5,172 employees, one AI assistant: this is what changed in the workplace

In every boardroom, there is talk about implementing artificial intelligence in the workplace, but what do we really know about what works well and what doesn't? In this blog, I discuss a case in which five thousand call center employees were given access to artificial intelligence. Scientists have investigated what this was like for these employees before the introduction, after the introduction, and what results were achieved. The researchers published this in early 2025 by Oxford University*. I will tell you what the impact was and what lessons you can learn to make the implementation run smoothly!  

What was the impact of implementing AI? 

Productivity went up. Employees solved an average of 15 percent more customer problems per hour. This was mainly because conversations were handled faster and employees were better at multitasking. 

The biggest gains were among beginners. Employees with less experience or lower skills in particular made great progress: up to 35 percent more resolved tickets per hour. AI acted as a kind of 'mental shortcut' that transfers successful patterns of top employees to others. 

The effect was smaller for the best-performing employees. Sometimes even a slight decrease in quality and customer satisfaction was measured, possibly because they started leaning on AI suggestions that did not match their own high standard. 

What changed for customers and employees? 

The tone of customers became more positive, and they asked for a manager less often. This indicates a smoother course of conversations and more trust in the officer. 

There was a decrease in turnover among employees with less than 6 months of experience. AI seems to make the job less frustrating for beginners, which increases the likelihood that they will stay. 

Employees who used AI suggestions more often learned faster. Even during AI failures, they proved to work faster and more effectively than before. This is evidence that they not only copied, but learned from the patterns that AI provided. AI trained them to become better at their job! 

How exactly did they implement it? 

The AI assistant participated in the chat channel and provided real-time suggestions for responses and relevant documentation. The employee could choose to adopt, adjust or ignore these suggestions. So the AI did not replace the employee, but really functioned as an assistant. 

The introduction was done step by step, which made it possible to compare performance before and after access to the assistant, and with colleagues who did not (yet) have access. This made the effect of AI easy to measure.  

To start, employees underwent a 3-hour training course on how to use AI for their work. They were trained online in small groups by two trainers. Afterwards there was little support. 

What does this mean for AI design and implementation? 

1. Design AI as an assistant, not a replacement. The employee's turn remained here . That helps quality and acceptance, especially in customer contact. 

2. Look beyond saving time. Also take into account what happens to the employees. Does it make them happy? Especially in sectors where there are many shortages in employees, such as healthcare or education, this can lead to more satisfied employees and less outflow.  

3. Think of AI adoption as a learning path. If you start today with an initial, well-designed application, you will not only build productivity. You also build "AI work skill" in your organization. Employees learn how to work with an assistant, how to weigh suggestions, how to keep control. And that makes it easier to implement new possibilities later on, even if AI continues to develop quickly and, for example, can do more with speech, summarizing or listening in. 

The core is simple: start, measure, improve. And design the work in such a way that humans and AI reinforce each other. 

*Source Oxford University Research 2025: Brynjolfsson, E., Li, D., & Raymond, L. (2025). Generative AI at work. The Quarterly Journal of Economics, 140(2), 889-942.