AI agents for customer service: what they handle, what they cost, and how to roll one out

An AI agent for customer service answers customers from your own help center, policies and order data, carries out routine requests within limits you set, and hands everything else to your team with a summary. Customer service is the most proven use of AI agents in 2026, because the work is high in volume, most of it follows written rules, and results are easy to measure.

This guide is for the person who runs support or operations. It covers what an agent can take on, what it should not, how to choose between your help desk's built-in agent and a custom build, and what it costs.

#What a support agent handles well

• Order and booking status. "Where is my order?" is the single most common request in many businesses. The agent looks it up and answers with the real tracking status.
• Returns, exchanges and cancellations inside policy. It checks the order against your rules, starts the return and sends the label. Anything outside policy goes to a person.
• Account help. Password resets, email changes, plan questions, invoices and receipts, after confirming who it is talking to.
• How-to questions. Answers from your help center and product documentation, with a link to the source.
• Changes to appointments and deliveries. Reschedule, change an address before dispatch, add a note for the driver.
• Triage and routing. Reads every incoming request, tags it, sets the priority and sends it to the right queue. This helps even when the agent answers nothing itself.
• Handover summaries. When a person takes over, they get the customer's issue, what has been tried and the relevant order details in a few lines.
• Out-of-hours and other languages. The agent covers nights and weekends, and replies in the customer's language.

#What should stay with your team

• Upset customers. An agent can recognize anger or distress. It should then hand over, not try to manage it.
• Exceptions. Making an exception for a loyal customer is a business decision.
• Large refunds and credits. Set a limit. Above it, a person approves.
• Complaints with legal or safety weight. Injury, fraud, chargebacks, threats of legal action.
• Vulnerable customers. Anyone who seems confused, in difficulty or at risk.

A good agent is judged as much by how well it hands these over as by how many requests it closes.

#The four parts of a support agent

1. Knowledge. Your help center, saved replies, policies and product information. Contradictions between them are the most common cause of wrong answers, so tidying this is usually the first job.
2. Data access. Read access to orders, bookings, subscriptions and customer records. Without it the agent can only talk in general terms.
3. Actions. A short list of things it may do, each with its own rule: start a return within 30 days, change an address before dispatch, refund up to a set amount.
4. Channels and handover. Web chat and email are the usual start. Messaging apps come next. Handover should move the whole conversation to your help desk, not ask the customer to start again.

#Built-in agent or custom build?

Most help desk platforms now include their own AI agent. Zendesk, Intercom, Freshdesk, HubSpot, Salesforce and Gorgias all do. That changes the first question from "should we build one?" to "is the one we already have enough?"

Start with the built-in agent when:

• You already use one of these platforms and your team is happy with it.
• The agent can reach the data it needs through the platform's existing integrations, for example your Shopify store.
• Your volume is modest, so per-resolution fees stay small.

Look at a custom agent when:

• Your order, booking or policy system is your own software, an ERP or an industry system the platform does not connect to.
• The agent needs to take actions with rules that are specific to your business.
• You run several brands, regions or languages with different policies.
• The agent should live inside your own product or app.
• Per-resolution fees at your volume would exceed the cost of a build. At 4,000 resolved conversations a month and about $1 each, that is around $48,000 a year.
• You have requirements about where customer data is processed and stored.

A common arrangement is both: the platform's agent for general questions, and a custom agent for the process that is particular to you, handing over into the same help desk. The differences between the two kinds of tool are covered in AI agent vs chatbot.

#Guardrails that protect customers

• Say it is an AI. It is good practice everywhere and a legal requirement in some places.
• Always offer a person. One message, such as "talk to a person", should be enough at any point.
• Confirm identity before account changes. The same checks your team uses.
• Never invent policy or promises. The agent quotes delivery dates, prices and terms only from your systems. If the answer is not there, it says so and hands over.
• Limits on money. Refunds and credits up to a set amount. Above that, approval.
• Care with personal data. Card numbers and sensitive details are kept out of the conversation and out of the logs.
• Everything logged. Conversations, actions and costs, so a wrong answer can be traced and fixed.

#How to roll one out

1. Pull three months of tickets and sort them by type. A few categories usually account for most of the volume. Pick the largest one with clear rules.
2. Fix the knowledge for that category. One correct, current answer per question.
3. Run in draft mode. The agent writes the reply and a person sends it. This costs little, and it shows you what the agent gets wrong.
4. Turn the misses into tests. Every wrong draft becomes a case in the evaluation set.
5. Go live for that category only, with handover for everything else.
6. Add the next category once the numbers hold for a few weeks.

#The numbers to watch

Be wary of any target for resolution rate that is set before you have looked at your own tickets. The right number depends on how many of your requests follow rules.

#What it costs

• Built-in agents are usually charged per seat or per resolved conversation. List prices of around $1 to $2 per resolution are common. At a few hundred conversations a month that is small. At several thousand it is a real budget line.
• A custom support agent typically costs $8,000 to $40,000 to build, depending on how many systems it touches and what it is allowed to do. Running costs are often a few cents per conversation in model usage, plus hosting and a few hours of upkeep a month.

The full breakdown, with a payback calculation you can copy, is in how much does an AI agent cost.

Our own custom AI agents start from $9,800 for one well-defined job and take three to five weeks, including the evaluation set, logging, a cost dashboard and a 30-day fix window. Every project is one fixed price, agreed before work begins.

#Where to start

Count your tickets by type for the last three months. If one or two categories dominate and the answers follow rules, an agent will help. If your help desk's own agent can reach the data it needs, switch that on first. If it cannot, or the numbers point to a custom build, book a free 20-minute call and we will tell you what it would involve. For a wider look across the business, the AI Opportunity Audit is a one-week review for $1,500.

This guide is part of our series on AI agents for business.

Common questions.

• What can an AI agent do in customer service?

  It can answer questions from your help center and policies, look up orders, bookings and accounts, carry out routine requests within limits you set, such as a return, an address change or a reschedule, sort and route incoming requests, and hand over to a person with a summary of the conversation so far. It works around the clock and in many languages.

• How much does an AI customer service agent cost?

  Agents built into help desk platforms are usually charged per seat or per resolved conversation, often around $1 to $2 per resolution at list price. A custom support agent typically costs $8,000 to $40,000 to build, plus running costs that are often a few cents per conversation. Which is cheaper depends mostly on your monthly volume.

• Will customers accept talking to an AI agent?

  Most customers care about getting the right answer quickly more than about who gives it. What they dislike is being trapped. Say clearly that they are talking to an AI, make it easy to reach a person at any point, and hand over without making them repeat themselves. Satisfaction scores for agent-handled requests are the number to watch.

• Can an AI agent handle phone calls?

  Voice agents exist and are improving quickly, but they are a separate piece of work from chat and email. Speech adds latency, accents, background noise and interruptions, and rules on automated calls apply. Most businesses get a faster and safer return by starting with chat and email, then adding voice for a narrow job such as appointment booking.

• How do you stop a support agent from giving wrong answers?

  Ground it in your own content so it answers from your help center and records and shows the source. Instruct and test it to say it does not know. Keep sensitive actions behind an approval step. And build an evaluation set from real past tickets, with the answers you would expect, that is run before launch and before every change.