AI Agent for Customer Service: What It Should and Shouldn't Do

Most companies that want an AI agent for customer service picture a bot that answers every customer on its own. That is the last step, not the first. The agents that work start smaller: they read the ticket, find the answer and hand a draft to a person.

I've put 35+ agents into real businesses, and customer-facing work is where I'm most careful, because a wrong answer to a customer isn't a bad draft. It's a promise your company now has to keep.

What does an AI agent for customer service actually do?

An AI agent for customer service reads an incoming message, works out what the customer needs, checks your systems for the answer and then takes one action: label, route, draft or reply. That last part is what makes it an agent instead of a chatbot.

A scripted chatbot follows a decision tree. It asks "Is your question about A, B or C?" and breaks the moment someone types something unexpected. An agent reads the actual message, looks up the order in your shop system, finds the relevant help article and writes a reply that fits.

Here is how the three options compare:

The useful version is mostly plumbing. Read the ticket, call two or three tools, produce one output. The model does the reading and writing. Your systems supply the facts.

Which customer service tasks should you hand to an AI agent first?

Hand the agent the tasks that happen every day, have one right answer and are cheap to get wrong in a draft. Triage and draft replies come first; autonomous answers come last.

I'd roll it out in this order:

1. Triage. The agent labels every incoming ticket by type, urgency and language, and routes it to the right person or queue. Nothing reaches the customer, so the risk is close to zero.
2. Summaries. For long threads, the agent writes a three-line summary at the top: who this is, what they want, what has already been tried. Your team reads that instead of scrolling.
3. Draft replies. The agent drafts an answer using your help content and the customer's order data. A person edits and sends.
4. Direct answers on narrow types. Order status, delivery times, opening hours, password resets. Only the types where drafts already go out unchanged.
5. Actions with approval. Changing an address, resending an invoice, booking a callback. The agent prepares the action, a person clicks approve.

Each step should run cleanly for a few weeks before you add the next. If you want a scoring method for which ticket types to start with, my process scoring method works for support too: how often, how clear, how gladly handed off.

What should a customer service agent never do on its own?

A customer service agent should never decide on refunds, compensation, policy exceptions, legal questions or complaints without a person approving it. Those are judgment calls, and the agent's answer becomes your company's answer.

This isn't a theoretical risk. In February 2024, a Canadian tribunal ruled in Moffatt v. Air Canada that the airline had to honor a bereavement refund its website chatbot had described, even though the actual policy said otherwise. The airline argued the chatbot was responsible for its own words. The tribunal didn't accept that.

So draw a hard line. A simple rule I use: if the reply sends money, promises something, admits fault or deletes data, a person approves it. Everything else can be drafted, and the boring repeat questions can eventually go out on their own.

The same goes for tone. An angry customer who gets a cheerful, generic AI reply gets angrier. Route strong negative sentiment straight to a person.

Can an AI agent replace your customer service team?

An AI agent can take over a large share of repetitive tickets, but it does not replace the people who handle exceptions, complaints and relationships. The realistic goal is a smaller pile of easy tickets, not an empty support desk.

The numbers in the market are big. In February 2024 Klarna said its AI assistant handled 2.3 million conversations in its first month, about two thirds of its customer service chats, which it described as the work of 700 full-time agents. In May 2025, Klarna's CEO said publicly that the company would make sure customers could always reach a human, after admitting the push on cost had hurt quality.

Gartner predicted in March 2025 that by 2029 agentic AI will autonomously resolve 80% of common customer service issues, with a 30% drop in operational costs. Note the word "common". The uncommon ones are where your reputation is made.

My take: plan for the agent to handle volume and your people to handle judgment. If you cut the team before the agent has proven itself on real tickets, you'll be rehiring under pressure.

How do you set up an AI agent for customer service, step by step?

You set it up by fixing your help content first, then connecting the agent with read-only access, running it in draft mode and only widening its permissions when the numbers say so. Here's the sequence:

Step 2 is where most of the work is, and where most projects skip. If the answer to "Can I return a sale item?" only lives in a senior colleague's head, the model will make one up. It'll sound confident. That's worse than no answer.

For the build itself, the seven build steps I use for any agent apply: one input, one output, fewest tools possible, test on real past cases. For customer service, "real past cases" is easy. You have years of tickets with the answers your team actually sent.

How do you know when the agent is ready to answer customers directly?

The agent is ready to answer a ticket type on its own when a person approves its drafts for that type without changes almost every time, over a few hundred tickets. One number, tracked per type.

Don't measure it across all tickets at once. An agent can be excellent on "Where is my order?" and poor on "Your product broke." A single average hides that. Give each type its own score and its own switch.

Once a type is live, keep checking. Sample a handful of sent replies every week. Watch for:

• Answers that quote a policy or price that changed
• Replies that are correct but miss what the customer was really asking
• Tickets the agent answered that should have been routed to a person
• Customers who reply "I want to talk to a human"

That last one is your best early warning. Make reaching a person easy and visible. An agent that traps customers in a loop costs you more than it saves.

What breaks AI customer service agents in production?

AI customer service agents break for the same reasons other agents do: outdated knowledge, access that's too wide, no log of what was sent and no human checkpoint on risky replies. The model is rarely the problem.

The specific version in support:

• Stale help content. You change the return window, nobody updates the help article, the agent keeps quoting the old one.
• Too much access. An agent that can issue refunds "just for small amounts" will eventually issue one it shouldn't.
• No record. A customer says "your bot told me I'd get a discount." Without a log of every reply, you can't check.
• No fallback. The order system times out and the agent answers anyway, without the data.

I wrote up the five failure patterns I see most separately. For support, add one more rule: whenever the agent isn't sure, it hands the ticket to a person and says so. A slower human answer beats a fast wrong one.

Start with triage this week. It touches no customer, it shows you where your tickets really go, and it gives you the data to decide what comes next.