Confronting algorithmic management using subject access requests
Insights using the case of food-deliveries
Luca Perrig 
Introduction
Digital platforms have long been considered a driver of marketization. By reducing transaction costs, ubiquitous internet access was thought to enable easier communication and hence an increasing reliance on markets, as forecasted by transaction cost economics (Gurbaxani and Whang 1991; Malone et al. 1987; Munger 2018). In addressing this evolution, sociologists have instead highlighted how digital markets are shaped by technical mediation (Bessy and Chauvin 2013; Gillespie 2010). In this perspective, platforms are seen as shaping markets, as they allow trust (Luca 2017), enhance competition (Velthuis and Van Doorn 2020), allow evaluation (Kornberger et al. 2017), and set rules (Kirchner and Schüßler 2019). It seems that this view of digital platforms as market enablers is shifting, however, and they are perceived as more authoritarian forms of economic organization. Platforms are now increasingly referred to as central planners (Lehdonvirta 2022; Pistor 2020) or as corporations acting as marketplaces (Frenken and Fuenfschilling 2021). The question of information on markets once more moves centre stage, reviving the debate around planning and free markets (Van Basshuysen 2022). In the early twentieth century, the ‘Socialist Calculation Controversy’ revolved around the question of knowledge and whether government planning was capable of generating sufficient information about the economy (Hayek 1945; Mirowski and Nik-Khah 2017; Preda 2009). With the advent of ubiquitous computing, platforms have vast quantities of data at their disposal that they can use to organise entire markets in ways that are reminiscent of centrally planned economies (Lehdonvirta 2022; Posner and Weyl 2018). This is most obvious in the gig economy, where workers on platforms such as Uber are not offered the possibility either of choosing their rides or negotiating their compensation, despite being considered ‘independent contractors’ in many jurisdictions (Hießl 2022).
This paper argues that gig economy platforms operating along these lines are attempting to automate a marketplace. By relying on digital data provided by diverse means – such as sensors, third-party APIs, or workers’ self-declaration – gig economy platforms are attempting to supersede a core function of marketplaces, namely human evaluation (Dieter et al. 2019). But can we even talk of a marketplace when evaluation is automated? Is ‘automated marketplace’ not an oxymoron? And how does this economic agencement affect the ultimate allocation of rides and money? This paper will address these questions through the case study of online food-delivery work.
As Aspers et al. (2022) rightly note, the literature on the platform economy has developed into two broad fields. The first considers platforms as spaces of trade and focuses on e-commerce. The second considers platforms as spaces of work and focuses on the gig economy. The literature addressing gig economy platforms focuses mainly on issues of autonomy and control (Schreyer 2024; Wood et al. 2019), surveillance (Newlands 2021), or resistance (Cant 2020; Woodcock 2021). These things have been essential in countering the platforms’ narrative that they are merely providing intermediation services (Aloisi and Stefano 2022; Prassl 2018). Our approach in this paper, however, is to endorse both the managers’ and the workers’ perspective. This provides us with a novel insight into the functioning of algorithmic management, which allows us to critically assess the platforms’ ambition of building an automated marketplace. It will most notably allow us more accurately to discern its limitations and the struggles that platforms face. In this sense, it contributes to a literature aiming to question the power of platforms, eschewing a deterministic view (Ferrari and Graham 2021; Joyce et al. 2023). For this purpose, the theoretical tools of market sociology are particularly useful, allowing us to understand what exactly platforms are building, if not a market.
The sociology of labour markets is rarely pursued through the lens of market sociology (Aspers et al. 2022). Studies of labour markets often rightly highlight their heavy social, biographical, cultural or institutional embeddedness (Sallaz 2013), and assign them to a separate area of study. It is indeed difficult to conceive of labour markets as a meeting of supply and demand. Networks affect the matching of job seekers and recruiters; interviews are core sites of discrimination; and women earn significantly less than men for the same job. Gig economy platforms explicitly promote their model as a remedy to these evils (Schor 2021). By automating the recruitment process, such platforms supposedly offer a marketplace devoid of human biases (Aloisi and Stefano 2022). Recent research, however, suggests that inequality and discrimination are observed in online labour markets, too, though in a different form (Fiers 2023). As we will see, marketplace automation requires extensive commodification, which is particularly delicate in the context of a labour market. This will lay bare the numerous instances of algorithmic fissures, in other words, failures in algorithmic management (Ferrari and Graham 2021). As we will see, it is these failures that lead platforms to come up with alternative – and more controlling – means of securing consent.
In order to study this novel economic agencement, we will follow Callon (2021) and his notion of market framings. In his conception, a market transaction can be successful only if it is appropriately framed (Callon 1998). Among the framings he identifies we will focus on the matching and pricing mechanisms that platforms implement. These are two of the most fundamental market functions that the gig economy platforms are attempting to automate, and they will constitute our main focus in the remainder of the article.
The structure of the paper is as follows. The next section provides some more insights from the literature, focusing more particularly on the automation of matching and pricing. The third section presents the methods and data on which the analysis will draw. Section four describes the functioning of algorithmic management in the case of food deliveries. The following two sections provide the empirical results – section five focuses on matching and section six on pricing. The last section concludes with a discussion.
Automating marketplaces
Digital platforms are increasingly automating their users’ decisions. Recent developments in the economy at large show that platforms are attempting more and more to come up with solutions rather than providing resources (Iliadis 2023). Decisions about what content to consume and at what price are handled by platforms in numerous instances: autoplay on YouTube serves up users’ next video, Amazon automates the prices offered by sellers on its marketplace, and generative AI is supposed to replace traditional search, to cite only a few examples. In an enthusiastic account of this perspective, (Posner and Weyl 2018, 292) tellingly anticipate that ‘people will not make choices but simply accept goods and services sent to them by computer programs’. In the same vein, some platforms are becoming more than infrastructures (Çalışkan 2021); they are rather evolving as automated marketplaces. As Pardo-Guerra (2019) notes, automation can indeed be considered an extension of infrastructure. As the market infrastructure is ever more pervasive, it ends up ‘automating away’ the market (Birch 2020). When transaction parties have no say in the valuation of the thing exchanged, what remains is indeed a form of economic agencement that cannot be called a market, and whose functioning still remains to be addressed in the literature.
As Viljoen et al. (2021) argue, the development of automated mechanism design closely follows the recent literature in the field of market design. Economics has long been used to legitimate market engineering, even when the resulting economic agencement can hardly be considered a market (MacKenzie 2008). Indeed, the definition of market has been considerably revised in economics. Following the teachings of Roth (2002), markets are not considered as spontaneously emerging institutions anymore, but as a form of economic organization that can be framed and improved (P. Aspers 2011; Callon and Roth 2021). Digital platforms have been prime institutions for the development of such economic engineering (Athey and Luca 2019; Nik-Khah 2023; Van Basshuysen 2022). As we briefly discuss below, this market engineering has been most notable in relation to matching and pricing schemes.
Automated matching
Algorithms handling the matching of supply and demand have been of great interest to economists in the past few decades (Roth 2018; Simioni and Steiner 2024). Their aim is to correct inefficient markets by conceiving matching algorithms that would maximize public utility. Some have been applied to job matching in labour markets, such as the market for medical internships or the one for economics PhDs. The schemes devised by economists in these instances often rely on market participants declaring their preferences, typically by ranking their preferred outcomes. An algorithm then tries to arrange the best possible allocation of jobs. Platforms in the gig economy similarly design algorithms to match workers and tasks. They depart from these textbook market designs in two regards, however. First, they do not seek to maximise public utility, but rather to maximize the number of transactions. Second, they do not inquire about workers’ preferences. Rather, they attempt to make inferences about workers’ propensity to work based on the data they gather about the market. In this regard, it is instructive to look at the uses of big data in hiring.
Hiring has a long history of using data as input for decision-making. Most often, this data is processed by humans, typically HR employees. This has been identified as a key source of discrimination or inefficiencies. Indeed, humans will assess job seekers’ moral character on the basis of their social media profiles or even from their credit reports (Kiviat 2017; McDonald et al. 2022). Automation is seen as a better alternative than human processing, and data-driven hiring systems are a growing industry. Processing data automatically is seen as a way to moralize the market and avoid discrimination or biases (Ajunwa and Greene 2019; Dencik and Stevens 2021). This process of building algorithmic identities for individuals in order to assess their adequacy for a job is similar to the profiles that gig economy platforms build for couriers (Cheney-Lippold 2017; Perrig 2023). These hiring systems use opaque algorithms and in this regard they are very much akin to the algorithms in use in the gig economy (Ajunwa 2020).
A key difference from traditional hiring in the case of the gig economy, however, is that this process is repeated many times per shift, as every ride is the occasion of a new ‘hiring’. Rather than one-shot hiring, couriers are constantly being evaluated by the algorithm for their adequacy to take the next job (Duggan et al. 2020). Hence, they have ample opportunity to become aware of the algorithm and find ways to position themselves better (Bucher 2021; Kellogg et al. 2020). For this reason, our study will lead us to study the reception of algorithmic management among couriers and discuss their algorithmic literacy. Ultimately, the allocation of rides is indeed best understood not only by studying the algorithm, but more importantly by studying its reception, as the outcome is very much shaped by the couriers’ anticipation of the automated matching mechanism (Bucher 2021).
Automated valuation
The platformization of the economy has opened out to encompass a wide range of pricing mechanisms on marketplaces. Auctions immediately come to mind, which are particularly well suited to e-commerce. Platforms allow for numerous types of auctions as they are able to closely monitor the information available to all parties (Milgrom 2019; Posner and Weyl 2018). eBay, for instance, has developed sophisticated bidding designs in order to ensure efficient resource allocation, which ultimately increases sales (Hasker and Sickles 2010). Auctions have also been leading the way to the automation of pricing, whether in the case of stock exchanges (Pardo-Guerra 2019) or online advertising (Jobin 2023; Viljoen et al. 2021).
Another, more recent form of digital pricing is data-driven (Ezrachi 2016). In this instance, the platform uses data and attempts to infer users’ propensity to sell or buy and estimate a market-clearing price accordingly. It is possible to distinguish three types of data-driven pricing, sometimes called algorithmic pricing (Brown and MacKay 2023; Leslie 2023; Seele et al. 2021). What these pricing mechanisms have in common is that they automatically set prices, which vary.
- Dynamic pricing evolves in accordance with an anticipation of supply and demand, in time or in space. The data necessary to track market developments thus comprises mainly past transactions, and contextual features of the market, such as the number of similar buyers and sellers at the same location or time. Uber is known for this type of pricing, which they call ‘surge pricing’, as it responds to sudden surges of demand at a particular place and time (Grohmann et al. 2022; Shapiro 2020; Van Doorn 2020). It is also notably used by Amazon, which offers its sellers the possibility to let the platform set the price according to the demand and supply of other identical products (Mattes 2022).
- Personalized pricing, also called behavioural price discrimination, is the practice of asking different prices depending on the buyer’s supposed willingness to pay. It thus requires personal data about individual buyers. Airlines are the typical example (Ezrachi, 2016), but the retail industry also uses tracking devices to personalise prices (Mellet and Beauvisage 2019; Turow 2017), credit scoring (Kear 2017), and the growing ‘insurtech’ industry notoriously personalizes its customers’ premiums (Krippner and Hirschman 2022; Moor and Lury 2018; Sadowski et al. 2024). Gig economy platforms do not personalize workers’ compensation, however, and the price of the same ride will be identical for two different couriers (Teachout 2023).
- Automated valuation refers to the use of data in order to set the prices of heterogeneous products or services automatically. By selecting measurable features of a good or service, a seller can thus set prices in bulk. It can be thought of as the opposite of personalized pricing. Instead of setting different prices for heterogeneous buyers, it consists of setting different prices for heterogeneous products. It is typical of the ‘proptech’ industry, which uses data about real estate (location, number of rooms, and so on) in order to value properties automatically (Fields 2022). Gig economy platforms use similar models to set workers’ compensation, most notably in the case of drivers and couriers. By measuring selected features of each ride (length, weather, time), they automatically set compensation.
Platforms may use a mixture of these ideal-typical pricing schemes. Transportation or delivery platforms such as Uber will, for example, combine automated valuation and dynamic pricing when setting the delivery fee. The pricing formula attempts to anticipate workers’ willingness to work, and an additional ‘surge price’ is sometimes offered in the form of a bonus. This article will focus on automated valuation, namely, how platforms try to anticipate workers’ willingness to work and put a price on it. In both matching and pricing, we thus observe platforms attempting to automate a marketplace.
Methods
This study draws on an ethnography of the online food delivery market conducted in Western Switzerland. It focuses on four different platforms operating in the region. The data were collected by way of participant observation and interviews.
Participant observation
The fieldwork started with an engagement as a bike courier for a period of six months. Working as a courier for such platforms is an individualistic experience, as couriers are spatially fragmented (Heiland 2021), so it might appear an unproductive way of carrying out the fieldwork. Being involved as a participant, however, has many advantages. It allows the researcher to experience the hardships of physical work. As already acknowledged by numerous social scientists, bodily engagement can in itself generate knowledge (Thanem 2019). As will become obvious in the analysis, pain and tedium are essential to the valuation of rides. Delivering meals is exhausting, and couriers will often seek to avoid difficult rides. Another important feature of immersive fieldwork is that it forces the ethnographer to live at the temporal pace of their field site (Dumont et al. 2023). It so happens that time is an important aspect of food deliveries. Of course, the urgency of the ride itself is expected, but less so are the long waiting periods between rides. As we will see, these will often prompt couriers to inquire about the functioning of the matching algorithm. Having first-hand experience thus brings many insights that are inherent in performing the activity. Other important information sources resulting from this participatory fieldwork are the messaging groups on which couriers discuss their experience of the job, share tips, help each other or complain about the working conditions. In the course of a year, messages were garnered from three groups, eventually totalling over 10,000. They were subsequently analysed for their content along with the interview transcripts. A second part of this participant observation took place in the offices of a local food delivery platform. It consisted of observing the daily work of the company employees in their open-plan offices, attending their meetings, and following them as they prepared the next update of their app. This provided exclusive access to the intricacies of algorithmic management. Most important for our purposes, it allowed observation of developers and managers in their struggle to automate a marketplace. This view will allow us to nuance the discourse of platforms as all-powerful organizations. Instead, this view of the inner workings of a food delivery platform showed that securing workers’ consent to work was a recurring concern of platform managers.
Interviews
Participant observation in this individualised field site comes with some limitations, of course, most obviously the fact that the researcher is not subject to their colleagues’ financial imperative. This is an important reason for complementing these observations with interviews. Some 24 couriers recounted their experience of work, their daily interactions with the app, and their beliefs about its functioning. Eleven managers were also interviewed, from four different platforms, who were given the opportunity to speak about their ongoing projects, as well as the history of their company and the challenges they had to overcome. These interviews were transcribed and then analysed using an iterative process between theory and data drawn from Corbin and Strauss (1990). The coding followed a sequential pattern (Tesch 2013), starting with descriptive codes such as ‘Introduction session’, ‘Waiting locations’, or ‘Graphical interface’. And then a second round of coding allowed the identification of recurring themes using pattern coding (Saldaña 2021). The resulting codes are more conceptual, with themes such as ‘Resistances’, ‘Beliefs’, or ‘Negotiations’. On this basis, a detailed account of the unfolding of a transaction was reconstituted, as recounted by all the parties involved.
The setting
Placing an order on a food-delivery app is an unusual economic transaction. One’s engagement in a market for food is easily perceived: one chooses a meal from a list, compares prices, and ultimately concludes the transaction. However, the sense of being engaged in a labour market is much less clear. No list of couriers is presented, nor the pay they will receive upon completion of the job. Gig economy platforms seem to be operating a largely automated marketplace. An algorithm takes care of matching supply and demand, and a formula is applied that determines a price for each job. These platforms are confident enough in the data they hold about the market to handle the matching and the pricing themselves.
Food delivery platforms organize what’s called a multi-sided market (Rochet and Tirole 2003). For this reason, they are sometimes referred to as marketplaces and distinguished from hierarchical organizations. While traditional delivery companies only had to enrol one side of the market, consumers, the gig economy platforms have to attract both sides, consumers and workers (see Figure \ref{Triangle]). On the consumer side of the transaction platforms use traditional marketing tools to entice them. The worker side is where the novelty lies, as platforms have to use the tools of algorithmic management in order to secure their consent (Lee et al. 2015; Perrig 2021). Algorithmic management refers to a sophisticated array of incentive devices that aim both to put couriers to work and serve as cognitive devices while on the job (Richardson 2020; Shapiro 2018). Those platforms thus differ from corporations in the sense that the workers’ consent is not legally enforced. Instead, platforms must make an offer for a job that couriers can reject. In this sense, for every incoming order the incentive devices are put to the test as the transaction is always at risk of being jeopardized by couriers rejecting the job.
This conception of algorithmic management is novel in that it takes the managers’ view and follows them in solving their own problem: securing transactions. This will prove a fruitful way of identifying the limitations of algorithmic management, as transactions indeed sometimes fail. Platforms sometimes run sophisticated calculations in order to find the right courier and set the right price, only to be rejected by the courier. In this endeavour, the matching and pricing mechanisms are essential. We are thus following Fourcade and Healy (2017) and attempting to ‘see like a market’. Such a view will force us to consider the burdens of algorithmic management, which will be conceived as a (too) ambitious attempt at automating a marketplace.
The matching algorithm
Building the optimal matching algorithm is difficult. One cannot simply assume that couriers are homogeneously spread throughout the city, waiting for an algorithm to send them the next ride. Instead, they constantly adapt their location and practices depending on their beliefs about the algorithm (Sun 2019). Platform managers thus attempt not only to optimize the allocation of rides, but also to provide the right incentives. The following paragraphs show how platform managers adjust their algorithm in order to best respond to these unruly couriers. The ultimate allocation of rides indeed can be understood only when this recursive relationship between couriers and algorithm is accounted for. The analysis should thus follow couriers in the building of their algorithmic imaginary (Bucher 2017).
The information
The journey starts with an information session. For couriers to start working, they need to identify themselves and pick up their gear at the company’s offices. This is a rare encounter with managers and an important moment for our purposes because the platforms will take advantage of this encounter to communicate important information before couriers go to work. Of course this includes information about the functioning of the app or platform. This moment is best recounted by a revealing exchange I witnessed at such a meeting between a prospective courier and a platform manager. It took place during a Q&A right after the employee’s presentation of the app. A dozen prospective couriers were present, including myself.
\begin{quote] - \textit{Courier:] Is distance [from the restaurant] the only criterion for attributing the rides?''\\ - \textit{Employee:]Yes.’‘\ - \textit{Courier:] Because I previously worked [for the same platform] in France and [the managers there] told me that if we stayed in front of a restaurant we wouldn't obtain the ride.''\\ - \textit{Employee:]Ok yes it actually is a bit more complicated. Several restaurant owners complained that there were sometimes up to twenty couriers waiting in front of their restaurants. So we tweaked the system and now in order to receive the ride you have to be close to the restaurant, but not standing in front of it for too long. So it’s really better to be constantly on the move.’’\footnote{Translated from French by the author.] \end{quote]
Several comments are in order here. First, one can sense that an understanding of the algorithm is an important skill on the job, as this experienced courier inquired specifically about it. This session is an important moment for couriers, as they have an exclusive opportunity to ask questions about the details of the job.
Second, and most obvious, the platform employee is not providing couriers with a full set of information. This is something that was observed of all the platforms studied. At this stage of their career, most couriers are still not aware of the subtleties of the job, and the platform provides them with an overly simple explanation of the matching mechanism. Why so? Was it to make the employee’s job easier and avoid getting into the complicated details of algorithmic matching? Is it explicit policy on the part of the company that employees should hide such information from couriers? We shall never know. Suffice it to say for our purposes that couriers get started with incomplete information.
Third, this exchange reveals that platforms are constantly revising their algorithms and responding to couriers’ reactions. When describing the evolution of their matching algorithm, platform managers would systematically recount how their initial formula started with a rudimentary rule, typically ‘send the ride to the courier closest to the restaurant’ before being improved. This is typical of start-ups in the platform economy, launching rapidly in order to reap the rents of a winner-takes-all market, and refining their technology on the fly (Bialski 2024; Luca and Bazerman 2020). This evolution is telling because the transition from ‘closest to the restaurants’ to ‘close but on the move’ has wide repercussions in terms of working conditions. At first, the instructions were clear and couriers knew the rules of the game. The update, however, brought radical uncertainty to their working conditions. What does ‘being constantly on the move’ mean? Couriers have to work with obscure instructions, and the platforms deliberately keep them in the dark, as shown by the following email, sent to all couriers a few weeks later:
\begin{quote] Monthly Q&A\ This month we address a question that many couriers have asked: ``How are rides attributed?’’\ Our reply: the algorithm attributes the ride to the courier that is able to get to the restaurant the fastest as soon as the order is ready. For this reason it is often optimal to be on the move while waiting for your order! \end{quote]
To say that the message is unclear is an understatement, and as we will see below couriers quickly learn to be sceptical of the platforms’ communications and attempt to organize in order to develop their own understanding (Ticona and Tsapatsaris 2023). For the moment, let us note that couriers start their job with little information about the functioning of the algorithm, because of the secrecy of platforms, but also because of its opacity (Burrell 2016). Let us now study how they encounter and inquire about it.
The encounter
As Bucher (2017) recounts in her study of Facebook users, algorithmic literacy starts with a first encounter. This experience then shapes the imaginary built around the algorithm. What does this encounter look like in our case? As we have seen, couriers start their job with little information. Most cannot even tell whether the matching is human or automated, and at this stage they do not think about it in that level of detail. The very first shift usually goes like this: the courier puts on comfortable clothes, leaves home and usually finds a comfortable bench, maybe in a park, to wait for their orders. Time passes, sometimes up to an hour, before they are sent a first ride. This may happen a couple of times, but usually by then the courier understands that something must be done about this. Rides are scarce, competition among couriers is fierce, and somehow they must try to figure out how to be more visible to this matching algorithm. Thus begins the realization that an algorithm is governing their work.
The first action they take will depend on the information they were given by the platform. Some will park in front of a popular restaurant and wait there. One courier succinctly recounted an experience shared by most of them. Here is an excerpt from my field notes:
\begin{quote] Fresh out of the information session, he follows the instructions and looks for the right spot to wait for his rides. He was told they were sent to the person closest to the restaurant. He chooses a central restaurant. It is his favoured choice as he knows the fast food is in high demand and the place has seats on its terrace, where he will be able to wait comfortably. While waiting, however, he notices colleagues coming from afar to pick up orders in this restaurant. Confronted with this new information, he ultimately revises his beliefs and starts looking for other clues that could help him understand this matching mechanism. His inquiry leads him to one certainty: ‘I tried analysing the rides I received, but it doesn’t depend on our location, as they told us.’ \end{quote]
This experience is common to many other couriers, who recount it in very similar terms. Mohammed, for example, puts it this way:
\begin{quote] What’s funny is that there are couriers who are coming from who knows where and take the order you were expecting because you were waiting in front of the restaurant for so long… [Laughs] It doesn’t compute! At first they said ‘we send the rides to the closest courier’. Now they say ‘we prioritize the fastest courier.’ I don’t know but I think they give the orders to their favourite couriers. \end{quote]
These first experiences of the algorithm force couriers to realize that they cannot rely on communications from the platforms. Instead, they have to develop their own understanding. Mohammed, for example, concludes by suggesting that some couriers are favoured in some way he cannot fathom. But what exactly does he mean? Is he suggesting some form of human intervention? The next section will recount how couriers try to find out the details of this algorithm.
The inquiry
A mixture of financial pressure and uncertainty regarding the functioning of the platform leads couriers to investigate how the algorithm works. They put more or less effort into the task, and most will not frame such efforts as tantamount to an investigation. However, they will casually adjust their behaviours and chat with colleagues about the distribution of rides. Two main sources of information can be identified in this endeavour: experiments and gossip.
As we have seen, couriers rapidly come to understand that distance from the restaurant is not the only variable in play. To refine their understanding of the algorithm, they pay close attention to the context in which they receive rides and engage in a form of algorithmic auditing, whose purpose is to unveil what variables are governing their working conditions (Buolamwini 2022). Some will conclude that delivery speed is a prominent factor, others their acceptance rate, and still others will bank on good customer ratings. They build these conclusions on differing experiences and sometimes very formal experiments, as shown by the following testimonies.
\begin{quote] Jorge and his three friends usually sit at a table in front of the same restaurant to wait for their rides. They do not wish to compete against each other and in order to share rides equally among them they tried aligning their smartphones on the table, one next to the other, aligned in the direction of the restaurant. Following their reasoning, the owner of the smartphone closest to the restaurant would receive the next ride. They gave up after a few tries because it didn’t work. \end{quote]
This experiment might seem naive, but it proved to be a quick and easy way to ascertain the validity of a claim. It is also telling in that it shows that couriers have a strong sense of justice that they wish to assert. The design of the algorithm and the limitations of the technology, however, do not allow them to affect the allocation as they might wish. Another group went to great lengths in order to test the algorithm.
\begin{quote] Taher works with three other friends. Together they quickly realized that it was not necessarily the closest one to the restaurant who would receive the next order. Taher is the quickest, with an average delivery time of 24 minutes, but they noticed he was not receiving more rides than the others. Realizing this, they give up their investigation and rely on finding a good spot where the number of rides they receive is satisfying. \end{quote]
As we will discuss, this experience is representative of a whole group of couriers who conduct an investigation for some time, only to give up and resign themselves to a merely satisficing solution.
Another important source of information among couriers is discussion with their peers. By cross-referencing their experiences, they can quickly reach a better understanding of the algorithm. This idea of an algorithmic gossip has been already addressed in the literature, mostly in the case of creative platforms, social media or search engines \citep[see][]{bishop2019; Cotter (2019); MacDonald (2021); Oeldorf-Hirsch and Neubaum (2023)]. Youtubers or Instagram influencers will often conspire in order to build a better understanding of the algorithm and improve their visibility. These strategies can be applied similarly to the workers of the gig economy (Sutherland et al. 2020; Ticona and Tsapatsaris 2023; Watkins 2022). The couriers will share tips and argue about their efficacy, most fiercely on messaging apps (Bonini et al. 2023). Here is an instance of this, prompted by an innocent question on a messaging group.
\begin{quote] \textit{–C1:] What do you do when you don’t receive rides?\ \textit{–C2:] Try disconnecting and reconnecting. Also, where are you located?\ \textit{–C3:] You shouldn’t wait home but close to the restaurants.\ \textit{–C1:] I’m waiting 200m from HolyCow.\ \textit{–C4:] I’m home and receiving rides lol\ \textit{–C1:] Does being on the move help?\ \textit{–C2:] Yes, it does help. \end{quote]
The couriers adopt different strategies in pursuit of their inquiry. When choosing the right location, some will for instance choose to always be on the move, others will settle to a convenient restaurant, and still others will give up, disillusioned and stay home in-between rides. What does that tell us ultimately about the allocation of rides? As expected, we cannot tell who is the right. However, we can say that the couriers’ reaction to the algorithm affects the allocation. A platform thus cannot expect couriers to receive their rides passively. Instead, they have to expect some kind of reaction and constantly revise their algorithm accordingly.
Data-driven price discovery
Once a ride has been sent to the right courier, the platform is only halfway to securing the transaction, as the courier now has to accept it. The payment they receive for the ride is crucial here, as it will be an important motivator. As independent contractors, couriers have no obligation to take the ride, and a contract is concluded only once they accept it. The moment they receive a ride offer, they are provided with selected information about the ride, including payment, and they must decide whether to accept it (see offers in Figure \ref{Offers]). Most of the time, couriers are under financial pressure. Food delivery is a notoriously precarious occupation, which relies on cheap migrant labour (Van Doorn and Vijay 2022). The couriers thus readily accept incoming orders without giving too much thought to the details of the job (Ivanova et al. 2018). During interviews, they would often say they accept all incoming rides, but would later recount how ‘this one time’ they did not. Because of bad weather, because they didn’t feel like going to a particular neighbourhood, or simply because they had just stumbled upon a friend, they would deem some rides undervalued and reject them. While these instances did not seem like much to the couriers, such rejections were a major concern for platform managers. A rejection is considered a failure of the automated valuation model, on the basis that the price does not adequately reflect the value of the ride. How does such a failure arise? In order to better understand what is at stake in this remote and automated valuation, we will need to study couriers’ preferences and how managers attempt to integrate them into the pricing formula.
In their quest for the right pricing formula, platform managers proceed by trial and error. As with the matching algorithm, the pricing formula is constantly improved. One can distinguish three stages in building a pricing scheme, which recount in chronological order how platform managers finetune the pricing formula.
It starts fairly naively. When launching their apps, platforms typically have little regard for this matter. They start by offering couriers a fixed price for each delivery. Every ride thus pays the same amount irrespective of its distinctive features. Of course, such a scheme could not work efficiently, as couriers would reject the more tedious rides, those that require longer travel for instance.
In a second stage platform managers start to improve the formula and thus engage in a form of data-driven price discovery. In order to find the ideal market-clearing price, they try to infer couriers’ preferences from the history of their previous rides. One feature is perhaps the most obvious: a reward for distance. The longer the ride, the less appealing. It should come as no surprise that it was the first feature managers identified, and the one that is included in the pricing scheme of all platforms. It is moreover not technically difficult to implement using the Google Routes API. Similarly, platforms face the problem of couriers not logging on in bad weather. Although most couriers are not bothered by light rain, they might well be put off by heavier rain or snow, let alone a storm or hail, when they might not even connect to the app. Furthermore, such circumstances tend to coincide with periods of high demand for deliveries, as people stay home and tend to want their meals to be delivered. Platforms thus include a measure of the weather using a weather forecast API and grant bonuses for rides accepted when weather conditions are bad. Again, this proves a powerful incentive and more couriers are thus willing to work.
A third stage comes with the fine tuning. Including compensation for distance and weather is good, but a significant number of rides are still rejected. Those that require climbing a hill for instance are still particularly loathed by couriers, and so are those that require carrying bulky loads, such as heavy beverages or large pizzas. The problem of elevation is easily solved, as the Google Routes API also offers the possibility of including it when asking about a distance. However, it entails an additional cost and platforms prefer to do without it. Measuring the bulkiness of every delivery is also theoretically possible, but implementing such rewards would require ‘measuring the volume or weight of every item on sale, which would be too complicated’, as a manager would conclude. Indeed, it would require that each restaurant provide the weight of each item on menus, which sometimes change. Moreover, relying on weight alone is insufficient to take account of volume or size or perhaps of special care needed for certain items. Because gathering and updating this data would be too time consuming, the bulkiness or heaviness of an order is never reflected in the compensation.
This has consequences for couriers’ ultimate decisions, as many find ways to identify undervalued rides. As can be seen in Figure 2, some platforms display information about the offered ride, including details of the meal to be delivered. Many couriers will then be careful to avoid accepting certain deliveries that they expect to be cumbersome. Other platforms are more cautious and retain much of the information about the ride, preventing couriers from evaluating the offer effectively. One manager interviewed on this subject tellingly declared ‘the less they know the better’.
Through this process of datafied price-discovery, platforms end up with a ‘good enough’ pricing algorithm (Bialski 2024). Whether because of measurement difficulties, unavailable data or cost considerations, it can only roughly estimate couriers’ propensity to work. As a last resort, platforms set up additional, broader incentive schemes aimed at gamifying their activities (Van Doorn and Badger 2021). In the price formula, for example, some extra money can be awarded when a certain number of rides are accomplished per week or day, in the hope that it will provide some incentive for couriers to accept incoming orders without being too picky.
Discussion
In this paper, we have studied the difficulties that platform managers face in their attempts at automating a marketplace. We could summarize them as emerging from two sources. The first is the fact that couriers’ behaviour is not easily anticipated. They are not passive, but instead try to develop strategies in order to make themselves more visible to the matching algorithm. They also have heterogeneous preferences when it comes to valuing a ride, which makes it difficult for platforms to set a price accurately. The second source of problems for the platforms is missing or unreliable data on the couriers’ activities on the streets. Far from being a standard task that could easily be captured in numbers, the work of couriers is not readily datafied. The value of a ride cannot be limited to a couple of features, such as distance and weather. While the amount of data that platforms gather about their workforce is worrisome from a privacy standpoint, it isn’t sufficient to fully automate the marketplace, despite their claim to the contrary. Moreover, much of the data they use is declared either by the couriers themselves, by restaurant owners, or provided by third-party APIs, which makes it costly and often unreliable or inappropriate. The result is that the compensation offered to couriers for rides often does not reflect their preferences, leading to rejections.
Despite these shortcomings, platforms are nonetheless successful in securing most of their incoming orders. One is forced to acknowledge that this wobbly economic \textit{agencement] somehow functions. How does it work? And what are platforms doing, if not organizing a marketplace? This paper concludes by suggesting that platforms are \textit{s(t)imulating] a marketplace. This portmanteau word conveys the idea that platforms are operating a novel economic \textit{agencement], a form of market-like mechanism which serves both to legitimise their position and optimise the enrolment of workers in each transaction.
First, food delivery platforms are not building marketplaces, instead they are \textit{simulating] them. While the idea of a marketplace is constantly used as a referent in their communications,\footnote{It is enough here to recall the words of Travis Kalanick, CEO of Uber: ‘[the platform] is not setting the price. The market is setting the price. […] We have algorithms to determine what that market is.’ Cited in \citet[p. 101]{lehdonvirta2022].] by automating the valuation of rides they instead deny market participants of a fundamental requirement of markets, which is that of evaluating the offer (J. Aspers Patrick;Beckert 2011; Callon 2021). More than simply providing a market infrastructure, they end up organizing a form of ‘centrally planned market’ (Lehdonvirta 2022), a form of economic \textit{agencement] that is neither market nor organization. Maintaining the appearance of a market is a matter of survival for platforms, a referent which allows them to legitimise their reliance on independent contractors (Pujadas and Curto-Millet 2019; Viljoen et al. 2021). In practice, however, we have seen that the algorithmic valuation of rides is clunky and neatly provides an instance of \textit{market design failure] (Rilinger 2023).
Second, platforms do not simply let transactions occur, instead they \textit{stimulate] them. Indeed, this is a case in which market design failure does not result in the market collapsing. Instead, transactions do occur. That is because this apparent fragility is compensated by their situation of intermediary, which grants them significant power (Bessy and Chauvin 2013; Gillespie 2010). Our research shows that active involvement in the unfolding of a transaction is indeed necessary for gig economy platforms to ensure that the market even holds together. On one hand, the matching algorithm cannot be entirely transparent, otherwise it could lead couriers to engage in anti-competitive gaming. Hence, platforms must obfuscate their inner workings, and constantly update their matching algorithm (Burrell 2016). On the other hand, the pricing formula is not sufficient to ensure the enrolment of workers in the transaction, as it only partly reflects their preferences. Hence, platforms must refrain from disclosing some information about rides, and implement gamification schemes in order to push them to work (Van Doorn and Badger 2021; Vasudevan and Chan 2022).
Moreover, this article contributes to a better understanding of platform control. Indeed, as Rilinger (2024) similarly suggests, conceiving of gig economy platforms through the lens of market design helps to reveal the numerous gaps and limitations in their attempts at simulating market-like trading conditions. Control is thus the logical recourse that platforms have to make up for these shortcomings. Indeed, their failure at designing an automated marketplace can be considered an important reason why they obfuscate information to workers (Rahman et al. 2023), as well as for platforms’ numerous anti-competitive practices and outright misdeeds (Grohmann et al. 2022; Isaac 2019; Shapiro 2023). Such abuses of power should indeed be considered an inevitable consequence of their business model, an indispensable complement of the illusory goal of market automation.
References
Citation
@article{perrig2023,
author = {Perrig, Luca},
publisher = {Cambridge University Press},
title = {Confronting Algorithmic Management Using Subject Access
Requests},
journal = {Economic and Labour Relations Review},
volume = {34},
number = {4},
date = {2023-12-12},
url = {https://www.lucaperrig.ch/Research/Perrig2023.html},
doi = {10.1017/elr.2023.55},
langid = {en},
abstract = {Gig work is defined by a (mostly) automated management,
operating remotely through an app. Without human interaction,
workers are left with only guesses about the functioning of the
algorithms they are subjected to. To better position themselves in
their competition for tasks, they try to influence the data profile
that platform build about them. Made of performance indicators,
personal information, and sensor data, these profiles are an
essential part of algorithmic management. This paper will identify
data profiles are core sites of the struggle in the gig economy. It
will discuss the benefits and limitations of bringing data at the
center stage through a workers’ inquiry of food-delivery platforms.
The analysis will distinguish three actors in this inquiry and
discuss their uses of data profiles : the couriers themselves, as
they attempt to make sense of algorithmic management; the
researcher, and how they can use personal data inorder to
reconstitute this field of struggle; and the trade unions, which can
provide a way to collectivise data governance and gain better
information in building their case in favour of gig workers’
rights.}
}