Keeping your voice warm with a customer who is frustrated and already raising theirs.
Sounding patient while you explain the same policy for the sixth time today.
Remaining firm with someone who has missed three payments and has a new excuse ready.
This is emotional labour – the unbilled work of managing other people's feelings while regulating your own. In service roles, it's a known part of the job description. It's also quietly draining the people doing it, with real costs to psychological wellbeing and, eventually, to the business itself.
Knowing how much this work takes out of people, the question then becomes: is it time to rethink who does it? Or, more specifically, what does it?
That's what Yuqian Chang, Ivey Assistant Professor of Marketing, and her colleagues tested in their new research, Artificial Intelligence, Emotional Labor, and Service Operations. Partnering with a loan collection agency, the researchers ran two large-scale field experiments involving more than 3,000 customers who had fallen behind on payments. Through outbound collection calls, they compared the performance of human employees with that of a hyper-realistic AI voice bot.
What they found challenges a common assumption about AI in business: that machines are built for logic, not the nuance of emotion. It turns out, they may be surprisingly good at both.
Faking a feeling is hard work. For humans, anyway.
In their first study, Chang and her co-researchers provided both the AI voice bot and human employees with the same training and emotional scripts before handling calls with 720 customers who had missed loan payments. They were instructed to use one of three tones: warm, firm, or neutral.
When a call required only a neutral tone, humans and AI performed about the same. But when the job called for a specific emotional response, the performance gap opened quickly.
Asked to sound warm and encouraging, human agents hit the right tone just 22 per cent of the time. Asked to sound firm and serious, they managed 19 per cent. The AI did it with 100 per cent accuracy, every time. And while human performance declined sharply within the first three hours of work, with most agents defaulting back to a neutral tone, the AI held steady, call after call.
It's a staggering – and almost alarming – gap. But Chang is quick to push back on the obvious conclusion: that AI is simply better than humans at expressing emotions.
"Sustaining a performed emotion, on demand, for hours, is genuinely hard," she said. "AI agents worked better here because they excel at consistency, at being told what to do and executing it without any frustration or fatigue.”
But accuracy isn't everything
The first study showed that AI could consistently perform the emotional cues it was given. But does that translate into better business results? To find out, Chang and her colleagues then examined whether expressing these emotions accurately actually helped the business in collecting overdue payments.
Here's where the story turns.
When the AI was instructed to sound warm with customers who had only recently fallen behind – and conversely firm with those who had missed payments repeatedly – it significantly outperformed human agents – collecting 49 to 94 per cent more than them.
But when the instructions were flipped – warmth for repeat offenders, firmness for a first-time slip – the AI underperformed humans. It executed the wrong emotion perfectly, and customers responded accordingly.
The finding points to an important distinction for leaders: AI will amplify whatever you tell it to do, with near-perfect accuracy. But accuracy doesn't matter if the emotion doesn't fit the situation.
As Chang explained, "A human might instinctively soften their tone or change their approach when something feels off in a call. AI doesn't do that – and that's where things can go wrong."
A smarter way to share the emotional load
While the research demonstrates AI's potential to take on emotional labour at scale, Chang and her co-authors see the bigger opportunity in rethinking how the work is divided between people and machines. For leaders looking to put these insights to work, Chang recommends a few practical moves:
- Map your pain points. Identify where employees repeatedly manage their emotions, and where emotional delivery affects outcomes or contributes to burnout. These are the areas where AI may ease the burden and free employees to focus on work requiring creativity, discretion, and sound judgment.
- Set clear emotional guidelines. AI follows instructions with precision. That's an asset when the instructions are right and a liability when they're not. Develop an evidence-based guide for matching the right emotion to the right scenario, and regularly review and refine it.
- Treat AI and humans as a tag team. Deploying AI as a blanket solution for emotional labour leaves money on the table. The research found that strategically assigning AI to some customer profiles and humans to others lifted repayment rates from 17.5 per cent (AI only) to 20.53 per cent.
- Monitor and tweak continuously. Treat emotional output as a quality metric, not a set-and-forget feature. Managers should keep retraining the model against operational data to keep the AI aligned with the business.
Underneath these tactics, however, is a bigger shift in mindset. The debate is no longer whether AI can perform emotional labour. It can – and, in the right conditions, better than humans. The real question now is whether leaders can deploy it thoughtfully enough to capture the upside without scaling the downside.
For more on the research behind these findings, read Chang’s paper, Artificial Intelligence, Emotional Labor, and Service Operations, co-authored with Zheng Fang of the Business School at Sichuan University, Xueming Luo of the Fox Business School at Temple University, Qingsheng Wu of the Antai College of Economics and Management at Shanghai Jiao Tong University, and Jaakko Aspara of the School of Business at Aalto University in Finland.