AI Now Scores Every Call, and Workers Cannot See the Rules

When call centres boomed at the end of the 1990s, scholars described them as “a situation in which the operator has an assembly line in the head”: programmes routed the calls and dictated the pace of work, while workers were closely monitored both by supervisors and by software that recorded their interactions with customers.

With the arrival of artificial intelligence, call centres are once again at the forefront of technology-driven workplace management. While software in the early years did not replace supervisors’ decision-making in managing the performance of frontline workers, AI technologies are now far more pervasive and play a central role in monitoring, evaluating, and rewarding employees.

Over the past year, we have been surveying and interviewing workplace representatives in UK-based call centres in the telecommunications and finance sectors, to understand how the introduction of AI at work is negotiated and managed, and how it is affecting workers.

Call centres now make extensive use of AI-powered software that performs sentiment analysis on calls between employees and customers. We run a survey in November 2025 with more than 700 workers participating across multiple call centres; our findings show that 81.3 per cent of agents in the telecoms and finance sectors are subject to constant voice recording, which typically includes sentiment analysis. This was ostensibly introduced to offer a more consistent and objective assessment: whereas supervisors previously listened to a sample of calls to evaluate agents’ performance, an AI can now analyse the totality of calls and allegedly reach a more accurate assessment of each worker’s overall average performance. Moreover, the sheer breadth of the analysed data makes it easier to identify recurring problems during calls and to pinpoint related training needs.

However, some workers have reported unexplained drops in their performance scores since the sentiment analysis software was introduced. Developers at external IT providers and managers within the organisations typically reassure employees that only the technical content of a conversation is analysed, but many of our interviewees, union representatives across eight call centres, expressed doubts about the broader potential of sentiment analysis tools. They shared concerns that the tone and quality of interactions might also be evaluated, and many suggested that the software may be sensitive to accents and could disadvantage neurodiverse employees. Above all, a lack of clarity and transparency in the scoring criteria leaves employees uncertain about why they are penalised and how to put it right.

The pervasive use of sentiment analysis is a source of considerable stress for workers, not least because, according to interviews with union representatives, sentiment-related performance metrics can affect the variable pay component by up to 40 per cent. In some call centres, employees can ask for negatively scored calls to be reviewed by a manager; in others, such recourse is hard to obtain. Many managers, moreover, appear to have only a limited understanding of how the algorithms shaping performance evaluations actually work.

In our interviews, we found that the ability of managers to correct AI mistakes is a key factor shaping employees’ perception of how fair the evaluation and reward processes are. We have seen that in workplaces where workers can more easily challenge the AI’s scoring and a manager can act as the “human in the loop”, feelings towards sentiment analysis tools appear less negative.

More generally, our findings suggest that fairness concerns extend well beyond the accuracy of performance scores. Workers and their representatives worry not only about whether AI gets evaluations right, but also about their ability to take part in decisions on the adoption and use of these technologies. In many instances, consultation took place only after decisions on AI adoption had already been made, leaving little scope for employees to shape how AI systems would affect performance management, rewards, or disciplinary processes.

Just as the call centres of the late 1990s became emblematic of a new era of digitally mediated management, today’s AI-powered workplaces offer an early glimpse of a phenomenon that could soon affect the wider economy: algorithmic systems might replace a substantial share of human-led decisions on supervision and performance evaluation well beyond call centres.

To avoid this scenario, more stringent legal requirements on meaningful human oversight should be introduced in the UK, which currently has no AI-specific legislation and has so far opted for a principles-based, sector-led regulatory approach. By contrast, the European Union adopted the AI Act in 2024 and is rolling it out gradually, including a requirement of human oversight for high-risk AI applications such as algorithmic management, which will apply from 2 December 2027, after the Digital Omnibus pushed back the original deadline .

Introducing a legal requirement of human oversight in the UK could address the problem of wrong and unchallengeable AI-led evaluations that call centre workers raised so often. For this to work in practice, managers would need adequate training and transparency about AI mechanisms and criteria, which could in turn improve their own experience in terms of skills development, autonomy, and self-efficacy.

Furthermore, social dialogue at workplace level should also be strengthened in the UK, where union voice is comparatively weaker than in many EU countries. Workplace-based committees made up of workers and their representatives should be set up to monitor the implementation of AI, and meaningful information, consultation, and negotiation rights should be granted whenever AI is introduced at work. These measures would improve transparency and worker participation, which are crucial not only for the wellbeing of employees but also for securing workers’ trust in their organisation’s processes.