Skip to main content
AI and machine learning July 20, 2021

Predicting burn-out risk in the digital era

Tired man sitting in front of computer monitor.
How artificial intelligence could help us prevent severe cases of burnout in organizations

“Nineteen years ago, my ‘perfect’ life fell apart. I had no energy. I could not handle sound, light, or movement. I could not focus or handle information. I could not eat, sleep, or cry. My hard disk was full and my batteries were flat. I felt like a total failure. I just wanted to disappear, sleep, and have no one expect anything of me, ever. I was convinced that I would never work again.”

Wenche Fredriksen shared her experience of burning out and the slow path to recovery on her LinkedIn page. [i] The effects of burnout on individual well-being can be devastating. Burned out employees are emotionally and physically exhausted, and find it difficult to focus on their work, their families, and themselves. Research has linked the experience of burnout to depression, mental and physical health issues, and suicide. [ii] It affects employees across professions and geography, [iii] even in countries reputed for their work-life balance like Sweden. [iv] 

The severity of burnout is slowly rising in our collective consciousness as a disease of our time (Berlin recently had a musical called “Das Burnout”). [v] Since we also live in a time where artificial intelligence (AI) promises to solve many of our problems, could it also help us prevent employees from burning out?

Prevalence and causes of burnout

Fredriksen’s story is far from unusual. Recent studies show that burnout affects 1 in 10 employees on average in Europe, [vi] which has not only dire consequences for the individuals but also a considerable economic cost. A recent article from Gallup estimates that burnout, depression, or anxiety disorders cost the German economy €9 billion annually in lost productivity.[vii] The WHO estimates that the cost of depression and anxiety disorders, which can result from burnout, is around $1 trillion each year to the global economy.

Burnout is an extreme manifestation of continuous exposure to a stressful workplace environment. Workplace stress is experienced by 1 in 4 employees regularly. The social and economic costs of this “milder” but more common form of burnout are still staggering. Research in the US estimates that workplace stress is responsible for an excess of 120,000 deaths annually [viii] and costs US employers over $300 billion each year. [ix] (These figures should not hide the difficulties and debates regarding the measurement of the prevalence and economic impact of burnout, due to issues of awareness, the use of different definitions, and the precision of our current means to assess them.)

While burnout is experienced by individuals, our current understanding of the phenomena points to organizational causes. [x] Burned out individuals are not weaker, less resilient, lower performing, or less competitive than the rest, but they can become so over time because they are being burned out. Burnout is also not a simple and direct consequence of working long hours. [xi] Burnout emerges from bad working conditions [xii] characterized by an imbalance between the demands placed on the individual and the resources that the individual has to meet these demands. Resources can be monetary, but also psychological or social, such as support from colleagues or from one’s supervisor. By contrast, demands are all those elements that require efforts from an employee – the amount, difficulty, and complexity of the tasks that have to be accomplished; the psychological, emotional, and cognitive efforts to handle difficult relationships at work; or the need to overcome internal or external bureaucratic hurdles. Employees who are subjected to this imbalance over time first experience emotional exhaustion, then cynicism, and finally lose their sense of accomplishment in their work. [xiii]

Organizations are typically not well prepared to deal with burnout. First, because of the widespread misunderstanding that burnout is a problem of the individual that can be solved by toughening up or learning to manage stress better. Which is why organizations handle burnout by encouraging employees to take a day off or to participate in yoga, meditation, or cooking classes. Yoga is great. But burnout is not a weakness of one individual, it is a faulty design of the system. These solutions can definitely help smooth the symptoms, but they will not address the causes of burnout. Second, organizations are typically unaware of the problem itself. The realization that burnout is a widespread and serious organizational health issue is recent (WHO first classified burnout as an occupational phenomenon in 2019). The consequences of burnout on employee productivity, absenteeism, and turnover – with their associated costs – are still poorly known or documented. In interviews with dozens of human resources managers across the world, I discovered that many HR professionals do not know the word, let alone its causes, or consequences. Third, detecting burnout is not as simple as it seems. The most sensitive organizations administer large-scale surveys to their employees to assess their engagement, psychosocial working conditions, or levels of stress. However, these surveys are typically better designed to obtain a broad understanding of the prevalence of burnout or of the level of engagement of employees, rather than to identify specific individuals who are likely to burnout and to intervene in a timely manner. As with Fredriksen, this leads to people discovering that they are experiencing burnout when they cannot work anymore, and they seek medical attention. At this stage, the route to recovery can take months or years.

How could artificial intelligence help?

Artificial intelligence will not solve the problem completely. It cannot replace the training of HR professionals in assessing the risk of burnout and intervening appropriately, their judgement in handling complex and sensitive situations, or their ability to help and support people. However, it will give us a better chance to solve it. Artificial intelligence can improve the information available to HR specialists, managers, and employees themselves to identify who is at risk of experiencing severe burnout. This can already be quite powerful. Imagine if we could predict who is likely to suffer from burnout two or three months ahead. Imagine if we could know which parts of the organizations, or which groups, are currently under extraordinary stress. This information would enable HR specialist, managers, or employees to talk to the people at risk, monitor well-being, seek help, and act preemptively.

Can we do this? Artificial intelligence algorithms can help us make sense of large volumes of data about our behavior, identifying specific patterns that can be related to an outcome of interest, like burnout. Incidentally, our life and work in organizations produce vast amounts of electronic trace data about our behavior, especially about our collaboration behavior. One specific source of data is in fact not only a silent record of our collaboration behavior but also a potential source of strain and stress itself: email. [xiv] Email is widely used (and overused) in organizations. Estimates place the number of emails received by each employee during a single day to be around 121, [xv] and a survey by Adobe shows that employees spend on average 3.1 hours checking work emails per day.[xvi] My own research shows that the number of emails per employee varies greatly, with some employees receiving and sending several hundreds of emails per day, every day. Even when emails from distribution lists, spam, or emails to and from external contacts are excluded – therefore limiting the exchanges to interpersonal collaboration with direct coworkers – I found that employees send and receive about 80 emails per day on average, with a maximum of over 400 emails in a single day by one employee.[xvii] Email is just one way in which we communicate at work, but it is a reliable indicator of collaboration, [xviii] it is ubiquitous, and is (relatively) easy to collect.

So, the more precise question is, can artificial intelligence algorithms use these traces of electronic exchanges to accurately predict employees at risk of burnout? Well, the evidence is still not conclusive, but it is promising. Across two different studies in medium-sized companies using increasingly sophisticated prediction models, my collaborators and I have been able to predict whether an employee is at risk of experiencing burnout with an accuracy of over 80 percent.[xix] We are currently replicating these studies using random forest prediction models in a third (larger) company with three aims in mind: one, show that we can replicate these prediction rates; two, identify which predictors are important across organizations; and, three, determine how much in advance we can reliably predict burnout. Our goal is to give organizations a tool that continuously and unobtrusively monitors email communications and raises flags in advance when an employee or a group of employees is at risk of burnout.

A different question is, should we do this? Clearly, continuously monitoring employees’ email communications raises privacy concerns. [xx] However, these concerns can be addressed by correct handling. First, the algorithms don’t need sensitive information such as the content of the emails. Only information about email traffic (i.e., sender, recipient, and time) are required to predict the risk of burnout. Second, while employees need to be identifiable in order to calculate a risk score, the score provided is simply a probability from 1 to 100 that does not say anything specific about the employee, apart from the risk of developing burnout. Finally, while scores are calculated for each employee, they could be reported at the group level, therefore removing any personally identifying information. Managers could then be notified of the possibility that one or several of their employees are experiencing burnout. More generally, in this case, as in most issues that touch individual privacy, the question is one of purpose. Is the purpose of using these three personally identifiable pieces of information sufficient to compensate for the associated loss of privacy?

The impact of lockdown

The increase in remote and virtual collaboration caused by lockdown makes this type of initiative even more relevant. On the one hand, remote work has further reduced the separation between work and personal life, and has resulted in increased stress levels. [xxi] On the other hand, the increase in the digitalization of the workplace also means that employees are communicating almost exclusively using digital communication tools, which enables these traces of interactions to be collected and to be even more representative of the collaboration flows that underlie the activity of any organization.

Organizations can and should prevent their employees going through experiences similar to Fredriksen’s, if not for their genuine concern about the well-being of their employees, at least because of the economic cost it generates for them. It is unlikely that one idea or solution on its own will solve the problem, and many approaches should be tried together. The idea that I exposed here – applying artificial intelligence algorithms to logs of email communications to detect employees or group of employees at risk of experiencing burnout – is a moonshot and still only a maybe. However, its risks and costs are minimal compared to not doing anything.


[i] Fredriksen, W. 2019. “Rising from the ashes – A story about burnout and the way back to life.” LinkedIn, September 18, 2019.

[ii] Carod Artal, F. and Vázquez-Cabrera, C. 2013. “Burnout Syndrome in an International Setting.” Burnout for Experts: Prevention in the Context of Living and Working. 10.1007/978-1-4614-4391-9_2.

[iii] ibid.

[iv] Savage, M. 2019. “Burnout is rising in the land of work-life balance.” BBC, July 26, 2019.

[v] Turner, Z. 2016. “‘Das Burnout’: An Epidemic in Germany.” The Wall Street Journal, May 23, 2016.

[vi] Sixth European Working Conditions Survey: 2015. Eurofound.

[vii] Nink, M. 2016. “The High Cost of Worker Burnout in Germany.” Gallup, March 17, 2016. 

[viii] Goh, J., Pfeffer, J. and Zenios, S. 2015. “The Relationship Between Workplace Stressors and Mortality and Health Costs in the United States.” Management Science, March 13, 2015.

[ix] Pfeffer, J. 2018. Dying for a Paycheck: How Modern Management Harms Employee Health and Company Performance―and What We Can Do About It. New York, NY: Harper Business.

[x] Demerouti, E., Bakker, A. B., Nachreiner, F., and Schaufeli, W. B. 2001. “The job demands-resources model of burnout.” Journal of Applied Psychology, 86(3), 499–512.

[xi] Wigert, B. 2020. “Employee Burnout: The Biggest Myth.” Gallup, March 13, 2020.

[xii] Wigert, B. and Agrawal, S. 2018. “Employee Burnout, Part 1: The 5 Main Causes.” Gallup, July 12, 2018.

[xiii] Demerouti, E., Bakker, A. B., Nachreiner, F., and Schaufeli, W. B. 2001.

[xiv] Newport, C. 2021. “E-mail Is Making Us Miserable.” The New Yorker, February 26, 2021.

[xv] Campaign Monitor. 2019. “The Shocking Truth about How Many Emails Are Sent.” May 2019.

[xvi] Naragon, K. 2018. “We Still Love Email, But We’re Spreading the Love with Other Channels.” Adobe Blog, August 21, 2018.

[xvii] Quintane, E. and Carnabuci, G. 2016. “How Do Brokers Broker? Tertius Gaudens, Tertius Iungens, and the Temporality of Structural Holes.” Organization Science. November 22, 2016. 

[xviii] Quintane, E. and M. Kleinbaum, A. 2011. “Matter Over Mind? E-mail Data and the Measurement of Social Networks.” Connections. 31. 22-36.

[xix] Estévez-Mujica, C. and Quintane, E. 2018. “E-mail communication patterns and job burnout.” PLOS One, March 8, 2018.

[xx] Hern, A. 2020. “Microsoft productivity score feature criticised as workplace surveillance.” The Guardian, November 26, 2020.

[xxi] Spataro, J. 2020. “A pulse on employees’ wellbeing, six months into the pandemic.” Microsoft 365 Blog/Work Trend Index, September 22, 2020.

Add new comment