Customer service agents who chose to consult frequent AI-generated feedback improved a performance measure in a workplace experiment, while adding a comparison with colleagues reduced the gain, researchers report.
The finding comes from a study in the International Journal of Research in Marketing of support staff at a US software company. The research was published online in September 2026 and is a corrected proof awaiting its final issue.
It offers a practical question for employers considering artificial intelligence: does a feedback tool help staff improve, and do they actually use it?
A faster way to offer feedback
Customer service managers often rely on surveys and coaching to understand how conversations went. Surveys can take time to collect and may reflect the views of people who are especially happy or unhappy, rather than every customer.
Researchers Hisham Abdulhalim, Byungyeon Kim, Elie Ofek, Adi Shalev and Talia Tron developed a model to predict a customer satisfaction rating from a call transcript and features of the conversation. These included speaking patterns, interruptions and call length.
The prediction is an estimate, not a direct statement from every customer. An AI system can be wrong, and its usefulness depends on the data and conditions in which it is tested.
Feedback was useful when agents viewed it
In the field experiment, the comparison group received the company’s usual coaching. Agents offered the new tool could also see daily feedback through a dashboard. Some saw their own estimated performance, while others were additionally shown a comparison with their team’s average.
The researchers reported limited overall effects from making feedback available. Performance improved in their analysis when agents chose to access it, but the improvement was smaller when peer comparisons were included.
That distinction matters. Accessing the dashboard was a choice, so people who used it may have differed from those who did not. The finding should not be read as proof that requiring everyone to check a score would achieve the same result.
The performance outcome was based on the model’s predicted ratings. Businesses should be careful not to treat a prediction as an independently observed customer response.
What managers can test
A team considering automated feedback could start with a limited trial and compare the tool’s assessments with actual customer surveys, supervisor reviews and evidence that problems were resolved. It could also ask agents whether the information is understandable and useful.
Speed is not the only relevant measure. A short call might reflect an efficient solution, but it might also leave a customer with an unresolved problem. Quality, repeat contacts and customer outcomes should be considered alongside call duration.
Our report on AI workflow examples examined a separate business experiment. Both studies address how organizations use AI, but neither establishes that a particular tool will work for every company.
One company, one setting
The study involved customer support for a business software product at one company. Three authors were current or former employees of the company, and the company reviewed and authorized use of the material, the paper says.
Those details do not invalidate the research, but they help readers assess its setting. Other businesses may use different systems, serve different customers or measure satisfaction differently.
Employers also need to consider privacy, appropriate employee notice and applicable rules before analyzing conversations. The strongest case for a feedback system is not that it produces a score quickly, but that staff can use reliable information to improve service.
