Insights · November 2026 · 7 min

What an agent-run support function looks like

A walk through the support function when an autonomous agent owns it, from the first message to the lesson it takes from each case.

Short answer: An agent-run customer service function reads every incoming message, classifies it, resolves what it can using connected systems, issues refunds or credits within written policy, escalates the rest to people with a full summary, samples its own work for quality, and learns from outcomes such as reopened tickets, satisfaction scores and retained customers.

Customer service is usually the first function companies think of for AI, and for good reason. The volume is high, much of it repeats, and the outcome of each case is visible within days. Most deployments so far have been chatbots that answer common questions and hand everything else to a person. That approach helps at the edges. It leaves the function itself unchanged.

An agent-run support function works differently. The agent owns the case from the moment it arrives until it is closed, and it is measured on the result. Below I walk through what that looks like step by step, then the metrics I would use to judge it.

How does an AI agent handle a support ticket from start to finish?

Intake and triage

Messages arrive by email, chat, web form, social channels and sometimes phone transcripts. The agent reads each one and attaches what it can find: the customer's account, recent orders, past tickets, plan level and any open issues. It then classifies the case by type, urgency and risk. A billing question from a long-standing customer two days before renewal is handled differently from a general question about a feature.

Triage is where a lot of human time goes today, and where an agent is reliably strong. It reads every message in full, applies the same rules each time, and does not get tired at the end of a shift.

Resolution

For most cases the agent can act directly. It can reset access, update an address, explain a charge by pulling the invoice, rebook a delivery or walk a customer through a setting. The difference from a chatbot is that the agent has permission to use the same systems a support representative uses, within limits, and it checks that its action worked before it closes the case.

I describe the general mechanics of how agents plan and use tools in how AI agents work. In support, the important point is that resolution means the problem is fixed, and an answer alone does not count.

Refunds and credits within policy

Money is where companies are rightly cautious. The approach that works is to write the refund policy down in plain rules, with limits by amount, customer tenure and reason, and let the agent act inside those rules. Say the policy allows the agent to refund up to a set amount for a late delivery on a first claim. Anything above that amount, or a third claim in a month, goes to a person for approval.

Writing the policy this way often turns up gaps. Many teams find that their real refund practice differs from the written one, and that different representatives apply it differently. Settling those questions before handover is part of the work, and it tends to make the function fairer to customers.

Escalation

Some cases should always go to a person: legal threats, safety issues, a customer in distress, a request that falls outside any written policy, or a situation the agent has low confidence about. The agent should escalate with a short summary of the case, what it has already checked, and what it recommends. A person picking up the case should never have to ask the customer to repeat themselves.

Good escalation is the clearest sign of a well-built agent. A system that escalates too little causes harm. A system that escalates too much saves little. The threshold should be tuned with real cases during the pilot.

Quality review

A human support team usually reviews a small sample of tickets each week. An agent function can review far more. A second process reads closed cases against the policy and a quality rubric, flags those that look wrong, and sends a sample to the retained team for a final check. Patterns in those reviews become changes to the policy or to the agent's instructions.

Learning from outcomes

Each case produces signals after it closes. Did the customer reopen it? Did they reply with thanks or with a complaint? Did they renew, downgrade or leave? An agent with a feedback loop can connect these outcomes back to the actions it took and adjust its choices over time. That is the reinforcement learning idea applied to support, and I explain it more fully in reinforcement learning for business agents.

What does the human team do in an agent-run support function?

The team gets smaller and its work changes. A few roles remain, and they matter more than before:

In my experience the best people for these roles are the strongest members of the existing team, because they already know where the policy is unclear and where customers get confused.

Which metrics show whether an AI support agent is working?

Many support dashboards track activity. An agent function should be judged on outcomes, because activity is cheap for a machine to generate. These are the measures I would track from the first week of a pilot:

I would avoid putting deflection rate at the top of the list. A system can deflect customers by making it hard to reach anyone, which looks good on a dashboard and costs the company later.

What are the risks of letting AI run customer service?

The main risks are an agent that sounds confident while being wrong, an agent that is too generous or too strict with money, and an agent that fails to notice when a customer needs a person. Each one has a practical answer. Ground every response in the company's own systems and documents. Put money behind written limits and approval thresholds. Make it easy for a customer to ask for a human at any point, and treat that request as a signal worth learning from.

There is also the question of how fast to move. I would hand over one channel or one category of case at a time, starting with the most routine, and widen the agent's scope as the metrics hold. The full sequence is in the buyer's guide to rolling out AI agents, and the cost side is worked through in how AI agents affect the bottom line.

Questions people ask

Can AI agents fully replace a customer service team?

They can take over most routine volume, but a small team should remain to handle sensitive escalations, own the refund and service policy, and review quality. Removing all human oversight is a risk few companies should take.

Can an AI agent issue refunds?

Yes, if the refund policy is written as clear rules with limits by amount, reason and customer history. The agent acts inside those limits and sends anything above them to a person for approval.

When should an AI support agent escalate to a human?

It should escalate legal threats, safety issues, distressed customers, requests outside written policy, and cases where its confidence is low. The escalation should include a summary so the customer never repeats themselves.

What metrics should I use to measure an AI customer service agent?

Track resolution rate without a person, reopen rate, time to confirmed resolution, satisfaction, policy adherence, escalation quality, cost per resolved case and retention of customers who contacted support.

How is an AI support agent different from a chatbot?

A chatbot answers questions and hands off the rest. An agent owns the case, uses the same systems a representative would, takes action within policy, confirms the fix and learns from what happened afterward.

Keep reading