AI in customer service works best as a coach for human agents and worst as a cheap substitute for them. Peer reviewed field studies show that assistants raise agent output, especially for newer staff, while customers still rate bots below people, punish disclosure and react badly when an angry complaint meets a cheerful humanlike bot.
Three assumptions sit underneath most decisions about putting software on the front line of service. The first is that a bot which gives the right answer is as good as a person who gives the right answer, so customers will not care who is on the other end. The second is that making a bot friendlier and more humanlike will make customers like it more. The third is that the gains from assisting agents will be shared evenly, or will flow mostly to the best people, and that the work left for humans will become lighter.
Each of these assumptions is partly wrong, and the ways in which they are wrong matter to anyone who designs service jobs, hires agents or answers to a board about service costs. The research in this article comes from field experiments inside working contact centers, laboratory experiments with thousands of consumers and several meta analyses, which are studies that pool the results of many earlier studies to find the overall pattern. Where the evidence is thin, comes from a single company or has not yet completed peer review, the article says so plainly.
Where Does the Popular Story About AI in Customer Service Go Wrong?
The popular story treats service automation as a cost line. A bot answers the easy questions, the number of calls reaching people falls, and the savings drop to the bottom line. In that story the customer is neutral. Customers want answers, the bot supplies answers, and nobody minds.
The trouble is that this story mixes up two different things. One is whether the software can do the task. The other is how people react when they know software is doing it. The research shows that these two questions have different answers, and that the second one often decides whether a deployment pays off.
There is also a gap in what has been measured. Much of the research on customer acceptance of chatbots consists of surveys that ask people how likely they are to use a bot. One 2023 meta analysis pooled 70 such studies covering 20,026 people and found that intention to use a service chatbot rose with how useful and how playful people found it, with their general attitude, with the support available to them and with the views of people around them. Culture and the type of task changed some of these relationships. That is useful, but a stated intention is not a behavior. People who say they would happily use a bot may still abandon one that misunderstands them.
Even the most widely cited thinking in this field began as prediction rather than evidence. A 2018 conceptual paper on service robots argued that machines would take over the simple and analytical parts of service while people would keep the emotionally and socially demanding parts. That forecast has turned out to be largely right, but with consequences for the remaining human jobs that few people anticipated. The sections that follow trace what the field and laboratory evidence now shows, starting with the agents themselves.
Do Agents Actually Work Better With an Artificial Intelligence Assistant?
The strongest single piece of evidence on this question comes from a contact center that introduced a conversational assistant built on a large language model, which is software trained on vast amounts of text to predict and produce humanlike writing. The assistant watched each customer chat as it happened and suggested replies and links to relevant technical help that the agent could use or ignore. Because the tool reached different groups of agents at different times, the researchers could compare agents before and after access with agents who had not yet received it.
The published study, which involved 5,172 customer support agents and was co written by the economist Erik Brynjolfsson, found that access raised the number of customer issues resolved per hour by 15% on average. That average hides the most important finding. Less experienced and less skilled agents improved both their speed and the quality of their work. The most experienced and most skilled agents became only slightly faster, and the quality of their work fell slightly. The gains were largest for problems that were moderately rare, the kind an individual agent had seen only a few times but that the system had seen often enough to learn from. The researchers also found signs that agents learned from the tool, and that international agents improved their written English.
The authors interpret this pattern as the tool spreading the habits of the most capable agents to everyone else. Put simply, the assistant had absorbed what good agents tend to say in difficult moments, and it passed that knowledge to people who had not yet built it up through experience.
A single company is a narrow base for broad conclusions, so a second study matters. A 2026 field experiment at the after sales chat operation of a large Chinese online retail platform randomly gave some agents access to a generative assistant, meaning a tool that writes new text, which drafted a diagnosis of the customer's problem and a proposed solution at the start of each chat. Agents could accept, edit or ignore the draft. The assistant made service faster and raised the ratings customers gave, although it did not change the objective measure of quality the researchers used, which was whether customers had to come back with the same problem. Agents with the tool answered faster and took a more active role, and customers had to explain less. This study is a working paper that had not completed peer review at the time of writing, so its findings should be treated as promising rather than settled.
Taken together, the two studies agree on the headline. When the software supports a human rather than replacing one, the human handles more customers in the same time, and customers notice the difference. Neither study supports the claim that the tool improves everyone.
Why Do the Best Agents Gain the Least?
The finding that top performers gain little, or even lose ground, appears in both field studies, and it deserves attention because it runs against intuition. Many managers expect a powerful tool to make their best people even better.
In the retail platform experiment, the lower performing agents gained the most, which narrowed the gap between the weakest and strongest members of the team. The top performers, by contrast, saw declines in both the ratings customers gave and the objective quality of their service. The researchers traced the problem to disruption of how these agents worked. When the strongest agents used the assistant, they spent more time away from their shifts, responded more slowly and saw more customers come straight back with the same issue. The tool broke the rhythm of people who already had a good rhythm.
The explanation is not mysterious. A tool that offers the typical good answer helps someone who does not yet know the good answer. For someone who already knows a better answer, the suggestion is a distraction that has to be read, judged and rejected. Every one of those small judgments costs time and attention.
This has two consequences for people managers. The first is that the case for these tools is strongest in teams with high turnover and many new hires, which describes a large share of contact centers. The second is less comfortable. If the tool works by spreading the habits of top performers, then those performers are supplying the knowledge that makes their colleagues faster, while their own measured output stalls. A pay or performance system that ranks agents only on calls handled will, after deployment, make the best agents look less exceptional than they are. That is a recipe for losing exactly the people whose conversations taught the tool what good service sounds like.
Why Do Customers Judge a Bot More Harshly Than a Person?
Turning to the customer side, the evidence on how people react when they know they are dealing with a machine is remarkably consistent, and mostly negative.
The clearest demonstration comes from a field experiment with more than 6,200 customers who were randomly assigned to receive a structured sales call from either a chatbot or a human worker. When customers were not told they were speaking to a bot, the bot sold about as well as proficient human workers and roughly four times as well as inexperienced ones. When customers were told at the start of the call that they were speaking to a bot, purchases fell by roughly 80%. Customers who knew they were talking to a machine were curter, ended calls sooner and rated the bot as less knowledgeable and less caring, even though it was the same software. Telling customers later in the conversation, and customers having prior experience with such systems, softened the penalty. This study concerned outbound sales calls rather than inbound service, but the mechanism it uncovered applies to both.
Why should a label matter so much? Laboratory and field research published in 2023 found that consumers rated service from bots lower than identical service from people, and that the main reason was a belief that firms automate to cut costs at the expense of customers. In other words, the customer reads the bot as a signal about the firm's priorities. That reading is not always wrong.
Empathy is the other part of the explanation. Three vignette experiments with 714 participants, in which people read a short written scenario and reacted to it, found lower satisfaction, lower intention to return and lower willingness to recommend after a chatbot interaction than after an identical human one. This held whether the outcome of the service was good or bad, and the gap was fully explained by how much empathy people felt the service provider showed. Making the bot's language more empathetic, though not making it look more human, closed the gap with the human agent.
The bot also changes how customers see the people who work alongside it. An automated sentiment analysis of consumer reactions in fashion and telecommunications retail found that overall sentiment toward bots was actually less negative than sentiment toward online human agents. The more striking finding was that sentiment toward the human agents became more negative after a retailer introduced a chatbot. Customers who had been through a bot first arrived at a person with lower patience and higher expectations. This is one study, and its method relies on automated sentiment scoring, but the pattern fits everything else in this section.
Should a Service Bot Look and Sound Human?
Firms often give their bots a human persona with a friendly label, a cartoon face and a chatty style, on the theory that humanlike bots will be liked more. Researchers call this anthropomorphism, which simply means giving a machine human features or treating it as if it had a mind. The evidence on whether it helps is mixed, and the mixture is informative.
The pooled evidence is mixed but informative. A 2021 meta analysis combined 108 samples covering 11,053 people who interacted with service robots and chatbots. It found that the link between how humanlike customers perceived a robot to be and their willingness to use it ran mainly through practical judgments, such as whether the robot seemed intelligent and useful, rather than through feelings of rapport. The effect also depended on the type of service. A later meta analysis of 82 samples from 42 articles found that humanlike chatbots improved customer outcomes from first contact through to purchase, but did not reduce customers' negative attitudes, and that the effect depended on whether the service ended well or badly. The most recent synthesis, published in 2026 and covering about 34,300 respondents, found that usefulness, ease of use, trust and satisfaction consistently drove acceptance, while humanlike features made bots more enjoyable without necessarily making them more trusted.
Small humanlike touches can change behavior in useful ways. An online experiment with 153 participants found that customers were more likely to agree to a chatbot's request for service feedback when the bot used humanlike language, and also when it first asked for a small commitment before the larger request, a sales tactic known as the foot in the door technique. The humanlike language worked because it made the bot feel more like a social presence.
The danger appears when things go wrong. A five study investigation, including analysis of a large real world data set from an international telecommunications company and four experiments, found that when customers arrived angry, a humanlike chatbot lowered their satisfaction, their overall view of the firm and their intention to buy again. Customers who were not angry showed no such penalty. The researchers traced the effect to inflated expectations: a humanlike bot promises human competence, and when it fails to deliver, angry customers feel the gap more sharply.
Emotional expression follows the same logic. Three laboratory experiments found that positive emotion expressed by a bot, such as enthusiasm or warmth in its replies, did less to improve service evaluations than the same expression from a human employee. Customers do not expect warmth from a machine, so warmth from a machine can feel odd rather than reassuring. How much this matters depends on whether a customer sees the service relationship as transactional or personal.
Service failure sharpens the point further. A 2025 hotel experiment compared failures by a human employee, a humanoid robot and a machinelike robot. The more humanlike the provider, the less customers forgave the failure, the more they expected in the way of recovery and the more dissatisfied they became. This held for failures in how the service was delivered, though not for failures in the end result.
The practical reading is that humanlike design is a bet on things going well. It helps a little when the service is routine and successful. It backfires when the customer is already upset or when the bot fails, which are precisely the moments when service matters most.
When Do Customers Actually Prefer the Machine?
Customers do not always prefer a person. In some situations the absence of a human is the point.
Five experiments in 2022 found that when people bought embarrassing products, such as treatment for a sexually transmitted infection or hemorrhoid cream, they preferred to deal with a service robot because they felt less judged by it. In a field test of an online health advertisement, people clicked more often when the ad featured a robot than when it featured a human, by 1.26% compared with 0.74%. The advantage disappeared for ordinary purchases, where people preferred a human, and it shrank when the robot looked very human, because a humanlike robot brought back the feeling of being judged.
A similar pattern holds for bad news. Experiments on unwelcome offers found that customers responded more favorably, with higher purchase likelihood and satisfaction, when a worse than expected offer came from an artificial intelligence agent rather than a person. Better than expected offers worked better from a person. The explanation is that people see machines as having weaker intentions, so they blame a machine less for bad news and credit it less for good news. The researchers themselves flagged the ethical risk of firms using bots to slip unpopular terms past customers.
This finding sits uneasily beside the empathy experiments described earlier, which found that chatbots were rated lower after both good and bad outcomes. The two sets of studies used different scenarios and measured slightly different things, so the honest conclusion is that the effect of outcome direction is not yet settled. What both agree on is that the identity of the messenger changes how a message lands.
Customers also soften toward bots when they can see what is in it for them. The cost cutting study described earlier found that the penalty for bot service disappeared when firms passed the savings to customers through lower prices, and reversed when the bot delivered clearly better service than human employees. The bot penalty is not fixed. It reflects a judgment about whose interests the bot serves.
Finally, blending humans and bots can close much of the gap. Three further studies with 643 participants found that customer trust in chatbot service builds in layers, from a general trust in service technology, to learned trust in chatbots as a category, to trust in the particular bot in front of them. A configuration that combined chatbot service with a light human touch matched pure human service on most customer outcomes.
What Happens to the People Who Still Answer the Phone?
The most neglected question in this field concerns the agents who remain. Here the evidence points in two directions at once.
The good news comes from the contact center study described at the start. Once agents had the assistant, customers were more polite to them and less likely to ask to speak to a manager, which in contact centers is called escalation. The same study, in its earlier working paper version, reported that agents with access to the assistant were more likely to stay with the company. In a sector where staff often leave within months, that matters as much as the productivity gain. The likely reason is simple: newer agents who are helped through difficult conversations have fewer bad days.
The bad news comes from what happens when bots take over the easy work entirely. A 2026 randomized experiment at the same large online retail platform tested a more autonomous system, sometimes called agentic software because it acts on its own rather than suggesting actions to a person. The software resolved the chats it was judged able to handle, while human agents supervised it and dealt with everything else. Customer ratings fell for the chats the software handled, even though those chats became shorter. When the software passed a chat to a human because the problem was technically beyond it, the human preserved service quality. When it passed a chat over because the customer was frustrated or angry, the human intervention worked less well. In those emotional handovers, agents sent fewer messages, took a smaller share of the conversation and were less active in seeking information and offering solutions. Stepping in early made a large difference. On the positive side, agents gave more attention to the chats the software could not handle at all. This study, like the other platform study, had not completed peer review at the time of writing.
Put these findings next to decades of research on emotional labor, which is the effort of displaying emotions a job requires, such as patience with a rude caller. A 2011 meta analysis of three decades of research found that faking emotions one does not feel, which researchers call surface acting, is consistently linked with emotional exhaustion and poorer wellbeing, while genuinely trying to feel the expected emotion carries much less cost.
The implication is uncomfortable. When software absorbs the routine, pleasant and quick interactions, the work left for people is concentrated in the angry, confused and emotionally loaded ones. Those customers have often already failed with a bot, and as the earlier sentiment analysis suggested, they arrive less patient than before. An agent who once had a mix of easy and hard calls may now face hard calls all day. The 2018 forecast that people would keep the emotional work was correct. What it did not spell out is that a job made entirely of emotional work is a different and more draining job.
What Does AI in Customer Service Mean for You?
If you lead a service operation or the human resources function that supports it, the evidence gives you a clearer starting point than most vendor promises do.
Your newest agents stand to gain the most from a well designed assistant. If your contact center hires constantly and loses people within months, an assistant that coaches agents through difficult conversations is likely to raise output, improve how customers treat your staff and may help you keep people longer. Your strongest agents may gain little and may even slow down, so you should let them switch the suggestions off and you should not judge them against a target that the tool has raised for everyone else.
Your customers will judge a bot by what they believe it says about you. If they think it exists to save you money at their expense, they will mark you down even when the bot performs well. If you hide the bot, you may win in the short term, but you are building service on a deception that the evidence shows customers resent once discovered.
Your angry customers need a person, and they need one early. A cheerful humanlike bot is the worst thing to put in front of someone who is already upset. Route anger and complaints to people quickly, and do not dress up the bot as a human to delay that moment.
Your remaining agents are doing harder work than before. If software takes the easy contacts, you have changed the job, and you need to change the job description, the selection criteria, the training, the recovery time between difficult contacts and the pay to match.
Key Takeaways
- Assistants that suggest replies to human agents raise output, with the largest gains going to newer and less skilled staff, while top performers gain little and sometimes lose quality.
- Customers penalize service they know comes from a bot, largely because they read automation as a cost cutting move and see bots as less caring.
- Humanlike design helps a little when service goes smoothly but backfires with angry customers and after failures, because it raises expectations the bot cannot meet.
- Customers prefer a machine in a few specific situations, such as embarrassing purchases, and they soften toward bots when the savings or a clearly better service reach them directly.
- Handing easy work to software leaves human agents with a heavier share of emotionally demanding customers, which research on emotional labor links to exhaustion.
- Several of the newest findings come from working papers at a single large platform, so the direction is clear but the size of the effects should be treated with caution.
Implications for Practice
Start with assistance rather than replacement. The strongest peer reviewed field evidence supports tools that sit beside an agent and suggest, not tools that take the conversation away. Pilot the assistant with a randomly chosen group of agents and compare them with a group who do not yet have it, so that you know whether the gains in your operation resemble those in the research.
Target new hires first, and redesign onboarding around the assistant. Because the gains are largest for inexperienced agents and for moderately rare problems, the tool can shorten the time it takes a new agent to become competent. At the same time, give experienced agents the option to turn the suggestions off, and track their quality as well as their speed so that a drop is caught early.
Revise how you measure and reward performance. If the assistant narrows the gap between strong and weak agents, a ranking based only on volume will undervalue your best people. Recognize the agents whose conversations represent the standard the tool spreads, for example by involving them in reviewing the suggestions and rewarding that contribution.
Be open about bots, and make the benefit visible. Disclosure carries a penalty, but the research on cost cutting shows that the penalty shrinks when customers can see what they gain, whether that is faster resolution, availability outside office hours or lower prices. Tell customers early that they are dealing with software and what it can do for them, and offer a clear route to a person.
Detect anger and hand it to people quickly. Use plain language rather than a humanlike persona for complaint and recovery channels. When a customer is frustrated, transfer the chat to a human early rather than after the bot has failed, because the field evidence shows that late emotional handovers get weaker human effort.
Treat the remaining agent role as a new job. Update the job description to reflect a workload weighted toward difficult and emotional contacts. Select for emotional resilience and problem solving, train agents in handling distress without relying on faked cheerfulness, build recovery time into schedules and review pay so that it reflects the harder mix of work. Monitor exhaustion and quitting among agents in the months after any deployment, not just handling time and customer ratings.
