Key takeaways
- You’re responsible for what your chatbot says. A Canadian tribunal made Air Canada honour a refund policy its chatbot got wrong.
- Under the EU AI Act’s timetable, Article 50 applies from 2 August 2026 and requires people to be told they’re talking to an AI unless that’s obvious.
- Write your escalation triggers down before launch: sensitive topics, money and refunds, repeated failure, low confidence and any request for a person.
- Measure quality from real transcripts (accuracy, handoff rate, repeat contacts), not from deflection alone.
Why the handoff matters more than the bot
Most chatbot failures aren’t wrong answers. They’re customers who can’t get out. The US Consumer Financial Protection Bureau described this in its June 2023 report on bank chatbots, warning about “doom loops” that are “often caused when a customer’s issue falls outside the chatbot’s limited capabilities.” The same report lists “wasted time, feeling stuck and frustrated” among the harms, and estimates that about 37% of the US population interacted with a bank’s chatbot in 2022 (a figure the CFPB attributes to Insider Intelligence).
There’s a legal angle too. In Moffatt v. Air Canada (February 2024), the airline’s website chatbot told a grieving customer he could claim a bereavement fare after travelling. The policy said otherwise. Air Canada argued, in effect, that the chatbot was responsible for its own actions. The British Columbia Civil Resolution Tribunal called that a “remarkable submission” and said: “It should be obvious to Air Canada that it is responsible for all the information on its website.” The customer was awarded $812.02. That’s a small sum, but the principle carries: treat every bot answer as your company speaking.
Tell people it’s an AI, at the start
Article 50(1) of the EU AI Act says systems that interact directly with people must be built so those people “are informed that they are interacting with an AI system,” unless that’s obvious from context. Article 50(5) adds that this has to happen “in a clear and distinguishable manner at the latest at the time of the first interaction.” Under the Act’s timetable, Article 50 applies from 2 August 2026. Our EU AI Act and Digital Omnibus guide covers how the wider timetable is shifting.
In practice that means a plain first message such as “Hi, I’m an AI assistant for Acme. I can help with orders and returns, or connect you with our team.” Don’t give the bot a human name and photo that suggest otherwise. In our view, this is good practice even if you don’t sell into the EU, because customers who feel tricked tend to stop trusting the answers. This isn’t legal advice; check your own obligations with a lawyer.
Ground every answer in your own knowledge base
The Air Canada bot’s answer contradicted a policy page it linked to. Grounding is the fix. Retrieve the answer from approved sources (help articles, policy pages, product data) and tell the model to answer only from them. If nothing relevant comes back, it should say so and offer a person rather than improvise.
- Give each policy one owner and one current version, and retire old PDFs so the bot can’t find them.
- Show a link to the source article under answers about refunds, prices or eligibility.
- Keep anything the bot must never promise (discounts, legal or medical advice, exceptions) on an explicit block list.
- Re-test the bot whenever a policy changes, not just at launch.
Escalation triggers to set before launch
Decide in writing when the bot stops and a human takes over. Here’s the starting set we’d use, adjusted to your business:
| Trigger | Example | What the bot does |
|---|---|---|
| Customer asks for a person | “Can I speak to someone?” | Hands off straight away, no second attempt to persuade |
| Repeated failure | Two rephrased questions with no useful answer | Offers a human and passes the transcript |
| Money, refunds or contracts | Disputed charge, cancellation terms | Answers from policy, routes any exception to staff |
| Sensitive or vulnerable situations | Bereavement, health, complaints, distress | Responds with care and escalates with priority |
| Low retrieval confidence | No matching article found | Says it doesn’t know and offers a person |
| Anger or legal language | “This is unacceptable”, “my lawyer” | Escalates to a senior team member |
A good handoff carries context. Pass the conversation, the customer’s details and the bot’s guess at the issue, so nobody has to repeat themselves. Out of hours, say honestly when someone will reply and capture a contact method. Don’t pretend a human is typing.
How to measure whether it’s working
Deflection rate on its own rewards the doom loop the CFPB warned about. A bot that never hands off scores well and serves people badly. Track these instead:
- Answer accuracy: review a weekly sample of transcripts against the source policy.
- Handoff rate and reasons: rising handoffs on one topic usually point to a gap in the knowledge base.
- Repeat contact: did the same customer come back within a few days about the same issue?
- Time to a human: how long it takes from a handoff request to a real reply.
- Customer feedback after the chat, read alongside the transcript rather than as a score alone.
Launch on a narrow scope first, say order tracking and returns, and widen it only once those numbers hold up.
Where TendWise fits
We design support assistants with disclosure, grounding and handoff built in from the first sprint. See our AI chatbot development service, or our notes on WhatsApp chatbot policy if that’s your main channel. When the bot needs to take actions as well as answer, agents versus workflow automation is the next read. Or book a strategy call to map your escalation rules together.
Frequently asked questions
Do I have to tell customers my chatbot is AI?
Can my business be liable for a chatbot’s wrong answer?
When should a chatbot hand off to a human?
What’s the best metric for chatbot quality?
Sources
- EU Artificial Intelligence Act: Article 50, Transparency obligations
- BC Civil Resolution Tribunal: Moffatt v. Air Canada, 2024 BCCRT 149
- Consumer Financial Protection Bureau: Chatbots in consumer finance