80% of customer service will be automated. And 64% of customers would rather it weren't.
The two predictions come from the same field of research and seem to contradict each other. They don't: taken together they say what will separate those who use these systems well from those who use them to save money.
The prediction says that by 2029 AI agents will resolve on their own 80% of common customer service issues, with a 30% reduction in operational costs (Gartner, press release of 5 March 2025). The word «common» carries the whole sentence and is cut from almost every quotation.
The same field of research produces a second prediction that circulates much less: 64% of customers would prefer companies not to use AI in customer service. Both are true and they talk about different things: the first about what the technology will be able to do, the second about how people react to the way it is used today.
The number that decides what to do isn't the prediction: it is the last hundred enquiries you received, sorted into three groups (answer always the same, needs looking up in a system, needs judgement). The proportion between the three is the only figure that concerns your company.
The most important word is «common»
The prediction is almost always quoted cut short, and the cut always falls in the same place. It doesn't say 80% of customer service will be automated: it says 80% of common requests will be. It is a precise category, and every company already knows which ones are its own: where are you, what time do you open, where is my order, does this product contain gluten, can the booking be moved.
These are the questions someone today answers for the thousandth time. They require no judgement, no negotiation, and no knowledge of the customer. They are also, almost always, the numerical majority of incoming volume and a clear minority of its value.
Rephrased like this, the prediction becomes much less spectacular and much more useful: in a few years it will be unusual to keep a person waiting for a question whose answer is always the same. More than a revolution, it is a shift of threshold, as happened with opening hours on the website or being able to book without phoning.
The other half of the data
The same body of research that produces that prediction produces another, which circulates much less because it sells much worse: 64% of customers would prefer companies not to use AI in customer service. And also: half of the companies that will have cut staff counting on AI expect to rehire by 2027.
It pays to hold them together, because both are true and they aren't about the same thing. The first describes what the technology will be able to do. The second describes how people react to the way it is used today. The 64% has no opinion on agent architecture: it has the memory of the last time it tried to reach a human being and couldn't.
The rejection isn't about automation, but about automation used as a wall instead of a door.
The difference, in practice, lies in three choices
They are three design decisions, all taken before building, and all free:
- The way out to a person is always visible. Not hidden after four steps, not conditional on phrasing the question the right way. A system that makes it hard to talk to someone isn't saving money: it is shifting the cost onto the customer, and customers all notice.
- The system states what it is. No made-up name, no fake hesitation to seem human. Transparency here isn't just a matter of style: the AI Act moves in the same direction, and a system built honestly today doesn't have to be redone tomorrow.
- What it doesn't know, it doesn't invent. An assistant that answers only by drawing on the company's documents, and that hands over when faced with a question outside its scope instead of improvising, makes far fewer mistakes. The mechanism is called RAG, and its main function isn't to seem more intelligent, but to know when to stay quiet.
This is a prediction, not a measurement: it describes a 2029 scenario formulated in 2025, and it is a kind of statement that should be treated for what it is. Predictions on adoption rates have a documented history of optimism, and this one also mainly concerns organisations much larger than an Italian SME.
It has no value as a promise. Nobody can tell you that 80% of your enquiries can be automated, for the elementary reason that they don't know which enquiries you receive. It has value as an indication of direction, and as an invitation to do something you can do straight away: count, on the last hundred enquiries you received, how many had the same answer.
The count that replaces the prediction
Industry predictions serve to decide whether to look at a topic. They don't serve to decide what to do, because they describe an average of companies that aren't yours. The number that decides what to do, in this case, is obtained in an afternoon.
Take the last hundred enquiries that arrived, from any channel, and sort them into three groups: those you gave an answer that is always the same, those that required looking something up in a system, and those that required judgement. The proportion between the three groups is your figure, and it is the only one that counts.
If the first group is large, you have a solvable problem and you know its size. If the third group is large, the honest conclusion is that right now an automated customer service system isn't your first intervention, and whoever proposes it to you is proposing it without having counted.
Questions and answers
Will 80% of customer service really be automated by 2029?
The prediction says something more precise: by 2029 AI agents will autonomously resolve 80% of common customer service issues, with a 30% reduction in operational costs (Gartner, press release of 5 March 2025). The word «common» carries the whole sentence and is cut from almost every quotation.
Common requests are a precise category every company recognises: where are you, what time do you open, where is my order, does this product contain gluten, can the booking be moved. They are the numerical majority of incoming volume and a clear minority of its value.
Is it true that customers don't want AI in customer service?
64% of customers would prefer companies not to use AI in customer service, and half of the companies that will have cut staff counting on AI expect to rehire by 2027.
The rejection, however, isn't about automation itself. The 64% has no opinion on agent architecture: it has the memory of the last time it tried to reach a human being and couldn't. It is about automation used as a wall instead of a door.
How do you build an automated assistant that customers accept?
Three design decisions, all taken before building and all free. The way out to a person is always visible, not hidden after four steps. The system states what it is, with no made-up names or fake hesitations to seem human, which is also the direction the AI Act is moving in. What it doesn't know, it doesn't invent: it answers by drawing on the company's documents and hands over when faced with a question outside its scope.
How can I tell how many of my enquiries can be automated?
By counting them, because the prediction describes an average of companies that aren't yours. Take the last hundred enquiries that arrived, from any channel, and sort them into three groups: those you gave an answer that is always the same, those that required looking something up in a system, those that required judgement.
If the first group is large, you have a solvable problem and you know its size. If the third is large, the honest conclusion is that an automated customer service system isn't your first intervention.
Sources
- Gartner, Gartner Predicts Agentic AI Will Autonomously Resolve 80% of Common Customer Service Issues Without Human Intervention by 2029, press release of 5 March 2025. The source of: 80% of common requests and the 30% reduction in operational costs.
- From the Gartner customer service surveys cited in the coverage following the prediction: the 64% of customers who would prefer no AI in customer service, and the share of companies expecting to rehire staff by 2027.
Count your enquiries before automating them.
The Diagnostic asks this too: what happens to an enquiry arriving outside working hours, how many times a week you rewrite the same information. Twenty questions, five minutes, no spam.