Algorithmic Management Consequences forWork Organisation and Working Conditions

Algorithmic management is a high-profile term but it is much less clear what the term means. If you ask technology companies, they will tell you that they can automate all the managerial processes. This way, you will end up with a very overexaggerated account of what these technologies actually do in the realm of workplace. The next stop are employers and managers who will say that the processes are not fully automated and that the algorithmic management assists them in day-to-day operations. If you ask workers, they usually do not have a full understanding of the degree to which technologies affect their work and to what extent the processes are automated. In order to develop survey questions which workers would understand and enable them to indicate algorithmic management, it is first necessary to grasp how algorithmic management is embedded in the world of work, and acquire a detailed account of related developments in the workplace.

If we understand algorithms as processes and sets of rules to be followed in calculations or other problem-solving operations (OEC Online,2021), there is nothing new about them — businesses have had processes of rules to be followed in making decisions for hundreds of years. In his famous writings on bureaucracy, Max Weber talked about step-by-step distributed and nominally objective procedures for selection and sorting, both typical for the decision-making in modern bureaucracies.

Following this line, if algorithms have been used by businesses for long time — what is then new about algorithmic management? The key aspect is an explosion in computing power and digital data collection through cameras, sensors, audio devices, biometrics, text, etc. The potential for using algorithms when you have so much more data has massively increased, making the use qualitatively different. The term ‘algorithmic management’ was the first used in a paper by Lee and colleagues (2015) to describe how Uber was coordinating and controlling their workforce, and since then, it has really taken off.

Algorithmic management is broadly studied in the context of platform work and warehouses (especially, Amazon). However, increasingly, aspects of algorithmic management have been identified in other conventional employment settings: retail, manufacturing, marketing, consultancy, banking, hotels, call centres, and among journalists, lawyers, and the police. In this blog, we will primarily discuss platform work and Amazon examples but also provide supporting illustrations from conventional employment.

At the moment, algorithmic management reshapes working conditions and organisations through the automation of:

a) direction (what needs to be done, in what order and time period, and with different degrees of accuracy);

b) evaluation (the review of workers’ activities to correct mistakes, assess performance, and identify those who are not performing adequately);

c) discipline (the punishment and reward of workers in order to elicit cooperation and enforce compliance) (Kellogg et al., 2020).

In ride-hailing and food delivery sectors, algorithms automatically direct workers by allocating fares and orders and then providing a route for workers to follow via a GPS map. Even when we look at this example, which seems like a quite clear-cut algorithmic direction, there are differences in the current body of research as to whether workers are receiving instructions or just a suggestion. Indeed, some research finds that these workers retain agency and platform algorithms are rather shaping actions rather than directing them; whereas in other studies, it seems like workers face evaluation and disciplinary measures if they do not follow the algorithmic direction. For example, some research highlights how in food delivery, a human supervisor contacts workers if they headed in a different direction or taken too long compared with the app’s estimates.

In conventional employment, a lot of research has been conducted on the Amazon warehouses. Workers are equipped with hand-held ‘scan guns’ or wearable devices that combine barcode scanners, motion and location tracking, and a display. Such devices automatically assign optimal items to workers and direct them around warehouses based on their location, increasing productivity and efficiency. Due to COVID-19 pandemic, there is a growth in installing cameras that use machine learning to alert workers when they are breaking social distancing rules in warehouses. Cameras are utilised in delivery vehicles to instruct workers to maintain safe distances, slow down, avoid unplanned stops or to take 15-minutes break if yawning.

There are examples of algorithmic direction in various sectors, including retail, manufacturing, and healthcare. For instance, a company Klick Health alerts workers when projects are behind schedule, notifies outstanding and urgent things to do and reduce distractions hampering productivity. A consultancy company, Publics, uses algorithms to create teams for new projects, meaning that the algorithm chooses different workers based on their skills and perceived performance in order to create the best teams combining account managers, coders, graphic designers and copywriters from across the organization. While an interesting aspect of scheduling software is the ability to automatically assign workers; when managers got interviewed, they said that they did not actually use this functionality. The managers preferred to use the system for the staffing overviews and predictions that it provided, but would then manually schedule workers themselves.

Moving on to algorithmic evaluation, use of customer ratings is wide-spread in the platform work. Some studies argue that, in addition to the customer ratings, Uber collects data on worker acceptance rates, their braking and acceleration speeds. Uber Eats goes further and collects metrics such as the number of weekend shifts completed after 8PM, their average weekly hours, how often workers have not shown up for a shift (if they have been assigned), data on when they log in late, preferences toward a particular shift, the number of orders carried out per hour as well as declined orders, actual time to reach the restaurants and customers (which are then compared to the optimal time based on the GPS and map data).

In online platform work, customer ratings play a central role, but the platforms also evaluate workers on the numbers of jobs they have completed, the length of their relationships with clients, as well as collecting data on their keyboard presses, and screenshots of work progress.

In conventional sectors, research from Amazon warehouses show that wearable devices record the number of products picked per hour, automatically comparing the ‘pick rate’ with a target based upon previously achieved pick rates. With workers having to hit the average from previous shifts. This is aggregated together to rank individual workers relative to their colleagues using metrics around individual speed, productivity, accuracy, and errors in real time, which are all bundled together into a single, composite assessment of their performance and matched to a normal distribution curve. The poor performance of the lowest 10% of workers is flagged up to human managers who then request workers speed up.

Similar methods are particularly applied in retail. In my own research, the system was officially referred to as “my guide system”, but the workers called it “my slave system”. Such systems allocate predefined time to put out different stock from the crate. If they do not perform as predicted, it would be flagged up to a human manager. In hospitality, research shows how TripAdvisor ratings and reviews have been incorporated into individual performance management of workers but also in weekly team meetings.

Frequently, workers are evaluated even before getting a job. Systems such as Equifax Kronos, SnagaJob, and Recruit use predictive analytics to algorithmically process and sort applications. Workers are evaluated by these systems and must comply with what the algorithms are deeming to be the right worker, the right fit for a particular job and a particular team.

Last but not least is algorithmic discipline that is closely related to the algorithmic evaluation. In ride-hailing platforms, workers are automatically deactivated if they have low ratings (below around 4.6, depending on the city and labour supply). Workers do not have a right to appeal but can ask for their deactivation to be reviewed. In some cities, workers can undertake Uber’s “quality improvements course” at their own expenses. However, currently there is a discrepancy between academic research (focusing on how workers are automatically deactivated) and industry claims. For example, Uber says that deactivations are all manually reviewed by a human manager; and thus, cannot be categorized as algorithmic management. This discrepancy blurs research attempts to establish what is really going on in the world of work, and undermines policy interventions in regard to improving working conditions.

In food delivery sectors, when companies like Deliveroo use a scheduling system, lower ranked workers are restricted in their ability to access shifts because workers with the highest rates get first pick of the most profitable hours. In online platform work, workers are usually not deactivated from the platform; yet their access to jobs is impacted. For example, at the top of 30-page searches are workers who are believed to be the best in terms of their job success score, their skills, hourly rate, amount of work completed. Clients are very likely to choose workers who are evaluated by the platform as being the best.

In terms of algorithmic discipline in conventional employment, for instance, Amazon relies on individual performance scores to decide who should be fired. As some scholars argue, “management consists of executing decisions based on data analytics”, meaning that managers are not really managing people, but rather carrying out the decisions made by the algorithms. On the other hand, Amazon challenges such notions. The company claims that managers are the ones who are deciding whether somebody should be fired or not. The system just highlights productivity issues. Earlier mentioned AI cameras installed in Amazon delivery vehicles have been used for disciplinary actions including firings, if workers have driven too fast, being involved in collisions, or made unscheduled stops. In hotels, workers have been fired because of ratings and reviews on Trip Advisor.

How to secure well-being of workers in the age of algorithmic management?

As for the potential consequences for working conditions, the increased potential to control workers as well as the quality of workers and product standards, creates more opportunities for what David Weil calls the “fissured employment relations” or the use of temps, outsourcing, franchising, labour brokers and platforms. Companies do not need to provide so much training to workers to ensure good performance, because workers are being directed and highly evaluated. Furthermore, workers who cannot meet requirements are subjected to disciplining or firing. All together this increases the potential for fissured employment relations.

Another potential consequence is reduced reliance on low-level managers and supervisors. The algorithms are taking the role of low-level managers and supervisors who are no longer required to use their own judgment but rather just passing on the decisions made by algorithmic management. The first study on algorithmic management by Lee and colleagues already identified this phenomenon reporting that “a few managers in each city to oversee myriads of workers (hundreds in a city and thousands of drivers on a global scale) in an optimized manner at large scale” (Lee et al., 2015).

Moreover, potential consequences include reduced intrinsic skill use and the standardization of the work; and reduced managerial agency and dispossession of workers knowledge. Such a situation may lead reduction in companies’ investment in skill development, because it’s not so necessary anymore.

Algorithmic application in the world of work is likely to lead to more ‘effort biased technical change’, causing an increased evaluation, discipline, and direction of workers — or in other words, intensification of work. In many jobs in warehousing, this leads to a reduction of worker autonomy; in some platform work, workers might experience an increase in autonomy due to a reduced role of managers. Even though platform workers are still tightly controlled by algorithms, they might perceive the overall situation as having more autonomy. Algorithmic evaluation can also create insecurity as well, e.g., anxiety over how a worker will be rated and profiled by algorithms. Overall, the workforce will be exposed to an increasing level of stress and anxiety, both of which are very harmful to wellbeing and health.

Algorithmic management practices may give rise to new practices of resistance and fuel workplace conflict, which Kellogg and colleagues (2020) call ‘algoactivism’. For example, Uber drivers resist the algorithmic allocation of work. They turn off their driver mode when in bad neighbourhoods, so they are not assigned rides, or they stay in residential areas to avoid bar patrons and they are frequently logging off to avoid long trips. Also, journalists have been found to resist algorithmic evaluation by manipulating the variables they enter into the evaluation systems. Workers are continually finding ways in which they resist the system and do act in a way which has not been perceived by the designers of technologies.

In addition to algoactivism, there are other available mechanisms to mitigate negative consequences on working conditions. From the legal perspective, GDPR is a useful framework for worker protection. Under the article 22, algorithmic management would be illegal because any serious decision significantly affecting workers as data subjects needs to be reviewed by a human. This is one reason why companies keep insisting how and why human manager are part of the process. Therefore, instead of having endless discussions on what is automated and not, we need to focus on collective responses. Collective bargaining and co-determination, especially when Swedish and German experiences considered, show the potentials for the collective regulation of the workplace. We should not see algorithmic technologies as separate to other issues of job and employment quality.