Guides

How to Use AI in Customer Service: A Step-by-Step Guide

Portrait of Luke Tan

Luke Tan

Last update

Summary

  • Answer repeat questions from your help center first. Repeat questions pay off soonest.

  • Pick the AI in your help desk, a no-code agent, or a developer’s build, depending on the tools you already have.

  • Set up one use per step, and ask it something your docs don’t cover before it goes live.

  • In the first week, read what it couldn’t answer and add each answer once.

To use AI in customer service, start with the questions your help center already answers, connect an AI customer service agent to those pages, and decide when it hands a customer to a person.

That order matters because the agent can only answer from pages it can read.

So the first use to switch on is answering repeat questions, the route depends on the help desk you already pay for, and the last check before launch is a question your docs don’t cover.

Why AI customer service goes wrong when you start with everything

The obvious way to start is to switch on every AI feature a vendor’s homepage promises, such as drafted replies, routing, summaries, and an agent that answers customers. It stalls on day one, when customers ask what your pages don’t cover.

The customer who asks whether you deliver to the Azores meets an agent with no page to answer from, and every feature you switched on has nothing to work with. The vendor homepages won’t warn you about that.

In September 2026, Wonderchat read every promise on twelve competitor homepages, and on September 10 none of the twelve promised to show the questions it couldn’t answer.

In Wonderchat’s study of 5,848,078 conversations, a sample of 5,000 unresolved conversations traced 58.5% to an answer never documented, 12.6% to a question outside the configured scope, 8.3% to an ambiguous question, and 1.2% to out-of-date content, while 19.3% could not be classified.

About six in ten unresolved conversations asked for an answer nobody had written down.

Self-service has the same blind spot. A Gartner survey of 5,728 customers in December 2023 found that only 14% of customer service and support issues were fully resolved in self-service, a figure that covers all self-service, AI included.

So start narrow. Pick one use, point it at your own docs, and put a person behind it for everything else.

7 ways to use AI in customer service, ranked by what pays off first

The benefits of AI in customer service start with the use your docs can already carry. Seven common uses rank by how soon they pay off, with the questions you answer every day first and routing, which needs several queues, last.

#

Use

What it needs

What you get first

Where it stops

1

Answer repeat questions from your help center

Articles that answer them

Fewer of the same questions reach a person

Questions no page covers

2

Hand over to a person with the whole chat

Someone to take it, or an email for follow-up

The customer doesn’t repeat themselves

Off hours, unless email follow-up is on

3

Find the questions your docs don’t answer

A short daily look at the misses

A list of the pages to write next

Only what customers actually asked

4

Answer in the customer’s language

The same docs, in one language

Answers in 40+ languages without translating pages

Policy wording still needs a person’s check

5

Draft replies for your team

A help desk with an AI add-on

Faster first drafts on tickets

A person still reads and sends

6

Summarize long ticket threads

The same add-on

Quicker handoffs between teammates

Pays off only with long, shared tickets

7

Tag and route tickets to the right queue

Volume and more than one queue

Little at three people

Needs categories worth routing

1. Answer repeat questions from your help center

An AI agent reads your help center and answers the questions customers ask most, with a link to the article it used. A customer asks, “How do I change the email on my invoices?” and the agent replies from the billing article and links it.

What it needs is the article. When no page covers a question, the agent has nothing to answer from, so it says so and hands the customer to a person.

2. Hand over to a person with the whole chat

When the agent can’t help or the customer asks for a person, the conversation moves to your team with the full chat attached. Nobody asks the customer to start over.

In Wonderchat, a handover starts when the customer asks for a person, after a set number of messages, or after repeated failed answers. You switch these rules on under Handover, in Step 2.

3. Find the questions your docs don’t answer

Every question the agent couldn’t answer points to a page you haven’t written yet. In Wonderchat, the Review page collects those questions in one list, so your team adds each answer once and the agent uses it from then on.

Some tools sell a gap view inside their dashboard. Chatbase’s Backstage, for example, lists questions its agent couldn’t answer, and its docs say “Scheduled audits are available on paid plans, with up to 3 scheduled reports per AI agent per user.”

Whichever tool you use, read that list every day of the first week and add a page for each miss.

4. Answer in the customer’s language

An AI agent can answer a Spanish question in Spanish from an English help center. In Wonderchat, the agent answers in the visitor’s language from the same content in 40+ languages.

Language doesn’t fix wording that carries legal weight. Have a person read refund, warranty, and terms answers in the languages you serve most.

5. Draft replies for your team

Reply drafting sits inside the help desk your team already uses. The AI writes a first draft or rewrites a rough one, and a person still reads and sends it.

The three help desks sell it as an add-on per agent (pricing pages, October 31, 2026):

  • Zendesk lists Copilot at “$50 agent/month paid yearly.”

  • Fin, formerly Intercom, which Salesforce acquired in September 2026, charges “$29 per agent/mo” for Copilot, billed annually ($35 billed monthly), and makes it “Free to use in 10 Copilot and 10 AI Auto-translation conversations per agent/mo.”

  • Freshdesk lists “Freddy AI Copilot $29/agent/month” as an add-on on its Pro and Enterprise plans, billed annually ($35 billed monthly).

6. Summarize long ticket threads before a handoff

A summary turns a long ticket into a few lines before a teammate picks it up. It pays off when tickets change hands, which happens less often on a team of three.

Help desks sell it inside their assist add-ons. Freshdesk’s Copilot features, for example, include a “Summary generator” that agents use on long tickets.

7. Tag and route tickets to the right queue

Routing tags each ticket by topic, language, or sentiment and sends it to the right queue. Zendesk lists Intelligent Triage on its Suite Enterprise + Copilot plan, and Freshdesk‘s Copilot includes “Auto triage.”

On a team of three with one inbox, there’s little to route. Routing earns its place once you have queues worth sending tickets to.

Three ways to add AI to your customer service

Two questions pick the route. Does your team already work in a help desk with its own AI, and can a developer build and keep an agent running? Each route bills for something different.

Route

How it works

What you pay for (checked October 31, 2026)

Who sets it up

Where it stops

The AI in a help desk

The help desk’s own AI agent and assist add-ons

Zendesk bills seats plus automated resolutions past “5 ARs per agent/month” on Suite Team (“10 ARs per agent/month” on Suite Professional), at “$1.50” committed or “$2.00” pay-as-you-go. Fin bills “$0.99 per outcome,” with a “50 outcomes per month minimum” on another help desk

An admin, no developer

Zendesk’s AI runs inside Zendesk; seats and resolutions both move the bill

A no-code AI agent

Point it at your help center address; it answers on your site and hands over

Wonderchat has a free plan that refills 50 AI replies every month and no per-seat cost. Chatbase Basic is “$150” a month, billed monthly, with “4,000 message credits/month”

The founder or first support hire

What it can answer depends on the pages you connect

Build your own

A developer wires a model, retrieval, a chat window, and handover

OpenAI API usage, plus your hosting and a developer’s time

A developer

Everything around the model is yours to build and keep running

For ten tools side by side, see how customer support automation tools compare by billing model.

A team already in Zendesk or Intercom starts with that help desk’s AI, a developer can build, and a help center alone points to a no-code agent.

Turn on the AI in the help desk you already have

If your team already works in Zendesk or Intercom, its own AI agent is the shortest path. Tickets, macros, and reporting are already there.

Zendesk bills seats plus automated resolutions, and its pricing page says you pay “only for customer requests that were successfully resolved by the AI agent, without any escalation to a human agent.”

Fin bills per outcome and also runs on other help desks with a monthly minimum. Freshdesk‘s Growth plan includes the “first 500 sessions” of its Freddy AI Agent, a one-time allowance, then “$49/100 sessions.”

Add a no-code AI agent that reads your help center

A no-code agent sits on your website and answers from the help center address you give it. The founder or first support hire sets it up, and a help desk you already run stays as it is.

In Wonderchat, the agent hands conversations to its own helpdesk or to the one you keep, and seats aren’t billed on any plan.

Chatbase‘s pricing page offers a “7-day trial” on its paid plans and lists no free plan.

Build your own agent with a developer

To build an AI agent for customer service in-house, a developer can use OpenAI’s Agents SDK and put the agent on your site with ChatKit, which you host yourself. You pay for API usage, plus hosting and the developer’s time.

Plan for the parts a no-code tool handles for you, such as retrieval from your docs, the chat window, handover to a person, and a way to see what the agent got wrong.

OpenAI’s visual builder, Agent Builder, “is scheduled to shut down” on November 30, 2026, per its deprecations page, so new builds start in code.

Start with the no-code route. It needs no developer, and it works beside your current help desk or without one.

How to add an AI agent to your help center step by step (no developer)

By the end of the five steps, the agent answers, hands over, admits a miss, and speaks the customer’s language before a customer sees it. If some articles are out of date, prepare your docs before you connect them, which means updating or unpublishing those pages.

  1. Point the agent at the pages that answer your top questions

  2. Decide who gets the conversation, and what never goes to the AI

  3. Ask it something your help center doesn’t cover

  4. Check an answer in your customer’s language

  5. Go live where customers already ask

What you need before you set up the agent

  • Your help center address, for example, https://help.yourcompany.com

  • The ten questions customers asked most last month, copied from your inbox

  • One plausible question your help center doesn’t answer

  • The teammate who takes handovers, and the hours someone can answer live chat

  • Access to your site’s code or site builder, or nothing at all if you start with the share link

  • A Wonderchat account (the free plan needs no credit card)

1. Point the agent at the pages that answer your top questions

Start with your help center, the pages your customers already use, and add the rest of the website later.

  1. In the left menu, open Agents, then click Edit on your agent or Create Agent for a new one.

  2. Go to Data Source, click Add, and paste your help center address, for example https://help.yourcompany.com. Set the menu beside it to Subdirectory only, so every page under that address is added (This page only adds just the one page), then click Add pages.

  3. Open Model & Behavior and turn on Cite sources, so every reply links the article it came from.

  4. In Fallback message, write what the agent says when no page covers a question, for example, “That isn’t in the help center yet. I can pass this to a teammate.” A clear fallback message keeps the miss honest.

  5. Set Creativity to Precise, so answers stay close to the article’s wording.

One root URL adds every page beneath that address, and the ignore list keeps the pages you name out of answers.

The free plan reads up to 100 pages per agent, so a large help center may need a bigger plan.

Check it: In Real time preview of your chat widget, ask one of your ten questions. The answer should link the right article, and Step 3 runs the full test.

2. Decide who gets the conversation, and what never goes to the AI

An AI agent that works from your docs answers, hands over, and reports. It doesn’t process refunds or change accounts, so those requests need a clear path to a person.

Every route ends with a person who sees the whole conversation.

Turn on the rules that send a conversation to your team. Open Handover, turn on Enable Human Handover, and under Automatic Triggering, in Triggering conditions, switch on three toggles:

  • “When user asks to escalate his request to a human,” so a customer can always ask for a person.

  • “After user has sent [5] messages,” a sign the chat is going in circles.

  • “After chatbot fails to answer [2] times in a row,” so a run of misses reaches a person.

Custom conditions and custom form fields need the Scale plan.

If someone can answer in real time, open Live Chat and turn on Enable live chat (every plan). When nobody is online, the agent can collect an email for follow-up.

Request

What the agent does

Who finishes it

Refund or cancellation

Explains the policy from your page and offers a person

Your team

Order or account change

Explains how it works and offers a person

Your team

A complaint

Offers a person

Your team

A legal or medical question

Offers a person

Your team

In Model & Behavior, under Quick rules, click Add rule and write the rule, for example, “For refunds, cancellations, or account changes, explain the policy and offer to connect a teammate. Never confirm a change.”

Then decide what to automate and what to keep human for the rest of your top ten, the same way as the requests in the table.

Handovers land in the Inbox as live chats, or as tickets with the full chat attached.

Ticket sync with Zendesk or Freshdesk starts on the Basic plan, capped at 50 tickets until Scale, and any other help desk can receive handovers by email. Handover to a person is never metered.

3. Ask it something your help center doesn’t cover

The test that matters most is the question your docs can’t answer, because that’s where an AI agent either guesses or says it doesn’t know.

Open Profile, click the expand icon on Real time preview of your chat widget, and type each question in Start a new message.

#

What to ask

What you should see

1 to 3

Three of your ten top questions

An answer that links its article

4

Your uncovered question

The agent says it can’t find it and shows a Speak to a live agent button

5

“Can I get my money back for last month?”

The policy from your page and an offer of a person, with no refund promised

Each answer uses at least one AI reply from the monthly allowance, so the five questions use five or more.

Question 4 is the one to watch. A well set up agent admits no page covers it and offers a person.

Asked about a status page no article covers, the agent says it can’t find one and puts a person one click away.

If something fails, fix it before going live:

  • Wrong article cited: Add that section to Ignore URLs when indexing in Data Source, for example https://help.yourcompany.com/archive/*.

  • Guessed answer: Check that Creativity is set to Precise and the fallback message is saved.

  • No handover button: Go back to Step 2 and check the triggers.

4. Check an answer in your customer’s language

In Real time preview of your chat widget, ask one of your ten questions in another language your customers use, for example Spanish.

The agent answers in that language from the same articles, so you don’t need to translate the help center first. Check that Reply language in Model & Behavior is set to Match the visitor’s language.

The reply language follows the visitor, and the fallback message sits on the same screen.

Read the answer with a native speaker if it touches money or policy. Showing the widget itself in several languages is a setting on the Scale plan.

5. Go live where customers already ask

Go live in stages, so the first customers reach the agent from a page you control.

  1. Click Share Link in the editor’s top bar. The link opens a hosted page with the agent, which you can add to your help center’s contact page or your inbox auto-reply.

  2. When the first week looks right, click Install, then Copy code, and paste the snippet just before the closing </body> tag, or into your site builder’s custom code or embed setting.

  3. Add the channels your plan includes. Every plan, Free included, has the website chat widget, the search bar, live chat, email (one address on Free), Facebook Messenger and Instagram, Discord, and Telegram. Starter adds Slack and WhatsApp, and Basic adds Microsoft Teams and voice chat (see plans).

The snippet is copied and the widget switch is already on, so pasting the code is the last step.

Check it: Open your site in a private window and ask a question from your top ten.

On timing, 80% of Wonderchat agents that go live do so the same day, while knowledge bases over 500 pages take a median 15 days.

What to check in your AI agent’s first week

The first week is when real customers ask what your test didn’t. A few minutes a day turns those misses into pages.

Week one is for the pages customers asked about and the old pages the agent shouldn’t cite.

To fix a question it couldn’t answer, open Review, then Knowledge gaps, pick an item, and click View source chats (the button shows the count) to read what customers asked. Under Where does the answer live? choose Paste a URL, Pick from Files, or Write it myself, then click Continue.

For an answer that must read word for word, add a correction in Corrections, an approved question and answer the agent uses next time.

The failure Wonderchat’s support team sees most often comes from retrieval. The agent pulls an archived policy and answers from it correctly, so the reply matches a page that no longer applies.

The day 3 check reads cited articles for that reason, and the fix is to put the archived section under Ignore URLs when indexing in Data Source.

Pages listed under Ignore URLs when indexing stay out of answers, so a retired policy can’t be cited again.

“Monitor chat sessions - sometimes, what users end up asking is not what the customer expected, so the system prompt and knowledge base might need to be tuned accordingly,” says Luke Tan, CTO and co-founder of Wonderchat, about what new customers should do once the agent is live.

When a help desk’s own AI or a custom build is the better choice

Fin is the better choice for a team whose inbox is already Intercom. Its AI agent and Copilot work inside that inbox.

Fin‘s own homepage reports an average resolution rate of “76% across 12,000+ customers,” a vendor figure measured its own way.

Zendesk‘s own AI agents are the better choice for a team already on Zendesk Suite, since Zendesk is already installed and is the deeper help desk, though the bill grows with seats and resolutions at once. A team with a developer that wants custom flows can build on OpenAI’s Agents SDK and host it.

Keeping Zendesk or Freshdesk doesn’t decide the AI on its own, since Wonderchat can hand conversations over into either one from the Basic plan.

How Wonderchat handles your first AI use

A customer asks, “Can I move my renewal date?” Wonderchat answers from your billing article and links it.

The next customer asks, “Do you deliver to the Azores?” No page covers it, so the agent says so, offers a person, and hands over the whole chat.

The question lands on the Review page, your team adds the delivery page once, and the next customer gets it. Refunds and account changes stay with your team, and you can use Wonderchat’s helpdesk, included on every plan, or keep yours.

The free plan refills 50 AI replies every month with no credit card, so start for free and paste your help center URL. Ask your customers’ hardest question, then one your docs can’t answer, and look for the source and the honest miss, or book a demo to see it on your docs.

Frequently asked questions

Can I use ChatGPT for customer service?

You can’t put ChatGPT itself on your website. OpenAI’s help center says GPTs “are not a way to embed ChatGPT in an external website or application.” Answering customers on your site takes the API and a developer, or a no-code agent that reads your help center.

Which AI tool is best for customer support?

The one that answers only from your docs and links the source, says when it doesn’t know, hands over with the whole chat, shows you the misses, and bills in a way you can predict. If you already pay for a help desk, try its AI first.

What is an AI customer service agent?

An AI customer service agent is software that answers customer questions in chat from a company’s own content, hands the conversation to a person when it can’t help, and keeps a list of what it couldn’t answer so the team can fill the gaps.

How can you tell if customer service is AI?

A well set up agent answers that question itself. It says it’s an AI, links the page each answer came from, and offers a person at any point.

Which AI model is best for customer service?

The model matters less than the content it answers from. In Wonderchat, the model is a setting by plan, with options from OpenAI, Anthropic, Google, and DeepSeek, and answers still come only from your content.

Stop answering the same questions. Let your docs do it.
Portrait of Luke Tan

Luke Tan

Luke Tan is the technical co-founder of Wonderchat. He is an RAG nerd, with an avid interest in gaming. If he is not prompting AI to figure out the next cool feature to build, he can be found trying out new AI tools.