Most support inboxes hold the same few questions asked a dozen ways, a smaller pile of requests that need someone to look something up or change something, and a handful of conversations where the customer is upset, worried or asking for an exception. An AI customer service agent can take a real share of the first pile, help with the second, and should mostly stay out of the third.
The hard part is rarely the model. It’s deciding which enquiries go where, and designing the moment a conversation passes to a person so the customer doesn’t have to start again. Get that moment wrong and customer service automation costs more goodwill than it saves. If agents are new to you, what AI agents are and where to start covers the basics first.
The short answer
An AI customer service agent should answer what your published information covers, help with account tasks once it knows who it’s talking to, and pass anything needing judgement or empathy to a person.
- Automate: questions your website answers, such as opening hours, delivery areas or what a service includes.
- Automate with guardrails: order status, booking changes and other reversible tasks, once identity is verified and the customer confirms.
- Keep human: complaints, refunds outside policy, disputes, payments, and anything involving health, safety, legal threats or a distressed customer.
- Always: let customers reach a person whenever they ask, and pass the conversation along so they never repeat themselves.
Sort your enquiries into three groups
Before choosing a tool, export a few hundred recent enquiries from your inbox, help desk, chat and forms, and tag each by what the customer needed, not how they phrased it.
Group 1: answerable from published information
Prices, opening hours, service areas, what’s included, lead times, how returns work, which documents to bring.
Should an agent handle it? Yes, and it’s the right first job, provided the agent answers only from approved content, links its source and admits when the content runs out.
The catch. Many “published” answers turn out to live in someone’s head, an old PDF or a saved reply. Close those gaps on the website first; preparing an AI knowledge base shows how to structure content so an agent finds the right answer.
Group 2: needs account data or an action
“Where’s my order?” “Can I move my appointment to Thursday?” The answer depends on who is asking, and sometimes something has to change.
Should an agent handle it? Often, in stages: read-only look-ups first, then well-defined actions that are easy to undo, such as rescheduling into an open slot. Each needs identity checked before personal data is shown (a one-time code to the email or phone on file, say), access limited to that task, the customer’s explicit “yes” before anything changes, and a log of every action. Payments and anything irreversible stay with a person. Connecting an agent to your website, CRM and booking system covers the integration side.
Group 3: needs human judgement or empathy
Complaints, exceptions to policy, disputes, bereavement, health worries, legal threats, safety issues, customers who are clearly struggling, and high-value sales conversations.
Should an agent handle it? No. Its job is to recognise these fast, gather the basics and hand over. People here want someone with authority to act, and an agent improvising on policy leaves the business answerable for what it said. In Moffatt v. Air Canada (2024), a Canadian tribunal held the airline liable after its website chatbot gave a grieving customer wrong information about bereavement fares. Limiting AI agent risks covers the controls.
Watch for conversations that change group
“What’s your returns policy?” is group 1 until the next message says the product arrived broken, for the second time. The agent should reassess on every message, and treat a change of tone or topic as a reason to hand over.
| At a glance | Published information | Account data or action | Judgement or empathy |
|---|---|---|---|
| Examples | Hours, prices, policies | Order status, rescheduling | Complaints, exceptions, distress |
| Agent’s role | Answers, with a source link | Looks up, acts after confirmation | Recognises, gathers basics, hands over |
| Needs | Current content in every language | Identity check, limited access, action log | Clear triggers, fast route to a person |
Where an AI customer service agent earns its keep
Three jobs tend to deliver the most value for the least risk.
Out-of-hours cover
Many enquiries arrive when nobody is working, including from time zones where your night is their office hours. An agent can answer group 1 questions immediately and turn everything else into a complete ticket for the morning.
- Be honest about availability. State your real hours, how the reply will arrive and a reply time your team can meet. Never imply someone is watching when nobody is.
- Keep an urgent route. For problems that can’t wait, such as a burst pipe in a managed building or a medical concern before a procedure, the agent should give the emergency or on-call number rather than log a ticket.
Triage and routing
AI can help even where people write every reply: classify each email, form or chat by topic, urgency and language, pull out the order number, spot repeat contacts and route it with a draft reply to check.
Support ticket triage is a low-risk place to start: with a person still sending every reply, a wrong label costs minutes rather than a customer. Track how often tickets are reassigned, and send anything the model is unsure about to a general queue. Sales enquiries need different rules, covered in AI lead qualification.
Answering in the languages your customers use
Current models handle many languages, though quality varies, especially in less widely used ones. Multilingual customer support works when:
- the agent replies in the customer’s language, detected from what they write, not their location or browser
- it answers from reviewed content in that language, not unchecked on-the-fly translation, above all for prices, policies and anything contractual; the multilingual SEO guide shows how to structure each language version
- the handover crosses languages cleanly, with a translated summary beside the original for a colleague who doesn’t speak it, and the customer told which language the reply will be in.
Test with real questions from your inbox in every language, not only the project team’s.
Designing the handover to a person
A good human handoff separates automation customers tolerate from automation they resent. Three things decide it: when it happens, what travels with it and what the customer is told.
When to hand over
Write the triggers down and test each one:
- The customer asks for a person, in any wording. Hand over at once, never loop.
- The topic is on your human-only list.
- Approved content doesn’t cover the question.
- The agent has misunderstood twice, or the customer keeps repeating themselves.
- The customer is frustrated or distressed, or mentions safety, health or legal action.
- The next step needs authority the agent lacks, such as an exception or goodwill gesture.
- Identity can’t be verified for a request that needs it.
Err towards handing over early: an agent tuned to hold on to conversations flatters the numbers while customers give up.
Carry the conversation across
So the customer never has to repeat anything, the handover should pass a structured package, not just a link to a chat log. Keep the conversation in the same channel where you can; if a chat has to become an email, quote the relevant part in the first human reply so the customer can see it was read. Test every channel you offer: context that survives in web chat can still get lost when a conversation starts in a messaging app.
What the person receiving it needs to see
Design the receiving screen as carefully as the chat window: someone picking up a handover should grasp the situation in under a minute.
| Field | Why it matters |
|---|---|
| Customer, language, channel, verification | Who they are and how to reply |
| One-line summary of the request | The situation without scrolling |
| Reason for handover | Whether to fix, reassure or decide |
| Answers given, with sources | Exactly what the customer has been told |
| Actions already taken | Nothing done, or undone, twice |
| Urgency and mood | Urgent or upset cases jump the queue |
| Transcript and customer history | Detail on demand |
Add a quick way to flag an agent answer as wrong, so the weekly review has something to work from.
Tell the customer what happens next
Say plainly who will pick it up, where and roughly when:
- Someone is available: “I’ve passed this to our bookings team with everything you’ve told me. Someone will reply here shortly.”
- Nobody is available: “Our team is back tomorrow at 9:00. I’ve logged your request, and they’ll reply by email.”
If nobody accepts a live handover within a set time, switch to the second message rather than leave the customer watching a typing indicator. How this looks in the interface, including AI disclosure, is covered in designing AI agent experiences people trust.
Improving the agent after launch
Customers will ask things nobody predicted, so plan the review routine before launch.
Read transcripts and escalations every week
Book a fixed slot to review:
- every handover, grouped by reason
- every “I don’t know” and every poorly rated answer
- customers who came back after the agent marked their question resolved
- a random sample of conversations it closed alone, where silent failures hide.
Fix the cause, not the symptom
- Content gap. The answer isn’t published. Add it to the website, not only to the agent’s instructions, so site, agent and team say the same thing.
- Retrieval problem. The answer exists but was missed, often because it’s buried in a long page or PDF. Restructure it.
- Scope problem. The agent attempted too much or handed over too soon. Adjust its instructions and triggers.
- Integration problem. A look-up failed or returned stale data. Fix the connected system.
Measure what customers experience
Track handover reasons, repeat contacts, ratings and the wait between handover and human reply. Distrust “deflection” (the share of conversations that never reach a person) as a headline figure: a customer who gives up counts as deflected too. If reviews keep pointing to one narrow, repetitive task, chatbot, workflow or AI agent explains why a simpler workflow may suit it better.
Frequently asked questions
Can an AI customer service agent replace a support team?
It shouldn’t be planned that way. It can absorb repetitive questions and out-of-hours cover, but people still handle handovers, review transcripts and keep the content it answers from accurate.
Should customers be told they’re talking to AI?
Yes, in the first message, with the route to a person kept visible. It sets fair expectations, and some laws expect it: the EU’s AI Act includes transparency obligations so that people know when they are interacting with an AI system, unless that’s already obvious.
What’s the safest way to start with AI customer support?
Launch on one channel, answering only questions your website already covers, with a visible route to a person and every conversation read for the first few weeks. Add account look-ups once answers are consistently right, and actions after that.
Start with the content it will answer from
An AI customer service agent can only be as accurate as the pages behind it. Before comparing platforms, check the content that answers what customers ask most:
Fix the gaps before any assistant starts quoting your pages. For an outside view of the foundations, a free website audit looks at how your pages perform on phones, whether they can be found and how easily visitors reach a person, the route every handover relies on. If answers are scattered across years of old pages and attachments, a website redesign can rebuild the content around the questions customers ask. Or talk to us about what your website needs.