AI chatbots for customer service have come a long way from the rigid menu bots of a few years ago. Modern assistants can understand free-form questions, look up answers in your help center and respond naturally at any hour. They can also get things badly wrong, confidently, in front of your customers.
Updated September 2026: we added details of the Air Canada chatbot ruling, the EU disclosure rule and the research behind retrieval-augmented generation.
Here is what separates the helpful implementations from the ones that backfire.
What AI handles well
- Routine questions with clear, documented answers: opening hours, shipping times, return policies
- Order status and account lookups, when connected securely to your systems
- Triage, gathering details and routing to the right team
- After-hours coverage for simple requests
- Drafting replies for human agents to review and send
Where it backfires
- Inventing policies. A chatbot that makes up a refund rule creates a real problem. In a widely reported February 2024 case, a British Columbia tribunal held Air Canada responsible for incorrect bereavement-fare information its website chatbot gave a customer. It rejected the airline’s argument that the chatbot was a separate legal entity and found the company responsible for all the information on its website, whether it came from a static page or a chatbot.
- Trapping customers. Nothing frustrates people more than a bot with no way to reach a human.
- Handling emotional situations. Complaints, bereavements and billing disputes need empathy and judgment.
- Pretending to be human. Customers should know they are talking to an automated assistant, and in some places, such as the EU, disclosure is a legal requirement: under the AI Act, people must be told when they are interacting with an AI system, a duty that has applied since August 2026.
Ground it in your own content
The single most important design decision is to have the bot answer from your approved content, such as your help center, policies and product information, rather than from the model’s general knowledge. This technique is often called retrieval-augmented generation, from a 2020 research paper. It reduces made-up answers and keeps responses consistent with what your staff would say.
Keep that content current. A bot is only as accurate as the help articles behind it.
Design the hand-off first
- Make “talk to a person” available at any point.
- Pass the conversation history to the human agent so customers do not repeat themselves.
- Automatically escalate complaints, legal questions and signs of frustration.
- Be honest about response times when no human is available.
A useful rule If a wrong answer would cost you money, a customer or your reputation, the bot should hand off rather than guess.
Measure the right things
Deflection rate, meaning conversations that never reach a human, is tempting to optimize but misleading on its own. Also track:
- Customer satisfaction after bot conversations
- Resolution rate: did the issue actually get solved?
- Repeat contacts about the same issue
- A sample of transcripts reviewed by staff each week
Start small
Launch on a narrow set of well-documented topics, review transcripts closely and expand gradually. Many teams get more value initially from AI that drafts replies for human agents than from a fully autonomous bot. For broader advice on getting started, see our small business AI guide.
Build the foundations first
AI customer service works when it answers from your approved content, admits what it does not know and hands off gracefully. Build those three things first and the chatbot becomes an asset rather than a liability.
Sources
- BC Tribunal Confirms Companies Remain Liable for Information Provided by AI Chatbot, American Bar Association, February 2024
- Regulation (EU) 2024/1689 (Artificial Intelligence Act), EUR-Lex
- Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks, arXiv, 2020
