An AI agent in customer service is not the mistake. The mistake is letting it repeat answers without ever giving the customer a way out. This week I lived through a perfect example of how NOT to use AI, and I want to share it so you don't repeat it in your company, academy or training centre.
What happened to me: six replies and no solution
I needed a copy of a credit agreement from a finance company. In the app and the customer portal the download failed and, through the chat, I was told that for contracts before a certain date I had to request it by email from customer service. So far, so reasonable.
I wrote the email... and an AI agent replied. Over almost three hours I exchanged six messages with it. The result:
- It explained how to download the contract from the app, even though I had already said I had tried and it didn't work.
- It suggested looking for the contract in the welcome email, even though I had explained it is an old purchase and I no longer have it.
- It told me that, because the purchase was old, the option wasn't available in the app and that I should reply to the same email... in other words, to the very agent that was answering me.
- It contradicted itself: first it said the problem was the purchase date, then that it was a specific fault with that contract.
- Every message ended with the same courtesy line, with no human contact and no deadline offered.
Nobody said "we've logged your request, you'll get it within 48 hours". Just loops. Wasted time for me and a poor brand experience for the company.
The 7 mistakes to avoid with customer service AI
1. Not offering a clear route to a human
A stuck customer should be able to reach a person in one step, not five. If the agent hasn't solved the problem by the second reply, it should escalate automatically.
2. Letting the AI reply without reading what has already been said
If the customer says "I already tried that", repeating the same step is disrespectful. The agent needs access to the case history.
3. Sending the customer back to the same AI
A channel that redirects to another channel that redirects back to the first is a circle, not a service.
4. Not connecting the AI to real systems
An agent that only "knows" the FAQ can't resolve specific cases. If it can't look up the file or start the process, all it can do is give directions.
5. Giving contradictory answers
Without an up-to-date, reviewed knowledge base, the AI improvises. And contradictions destroy trust faster than an "I don't know".
6. Not saying it's an AI or what its limits are
Transparency is not just good practice: in the European Union, AI regulation is moving towards requiring that people know when they are talking to an automated system. It's worth having an adviser check this for your specific case.
7. Measuring cost savings instead of resolution
If the only metric is "tickets handled", the AI looks like a success even when the customer gets nothing. Measure first-contact resolution and time to solution.
How to do it well: quick checklist
- Automatic escalation: after 2 unresolved replies, or if the customer writes "this doesn't help", it goes to a person.
- Case memory: the agent reads the whole thread before replying.
- Real action: it can open a ticket with a tracking number and a deadline.
- Transparency: it identifies itself as AI from the first message.
- Weekly human review of failed conversations to improve the agent.
A practical example for an academy or SME
Imagine a language academy that automates enrolment enquiries with an AI agent. A good design: the agent answers timetables and prices; if the student asks for a certificate or a group change, it creates a task in the system (for example in Odoo), tells the student that someone from the office will handle it and gives a deadline. With a simple automation (Make or Zapier) the alert reaches the team without anyone having to copy it over.
Frequently asked questions
Is using AI in customer service a bad idea?
No. It works very well for repetitive enquiries. It fails when it is used as a barrier instead of a first filter.
When should a case go to a person?
When the customer repeats their request, when there is a specific technical error, when money or legal documents are involved, or when the customer asks for it.
How do I start doing this well in my team?
First define what the AI can and cannot resolve, then train your team to supervise it. That is the job of responsible AI coaching.
Want to use AI without losing customers?
At Meta Academy AI we help companies, academies and training centres roll out AI with judgement: what to automate, what to leave to people and how to train the team. Start with a free preliminary study and also read our AI usage policy for companies.
Caroline, director of Meta Academy AI and Meta Language Academy.