- An AI customer service agent answers questions and completes routine requests through your systems, such as order tracking, booking changes and address updates.
- Your business is responsible for what it tells customers, whether the answer comes from a web page or an AI agent.
- Hand off to a person when the customer asks, when the agent is unsure or stuck, for complaints and vulnerable customers, and for decisions with a significant effect on the customer.
- A good handoff passes the full conversation and a summary, so the customer never repeats themselves.
- Tell customers they are talking to an AI. In California, a bot that misleads people about being a bot to influence what they buy breaks the law unless it discloses that it is a bot.
- Measure resolution the customer confirms, handoff reasons, repeat contacts and a weekly review of real conversations.
What should an AI agent handle in customer service?
Start with requests that are frequent, follow a clear policy and can be answered or completed from your own systems:
- Status questions: where an order, parcel, booking or repair is, read live from your order or tracking system.
- Simple changes: rescheduling a booking or delivery slot, updating an address or contact details, cancelling within policy.
- Logging requests: opening a return, fault report or maintenance request with the details your team needs, and confirming the next step.
- Policy and product questions: answered from your approved help content, with retrieval so the agent quotes your policy instead of guessing.
- Triage for your team: classifying incoming emails and tickets, collecting missing details, and routing them to the right queue with a summary.
The difference between a chatbot and an agent is that the agent can take actions through tools, not only answer. If you are still deciding which kind you need, read AI agent use cases and where they fail and how to build an AI chatbot.
Are you responsible for what your AI agent tells customers?
Yes. A Canadian tribunal held Air Canada responsible for wrong information from its website chatbot, treating the chatbot as part of the company’s own website.
Yes. In Moffatt v. Air Canada, 2024 BCCRT 149, British Columbia’s Civil Resolution Tribunal rejected Air Canada’s argument that its chatbot was effectively a separate legal entity responsible for its own actions. The tribunal said the chatbot was still just part of Air Canada’s website, that Air Canada was responsible for all the information on its website whether it came from a static page or a chatbot, and that it had not taken reasonable care to make sure its chatbot was accurate.
When must an AI agent hand off to a human?
When the customer asks, when the agent is unsure or stuck, for complaints and vulnerable customers, for out-of-policy requests, and for decisions with a significant effect on the customer.
- The customer asks for a person. Make this possible at any point, in plain words, without a maze.
- The agent is unsure, cannot find the answer in your content, or has failed to resolve the request after a couple of attempts.
- Complaints, strong frustration, legal threats, or anything that may need an apology and judgement.
- Signs of a vulnerable customer or a safety issue.
- Requests outside policy and exceptions, which need a person’s authority.
- Decisions that have a legal or similarly significant effect on the customer, such as closing an account or refusing a service.
That last point has a legal basis in the UK. Under Article 22A of the UK GDPR, a decision is “based solely on automated processing” when there is no meaningful human involvement, and it is a significant decision when it has a legal or similarly significant effect. Article 22C requires safeguards for such decisions, including letting the person make representations, obtain human intervention and contest the decision. These articles were substituted by the Data (Use and Access) Act 2025.
How do you design a good AI-to-human handoff?
Pass the full conversation and a summary to your team, keep the customer in the same channel, and tell them what happens next. In practice:
- Pass the whole conversation, the customer’s identity, the details collected and a short summary of what was tried, into your helpdesk or CRM.
- Keep the customer in the same channel. Moving them from chat to a phone queue makes them start again.
- Tell the customer what happens next: a person joining now, or a reply by a stated time if the team is offline.
- Out of hours, create a ticket with everything collected instead of leaving the customer at a dead end.
- Let your agents see what the AI said and correct it, and feed recurring corrections back into the help content.
The most common failure is the handoff, not the AI. If customers have to repeat themselves to a person, they will judge the whole service by that moment.
Which rules apply to AI customer service in the UK, US and Canada?
Your business is responsible for what the agent says, California prohibits bots that mislead people about being bots, and UK GDPR requires safeguards, including human intervention, for significant automated decisions.
- Canada: the Air Canada decision is a small claims ruling, but it shows how a tribunal treats a company’s chatbot: as part of the company’s own website.
- United States: California’s Business and Professions Code section 17941 makes it unlawful to use a bot online with intent to mislead someone in California about its artificial identity in order to influence what they buy or how they vote; a bot that discloses it is a bot is not liable. The disclosure must be clear and conspicuous.
- United Kingdom: conversations contain personal data, so UK GDPR applies, including the automated decision-making safeguards above when the agent makes significant decisions on its own.
- Voice and WhatsApp: AI voices on outbound calls fall under US calling rules, covered in the AI receptionist guide. WhatsApp has its own rules for AI and handoff, covered in the WhatsApp AI agent guide, along with Gulf data laws.
Whatever the law requires, telling customers they are talking to an AI and how to reach a person is good practice everywhere. None of this is legal advice; confirm your position with your adviser.
Should your customer service AI agent use voice or chat?
Start where your customers already contact you and where requests are most repetitive.
Chat and email are easier to start with: the agent has time to look things up, customers can read links and confirmations, and transcripts are easy to review. Voice agents suit businesses whose customers mostly call, but they need faster responses, careful confirmation of names and numbers, and a smooth transfer to a person. Many businesses run both on the same back end, so the agent gives the same answers in every channel.
How do you measure an AI customer service agent?
Measure confirmed resolution, handoff rate and reasons, repeat contacts, customer satisfaction, and accuracy in a weekly review of real conversations.
- Confirmed resolution: conversations where the customer confirms the issue is solved, not just conversations that ended.
- Handoff rate and reasons: why the agent handed off, so you can fix gaps in content or tools.
- Repeat contacts: customers who come back about the same issue within a few days, a sign the first answer did not work.
- Customer satisfaction: compared between AI-handled and person-handled conversations.
- Accuracy review: a weekly sample of real conversations checked by a team lead against your policies.
How do you roll out an AI agent for customer service?
Automate two or three frequent request types first, test on real past conversations, launch in one channel with daily review, then expand. Step by step:
- Pull a month of tickets, chats or call notes and group them by reason. Pick two or three frequent, policy-driven reasons to automate first.
- Clean up the help content those answers depend on, and name an owner who keeps it current.
- Connect the systems the agent needs through narrow tools, such as reading order status or changing a booking, with read-only access where possible.
- Write the handoff rules and connect the handoff to your helpdesk or CRM.
- Test with real past conversations, including angry, confused and off-topic customers, before going live.
- Launch in one channel, or to a share of conversations, with daily review at first. Expand when confirmed resolution is steady and handoffs work.
Related reading
For phone calls and bookings, read AI receptionist for clinics and small businesses; for qualifying sales enquiries, AI agents for real estate leads; for messaging, WhatsApp AI agents for business. For grounding answers in your own content, see what is RAG. The tracking and chat systems mentioned above are in the Shipping Hack, Pacifrica Express and ServiPR case studies. Sigi’s AI agent development, RAG assistant and AI integration services cover this work.

