AI Customer Service Agent: How to Set One Up

An AI customer service agent is software that reads a customer's message, looks up the facts in your systems (order status, account details, policies), and either resolves the request or hands it to a person with the context already gathered. Set up well, it takes the repetitive half of your support queue: where is my order, how do I reset this, can I change my booking, what is your returns policy. Set up badly, it becomes the thing customers complain about on review sites. The difference is not which model you use. It is which requests you give it, which systems it can see, and how easy it is to reach a human.

This post is part of our guide to rolling out agents, which starts at wizeb.com/blog/how-to-implement-ai-agents-in-business. If support is one of several jobs you are weighing, wizeb.com/blog/first-ai-agent-for-business helps you pick which to build first.

The Problem: Support Volume Grows Faster Than Your Team

Every growing business hits the same wall. Orders double, and so do the emails asking about them. You hire another support person, they take six weeks to get up to speed, and by then volume has grown again. Most of what they answer is the same twenty questions, looked up in the same three systems.

Analysts expect agents to take a large share of that work. Gartner predicts that by 2029 agentic AI will autonomously resolve 80% of common customer service issues without human intervention (gartner.com/en/newsroom/press-releases/2025-03-05-gartner-predicts-agentic-ai-will-autonomously-resolve-80-percent-of-common-customer-service-issues-without-human-intervention-by-20290). That is a forecast, not a result, and the word "common" carries a lot of weight.

The other side of the ledger matters just as much. A Gartner survey of 5,728 customers found that 64% would prefer companies did not use AI for customer service, and the top worry, at 60%, was that it would be harder to reach a person (gartner.com/en/newsroom/press-releases/2024-07-09-gartner-survey-finds-64-percent-of-customers-would-prefer-that-companies-didnt-use-ai-for-customer-service). Customers are not against fast answers. They are against being trapped. Design for both facts at once.

AI Customer Service Agent vs Chatbot

The old support chatbot matched keywords to scripted answers and sent you a help-centre link. An AI customer service agent does three things a chatbot cannot:

• It looks things up. It checks the order in Shopify, the subscription in Stripe, the ticket history in your help desk, and answers with the customer's actual situation, not a generic article.
• It takes actions within limits you set. It can update a delivery address before dispatch, rebook an appointment, issue a refund under a set amount or resend an invoice.
• It knows when to stop. It hands over to a person when the request is outside its rules, when the customer is upset, or when it is not confident, and it passes along a summary so the customer does not repeat themselves.

Our explainer at wizeb.com/blog/ai-agent-vs-chatbot-vs-automation covers the difference in more detail.

What Your AI Support Agent Should Handle (and What It Should Not)

Pull the last three months of tickets and sort them by type. The right starting set is high-volume, answerable from your systems and low-risk if something goes slightly wrong.

Good first requests

• Order and delivery status ("where is my order?" is often 20–40% of e-commerce and distribution tickets).
• Account and access questions: password resets, login help, updating contact details.
• Policy questions: returns, warranties, opening hours, pricing tiers, shipping regions.
• Simple changes: rescheduling a booking, changing an address before dispatch, cancelling within a free window.
• Triage: reading every new ticket, tagging it, pulling the relevant order and drafting a reply for a person to send.

Keep with people (for now)

• Complaints, threats to cancel and anything emotionally charged.
• Refunds or credits above a set amount, and anything that changes a contract.
• Legal, medical, safety or regulated advice.
• Your top accounts, where a personal relationship is part of what they pay for.

How to Set Up an AI Customer Service Agent in Six Steps

1. 1Pick the channel and the job. Start with one channel (usually email or web chat, where you can review answers before they go out) and one request type. Phone comes later; a voice agent is the same design with telephony on top, as wizeb.com/blog/voice-ai-agents-business-case-2026 explains.
2. 2Connect the systems it needs, read-only first. Help desk (Zendesk, Freshdesk, Gorgias, Help Scout), order or booking system, CRM. Give it the narrowest access that does the job; write access comes after it has earned trust.
3. 3Write the rules down. What it may answer, what it may change, refund limits, tone, and the exact conditions for handing over. Most of this lives in your best support person's head; spend a morning getting it out.
4. 4Load the knowledge. Help-centre articles, policies, product data and a few hundred good past replies. Fix contradictions before you load them, or the agent will repeat them. Our guide at wizeb.com/blog/train-ai-agent-on-company-data covers this step.
5. 5Test on real past tickets. Run it against 200–300 resolved tickets and score every answer: correct, correct but needs editing, or wrong. Do not launch below the pass mark you agreed in advance.
6. 6Roll out in stages. Draft-only first, where your team approves every reply. Then let it send answers for the request types that scored highest, with a person checking a daily sample. Widen the scope one request type at a time.

What It Costs

Off-the-shelf help-desk AI add-ons are usually priced per resolution or per seat. They are quick to switch on and a good fit if your requests are standard and your data already lives in that help desk. Check the per-resolution price against your volume: at a few thousand resolutions a month, those fees can pass the cost of a custom build within a year.

A custom support agent that reads your help desk and one or two other systems, drafts or sends replies and hands over cleanly is typically a single-task project: about $5,000 to $15,000 to build over three to six weeks, plus a few hundred dollars a month to run. Add more systems, more actions it can take on its own or a phone channel and it moves into the next tier. The full breakdown, line by line, is in our cost guide at wizeb.com/blog/ai-agent-cost-guide.

The Metrics That Tell You It Is Working

• Resolution rate: the share of conversations closed without a person, that did not reopen within seven days. Deflection that comes back as a second ticket is not resolution.
• Accuracy on a weekly sample: have a team member score 30–50 agent replies a week.
• First response time and time to resolution, before and after.
• Customer satisfaction on agent-handled conversations versus human-handled ones. If it is lower, look at the handover, not the model.
• Cost per resolved ticket, including model, help-desk fees and review time. Our guide at wizeb.com/blog/how-to-measure-ai-agent-roi shows how to set up the comparison.

Example: A 25-Person Online Retailer

This is an illustrative example, but the numbers are typical. A UK online homeware retailer with 25 staff gets about 3,000 support emails a month through Gorgias. Three people answer them. When they sort a quarter of tickets, 38% are order-status questions, 14% are returns-policy questions and 9% are address changes before dispatch.

They build an agent that reads each new email, checks the order in Shopify and the tracking with the carrier, and drafts a reply. For two weeks the team approves every draft; 91% go out unedited on order status, 84% on returns. They switch on auto-send for order status and returns, keep address changes on draft-and-approve, and route anything mentioning "damaged", "refund" or "complaint" straight to a person with the order attached.

Three months in, the agent resolves about 1,300 emails a month on its own, first response time on those drops from around five hours to under two minutes, and the team spends the time it saved on damaged-item claims and wholesale accounts. The build cost about $11,000 and runs at around $400 a month. The lesson is not the exact numbers: it is that the agent started narrow, earned auto-send one request type at a time, and never blocked a customer from reaching a person.

Common Mistakes to Avoid

• Launching on every request type at once. You cannot tell what is failing, and one bad week sets adoption back months.
• Hiding the human option. It raises complaints and churn, and it is the first thing customers mention in reviews.
• Loading a help centre that contradicts itself. The agent will confidently quote whichever article it finds first.
• Giving write access on day one. Earn it with a scored test set and a draft-only period.
• Measuring deflection instead of resolution. A ticket the customer reopens tomorrow cost you twice.

How WizeB Builds AI Customer Service Agents

WizeB builds custom AI agents, workflow automations and voice agents for growing businesses in the US, UK and Europe. For support, we start with your own ticket history: we sort it, pick the request types worth automating, connect the agent to your help desk and order or booking system with the narrowest access that works, and score it against your past tickets before it sends a single reply. You get a fixed price and timeline up front, a monthly running cost at your real volume, and you own the prompts, workflows and test sets. See how we work at wizeb.com/services/ai-agents.