What customers say about AI support
Gartner surveyed 3,566 B2B and B2C customers in February and March 2026. Of them, 87% say a company that uses generative AI for customer service must provide access to a human agent. (Gartner, August 2026).
Read it correctly: customers are not against AI. Gartner reports that people reluctant to use AI would most often change their mind if they could switch to a human. The problem is AI as a wall, not AI as a first line. Gartner’s Eric Keller advises against making GenAI a mandatory first step for every issue.
When AI should hand over
Write handoff rules before launch instead of leaving them to the model. Our suggested starting set:
- The customer asks. “Human”, “agent”, “operator” or a “Talk to a person” button passes the conversation at once, with no follow-up question.
- Sensitive topics. Complaints, payment disputes, contract cancellations, legal questions, damaged goods. AI can collect the facts; a person answers.
- Low confidence. If the answer is not in the knowledge base, admitting it beats guessing.
- A stuck conversation. The customer rephrases the same question a third time or writes angrily: hand over.
- High-risk actions. Refunds, changing a delivery address after dispatch, correcting an invoice.
Adapt the list to your process. The principle matters most: the way to a person is not the last step after three failed attempts.
What to pass along
Repeating themselves is what annoys customers most. When the conversation reaches a team member, they should see:
- who the customer is and which order or contract it concerns;
- a short summary and the customer’s last question;
- what the AI already tried and anything it promised;
- urgency and the reply deadline.
If this stays in a chat window the team cannot open, the handoff is only a label on a button.
What to do outside working hours
A person is not always available. Say so honestly and create a ticket with a deadline: “We will reply by 12:00 tomorrow.” Customers tolerate waiting better than an endless AI loop. The ticket must not vanish into a shared inbox; it needs an owner and a deadline the system tracks.
Say that it is AI
A person talking to an AI system should know it is not a person. That is the EU AI Act, Article 50, for systems that interact directly with people, unless it is obvious from the context (AI Act, Article 50). Even where it does not strictly apply to you, “I am an AI assistant and can pass you to a colleague” is a better opening than pretending.
What to measure after launch
- how many conversations go to a person, and for which reasons;
- time from handoff to the first human reply;
- how many customers write again about the same issue;
- how many handed-over conversations were about something the AI should have known (this shows what to add to the knowledge base).
A high handoff count is not a bug to drive to zero. If it is zero, customers may simply not find the exit. For where a chatbot pays off and where it does not, see AI chatbot for a business website.
How it looks in practice
In our ticket system FDesk, every e-mail becomes a ticket. AI sets the topic and urgency and drafts a reply, but only a team member sends it, and SLA tracking shows what is overdue. The person stays in the process by design. On the phone, DeskVoice answers calls, books clients and hands the call to a person when needed; data stays in the EU. More on the phone case: AI phone assistant instead of a call centre.
Conclusion
If your support gets many repeated, simple questions, run AI as the first line, but write the handoff rules first, give the team full context and set a deadline for out-of-hours tickets. If most questions are complaints or individual calculations, start with AI that only sorts and drafts replies. We map your process and check where AI helps and where it only gets in the way: discuss your project.
