track how often drafts go out unchanged, edited or discarded.
4. Allow automatic replies only for narrow, low-risk topics, such as order status or opening hours, and only once reviews show a low error rate.
5. Review a random sample every week, including tickets that were never escalated. That is where you find the mistakes the system was confident about.

To estimate what this could save your support team, describe the workflow in the ROI calculator. If you are considering an agent that also takes actions, such as issuing refunds, read AI agents explained first, or browse common workflows by business area.