Yes, for repeating factual questions with answers you've approved, and to help your team reply faster. No, for complaints, emotional situations and anything involving exceptions or money, unless a person stays involved. Always offer an easy way to reach a human, and remember that you remain responsible for what it says.
When does it make sense?
| Situation | AI alone | AI helps a person | A person alone |
|---|---|---|---|
| "Where's my order?" | Yes, from your system | ||
| Opening hours, returns policy | Yes, from approved text | ||
| Product question with a known answer | Often | Yes | |
| Refund or exception | No | Yes, it prepares | |
| Complaint | No | Yes, it summarises | Often |
| Upset or vulnerable customer | No | Yes | |
| Legal or safety matter | No | Yes |
Our advice is to begin in the middle column. AI that drafts for your team carries little risk, and it shows you quickly where the answers are good. Only then let it answer the first rows by itself.
Who's responsible when it's wrong?
You should assume you are. The best-known example is Moffatt v. Air Canada. A chatbot on the airline's website gave wrong information about bereavement fares, and a British Columbia tribunal held the airline liable for negligent misrepresentation. It rejected the argument that the chatbot was a separate entity. The chatbot was treated as part of the website. It's a small Canadian case and not EU law, but the logic is worth remembering: your bot speaks for you.
So limit what it can say. It should answer from your approved sources and say so when it doesn't know. It shouldn't make up policy or promise exceptions.
Do you have to tell people it's AI?
In the EU, this is now part of the rules. Article 50 of the AI Act requires providers to design AI systems that interact directly with people so that those people are told they're dealing with AI, unless that's obvious. The Commission summarises it as humans being made aware they're interacting with a machine, and says the transparency rules come into effect in August 2026. Check with your supplier or adviser how this applies to your setup. Beyond the law, we'd just say it: people mind being misled more than they mind a bot.
What makes a good hand-off?
- A person is one clear step away, always, without having to argue with the bot
- The conversation goes with it, so nobody has to repeat their story
- It hands off by itself on complaints, repeat questions or low confidence
- Outside office hours it says when someone will respond, and then that happens
The warning sign is a setup that measures success by how few people reach your team. That number can look great while customers give up. Track unresolved questions and repeat contacts as well.
How do you test before going live?
- Collect real questions from your inbox, including strange ones.
- Write the right answers next to them, with your team.
- Run them through and review each answer. Wrong is worse than "I don't know".
- Try to break it: ask for discounts, off-topic advice, other customers' details.
- Launch on one channel or topic, and read the conversations each week.
- Run the same test after every change to content or model.
What should you build?
Your helpdesk may already offer this, so look there first. If visitors only need a route to you, a form may do. Custom work makes sense when answers have to come from your own systems or the AI has to take actions. Then you need an agent with approval moments, like our Business Agent, from €7,950, excluding VAT.
