AI Customer Support: Triage, Drafting and the Line a Model Should Not Cross

AI customer support is the placement of a model on the reading, sorting and drafting steps of a support queue: classifying every incoming ticket, summarising it, routing it to the right person, and drafting a reply from the company’s own help documents. The line that decides whether it works is not model quality. It is the rule for which tickets a model may answer alone and which it must hand to a person with a draft attached.

Direct Answer

AI customer support pays when the model triages every ticket and drafts replies from your own documentation, and a person handles anything about money, safety, complaints or a promise. Triage removes the reading work: a support agent starts the day with a sorted, summarised queue instead of an inbox. Drafting removes the typing work: the agent edits a reply that already cites the right help article. Autonomous answers are reserved for a defined set of questions with a clear source, labelled as such, with a person one click away. Measured on first-response time and second-contact rate, this is where support teams see the return.

Key Takeaways

• Triage first. Sorting and summarising every ticket is low-risk and removes the most hours.
• “Chatbot” and “agent” are marketing words for how much the system is trusted to act, not two different products.
• Drafts come from your help centre and past tickets, with the source shown. Never from general knowledge.
• Money, safety, complaints and promises are answered by a person, always.
• Label model answers and keep a human one step away. Trust is the asset, and it is lost on the first confident wrong answer.

Is this an “AI chatbot” or an “AI agent” for support?

Both terms describe the same system from different angles, and treating them as a product choice is where rollouts go wrong. The real variable is not the label. It is how much of the account the system can see and how many of its own actions it is trusted to finish without a person.

A chatbot, in the strict sense, answers from a script or a knowledge base. A customer types a question, the system retrieves the closest documented match, and sends a reply. It has no view of the order, the account or the open ticket, so it can describe a refund policy but cannot tell a customer whether their specific refund has been issued.

An agent for customer support can see that record. It reads the ticket, looks up the order or account state, and drafts or takes an action inside the help desk rather than just answering from text. That is the same triage-and-draft placement described above, named differently by vendors selling it. The model decides what a ticket is and prepares the next step; it does not decide, on its own, which of those steps are safe to finish without review.

So “the best AI agent for customer support” is rarely a product question. It is a governance question about that team’s own tickets: which actions are reversible and cheap enough to let the agent finish alone, and which ones — a refund, a cancelled order, a changed address — need a person to confirm before anything is sent. A team that answers that question for its own ten most common ticket types has done the actual work; a team that just buys “the agent with the best reviews” has skipped it.

Example: a customer asks “where’s my order.” A scripted chatbot matches the question to a shipping-policy answer and sends it regardless of whether that order shipped on time. An agent reads the actual order record, confirms the real status, and either answers directly — if that ticket type is cleared for autonomous replies — or drafts the answer with the order status attached for a person to send. The difference a customer feels is not the word used to describe the system. It is whether the reply is about their order or about orders in general.

What can a model safely do in customer support?

The table is the whole design. IBM’s overview of customer service describes the function; the AI question is only which rows the model owns. Most failed support automations skipped the table and let the model answer everything, and the first refund it promised that the policy did not allow ended the project.

How does AI ticket triage work?

Every ticket that arrives is read by the model before any person sees it. The model produces four things: the topic, the urgency, the sentiment, and a two-line summary. It then routes the ticket to the queue that matches, attaches the summary, and, for the ticket types it is allowed to draft, adds a draft reply that quotes the relevant help article.

The agent’s morning changes from “read forty emails and decide what they are” to “work through a ranked queue where each item says what it is and has a draft attached”. Nothing has been sent. Nothing has been decided. The reading and typing have been removed, and those were most of the hours.

Escalation is a rule, not a judgment call by the model. Any ticket mentioning a refund, a charge, a safety word, a legal word, or carrying strong negative sentiment goes straight to a person with no draft sent. IBM’s human-in-the-loop overview describes the general pattern; the support-specific placement rule — which actions stay human regardless of how capable the model gets — is in Where human-in-the-loop actually belongs in an AI workflow. The escalation list is written by the team, and the model cannot override it.

How to set up AI customer support

1. Export a month of tickets and classify them by hand. Topic, urgency, whether the answer exists in the help centre. This tells you which rows of the table are worth automating.
2. Record the baseline. First-response time, resolution time, second-contact rate, agent hours per week on reading and typing.
3. Write the escalation list. The words and ticket types that always go to a person. Keep it short and absolute.
4. Turn on triage only. Classification, summary, routing. No replies. Run for two weeks and check the classification against the team’s judgment.
5. Add drafting for the sourced rows. The model drafts from the help centre with the citation visible. Agents edit and send. Developers wire the help-desk tool through its API using the model provider’s tool-calling; OpenAI’s function calling guide is the reference.
6. Turn on autonomous answers for a defined set only. The ten “how do I” questions with the clearest documented answers, labelled, with a “talk to a person” link.
7. Measure again after a month. The same numbers, plus customer rating on model-assisted tickets versus human ones.

What AI customer support is not

It is not a chatbot that replaces the support team. The team still owns every decision that touches money, safety or trust, and every ticket the model could not source.

It is not a way to hide the queue. A model that answers instantly with something plausible and wrong has not resolved a ticket. It has created a second one and lost some trust in between.

And it is not a general-knowledge model answering from what it remembers. Every answer comes from your documents or your records, with the source visible, or it does not go out. The fuller, action-taking version of this — where an agent can also change the order, not just describe it — is covered in AI customer support agent.

Summary

AI customer support works as triage plus sourced drafting, with a written escalation list that keeps money, safety, complaints and promises with a person. Whether a vendor calls it a chatbot or an agent, the question that matters is the same: which of its actions are reversible enough to run without review. The model reads and sorts every ticket, drafts the replies that have a documented answer, and answers alone only a defined set, labelled, with a person one click away. Measure first-response time and second-contact rate for a month before and after. The wider library is under AI for business; the engagement that builds this inside your help desk is operations and internal AI.