AI for customer support: chatbots, triage and handover to humans

Customer support is one of the most common places businesses put AI to work, and one of the most visible when it goes wrong. A support chatbot speaks for your business. If it promises a refund your policy does not allow, the customer will expect you to honour it.

This guide covers the three main ways AI helps a support team, how to keep a bot grounded in your own policies, when and how to hand over to a person, and what to measure. It is written for small teams using helpdesk software with AI features, or building a simple bot on a general AI model.

Three jobs AI can do in support

1. Self-service answers: a chatbot on your website or in your helpdesk answers common questions directly: delivery times, returns, opening hours, order status.
2. Triage: AI reads incoming emails and messages, tags the topic, judges urgency and routes each one to the right queue or person.
3. Agent assist: AI drafts replies, summarises long threads and suggests help articles, and a person sends the final answer.

Agent assist carries the least risk because a person checks every reply before it goes out. Self-service carries the most, because the customer sees the output directly. Many small teams start with triage and agent assist, then add a chatbot for a narrow set of questions.

Anthropic’s customer support use-case guide suggests AI fits best where there is a high volume of repetitive queries, a need for round-the-clock availability, sharp peaks in demand, or a need for a consistent brand voice.

The Air Canada lesson

In Moffatt v Air Canada (2024), decided by British Columbia’s Civil Resolution Tribunal, a customer asked the airline’s website chatbot about bereavement fares. The bot told him he could apply for the discount after travelling. Another page on the same website said otherwise, and the airline refused the claim. According to a summary by the law firm McCarthy Tétrault, Air Canada suggested the chatbot was responsible for its own actions. The tribunal rejected that, finding the airline had not taken reasonable care to make sure the bot was accurate, and stated: “It should be obvious to Air Canada that it is responsible for all the information on its website.”

That was a Canadian decision, but the principle is easy to apply anywhere: customers will treat what your bot says as what your business says.

Ground the bot in your own policies

A general AI model knows nothing about your returns window or delivery partners. Anthropic’s guide notes that the model needs the relevant information in its context, which usually means connecting it to your help centre or policy documents, a method known as retrieval-augmented generation. For live, customer-specific data such as order status or account details, the guide says embedding-based retrieval is not enough and recommends tool use, meaning the bot looks the information up in your order system.

The guide’s example guardrails are a good starting point. They include instructions not to speculate about future plans and not to “make promises or enter into agreements it’s not authorized to make.” It also recommends checking that responses align with company policy and stripping out personal data unless it is needed.

Anthropic’s guidance on reducing hallucinations adds two simple instructions worth including: answer only from the provided documents, and say so when the answer is not there.

Tell customers they are talking to AI

Being upfront avoids confusion and complaints. If you serve customers in the EU, it is also a legal point. Article 50(1) of the EU AI Act requires providers of AI systems that interact directly with people to design them so that people “are informed that they are interacting with an AI system”, unless that is obvious. The European Commission’s AI Act page says the transparency rules apply from August 2026. The UK has no equivalent AI-specific rule, but a clear label such as “AI assistant” is good practice anywhere.

Design the handover to a person

A good handover is the difference between a helpful bot and a trap. Decide the triggers in advance:

• the customer asks for a person, in any wording;
• the bot fails to resolve the question after two attempts;
• the topic is a complaint, a refund above a set amount, a legal threat, a safety issue or a data request;
• the customer mentions illness, bereavement or other signs they may need extra care;
• the bot cannot find the answer in its sources.

When the handover happens, pass a short summary and the conversation history to the agent so the customer does not have to repeat themselves. Tell the customer honestly what happens next and when, especially outside opening hours.

Measure what matters

Anthropic’s guide lists criteria at two levels. Task-level measures include response accuracy, topic adherence and escalation efficiency. Business measures include deflection rate (queries resolved without a person), customer satisfaction and average handle time.

Watch deflection rate with care. A bot that makes it hard to reach a person will score well on deflection and badly on satisfaction. Track them together, and read a sample of conversations each week, especially those that ended in a handover or a low rating.

A worked example

An example (illustration only): an online clothing shop with three support staff gets about 1,200 messages a month. It sorts the common intents and decides how each is handled.

The bot is labelled as an AI assistant, offers a “talk to a person” button on every screen, and passes a summary to the helpdesk on handover. The team reviews 20 conversations a week and updates the help pages when the bot cannot find an answer.

Common mistakes

• Launching a bot on everything at once instead of a few well-documented topics.
• Letting the bot answer from general knowledge instead of your policies.
• Hiding the route to a person.
• Out-of-date help pages, which the bot will repeat faithfully.
• Measuring deflection alone.
• Letting the bot promise refunds, discounts or exceptions.