Anyone who has watched a livestream freeze mid-concert or lost a multiplayer game to a server crash knows the frustration that follows. What happens next, whether a company can smooth things over or loses the customer entirely, has always depended heavily on the person handling the complaint. But as companies race to hand off customer service to generative AI systems like ChatGPT-powered chatbots, a new question has emerged: when a service goes wrong, does a customer actually want to talk to a machine?
A set of studies published in Psychology & Marketing suggests the answer depends heavily on how badly things have gone. For minor hiccups, generative AI can leave customers feeling better than a human representative would. But when the first attempt at fixing the problem also fails, the calculus flips, and only specific interventions can bring the AI back into contention.
The question behind the research
Hesam Olya of the University of Sheffield and colleagues from the University of Birmingham and Texas A&M University set out to examine an area that has received little empirical attention. Earlier research on AI in customer service focused on simpler, rule-based chatbots, the kind that give you a menu of clickable options and often leave you shouting “representative” into the void. Those systems tend to lose to human agents in head-to-head comparisons because they feel mechanical and impersonal.
Generative AI is a different animal. Powered by large language models, these systems can hold conversational exchanges, recognize emotional cues in a customer’s message, and adapt their responses on the fly. The research team wanted to know whether this newer breed of AI changes the traditional wisdom that humans are better at handling complaints, particularly in emotionally charged situations.
They focused on live-streaming services, an industry projected to hit $256 billion by 2032. Live events like concerts and multiplayer gaming sessions carry high emotional stakes: customers are paying for a real-time hedonic experience, and when it breaks, they notice immediately.
How the researchers built their case
The team used a layered approach. They started with qualitative interviews to understand how consumers think about AI versus human service agents. They then ran a lab experiment using facial recognition software (FaceReader) to capture participants’ spontaneous emotional reactions when interacting with different types of agents, providing an objective check on self-reported feelings.
The main evidence came from four online experiments with a total of more than 1,200 U.S. participants recruited through the Prolific platform. Each experiment used video vignettes showing a service disruption during a livestreamed video game session or concert, followed by a recovery attempt from either a basic chatbot, a generative AI chatbot, or a human agent. Participants then reported how likely they were to buy the service again and how they felt emotionally.
The researchers built their theoretical case around cognitive appraisal theory, which holds that emotions arise not directly from events but from how people evaluate those events. A key concept they measured was “promotion emotion,” a category of high-energy positive feelings like excitement, delight, and cheerfulness. This is distinct from mere satisfaction, which tends to be a calmer, more settled feeling. Promotion emotion, the authors argue, is what livestream customers are really after, since these services exist to deliver pleasure rather than just utility.
What the experiments revealed
In the first experiment, when a single service failure was cleanly resolved, generative AI recovery outperformed both the basic chatbot and the human agent. Participants who interacted with the generative AI reported higher repurchase intentions (an average of 5.04 on a 7-point scale) than those who dealt with a human (4.53) or a basic chatbot (3.95). The generative AI also produced the strongest promotion emotion, and that emotional lift was what drove the difference in behavior. The researchers ruled out frustration and helplessness as alternative explanations.
Then came the twist. In the second experiment, the team introduced what the field calls a “double deviation,” meaning a failure followed by a failed recovery attempt. The customer reports the problem, tries the fix, and it doesn’t work. Roughly 68% of customers report experiencing this kind of compounded failure in real life.
Here the advantage flipped. After a double deviation, human agents produced higher repurchase intentions (4.63) than generative AI (4.03). The AI’s earlier edge evaporated once things got complicated. The researchers interpret this as a signal that customers begin to suspect the company is prioritizing cost-cutting over their interests when a firm keeps sending an AI back after the first attempt has failed.
Two ways to rescue the AI
The final two experiments tested whether specific interventions could bring generative AI back into competition after a double deviation. The first tested empathy. When the recovery agent (human or AI) displayed low empathy, sticking to cold, fact-based responses, humans still outperformed the AI. But when the agent expressed high empathy, acknowledging the customer’s disappointment, apologizing repeatedly, and using warm language, the generative AI closed the gap. Repurchase intentions for the AI (4.81) became statistically indistinguishable from those for the human (4.66).
The second intervention was monetary compensation. In this experiment, participants had paid $20 for a live concert stream that got interrupted twice. When no refund was offered, the human agent produced better outcomes than the AI, echoing the earlier double-deviation results. But when a $10 refund was offered, the pattern reversed. Generative AI now produced higher repurchase intentions (5.47) than the human agent (4.80). The authors interpret this as compensation neutralizing the suspicion that using AI represents corner-cutting, allowing the AI’s speed and efficiency advantages to come through.
What this means for companies deploying AI
The findings point toward a tiered approach for firms integrating generative AI into customer service. For straightforward failures like a temporary lag or a minor bug, generative AI can serve as what the authors describe as “a first line of defense,” resolving issues instantly and leaving customers in a better emotional state than a human agent might.
When issues escalate into double deviations, companies face a choice. They can hand off to a human agent, who will likely be perceived as more sincere and accountable in these emotionally loaded moments. Or they can continue with generative AI, but only if they equip it to signal genuine empathy or pair its response with tangible compensation like a refund, bonus credits, or virtual rewards.
The researchers also note some limits to their findings. The experiments used scenario-based videos rather than real-time interactions, and some of the statistical effects were modest in size. The study focused on live-streaming contexts, so the results may not transfer directly to higher-stakes domains like medical or legal services, where financial compensation cannot easily undo harm. Participants were also not screened for familiarity with livestreaming, which could affect how realistic the scenarios felt for some. The authors suggest field experiments and cross-cultural replications as next steps to test how well the patterns hold up in the wild.
