Ok, but is the price right?

Companies increasingly delegate their prices to algorithms trained on data on customer preferences. In some cases, these algorithms rely on a machine learning process to develop sophisticated pricing strategies that respond to both customer and competitor behavior.
Artificial intelligence algorithms are gradually coming to occupy a prominent place in the pricing methods of various companies. This is particularly true in more technological and web-operating companies, but more generally any sector where there is a large amount of detailed data available on demand and potential competition lends itself to this type of evolution. Fueled by this data, AI algorithms are designed to learn which pricing strategies will yield the greatest returns based on a trial-and-error learning process.
But what can be expected from the growing trend to delegate pricing decisions to algorithms? Empirical evidence on the impact of AI algorithms is still rather scarce. However, there are cases that have already received considerable attention. For example, in the German gasoline retail market, algorithmic pricing software has become widely available since around mid-2017. To understand what this means, we have analyzed detailed data regarding the phase before and the phase after the advent of algorithms. The surprising result was a steep increase in prices: the introductionn of algorithmic intelligence in pricing at German gas stations appears to have coincided with an increase in margins of up to 38%!

A problem with this type of study is that both the exact timing of the adoption of the algorithms by the various gas stations and the exact details about the type of algorithms used are necessarily unknown to outside observers. For this reason, in the study of the German gasoline market, the authors identified which petrol stations had adopted algorithmic pricing software and when they had done so through a statistical procedure.
This is an interesting idea because it takes advantage of what should be the peculiarities of an algorithmic pricing mechanism. In particular, compared to the prices established through human intervention, the prices established by the algorithms should differ in: the number of price changes made during the course of the day, the average size of price changes and the response time of the price update of a given station compared to a competitor's price change. Precisely these measures are the ones on which pricing software companies advertise the ability of their algorithms to generate positive impacts for distributors.
But how and why does AI induce higher prices? In the case of the German gasoline market, margins have been found to rise gradually, suggesting that algorithms need time to train themselves and converge towards tacitly collusive strategies in which AI determines less aggressive pricing than would be expected under normal market competition. That is, the algorithms learn pricing strategies that are more cooperative and lead to higher price levels. If expected profits for companies are larger, economic harm for consumers is just as clear.

This kind of result explains the vigorous legal and political debate on the collusive potential of AI applied to pricing that is currently underway. In particular, the potential of using algorithms as a means to facilitate explicit or tacit collusion has been a popular talking point among antitrust authorities, business organizations and competition law experts in recent years. Above all, there is strong skepticism about the ability of current competition law to address any tendency of AIs to induce super-competitive pricing as the law is designed to punish explicit agreements between companies.
There is still no general consensus on what AI will mean for pricing systems and what the ultimate consequences will be for companies and consumers. Given the lack of empirical evidence available, current research is mostly directed towards theoretical and experimental methods to evaluate the impacts of AI. The results are interesting but full of ambiguities. If some studies underline the collusion in prices deriving from learning reciprocal pricing strategies, others indicate on the contrary that AI, by allowing companies to better predict demand, pushes them towards more aggressive pricing strategies, thus lowering prices and making collusive behavior less likely.
Finally, we must not forget the fundamental role of data, real fuel in the engine of AI algorithms: whoever controls the data controls the type and effectiveness of the AI algorithms that have to work on this data. This very aspect is at the center of a recent study I conducted with Michele Rovigatti (Bocconi) and Gabriele Rovigatti (Bank of Italy). The results, presented for the first time a few days ago at a conference at Yale University, illustrate the problem of a digital platform (such as Google for example) which, through the type, quality and frequency of the data it releases to advertisers active in the auctions where it sells advertising space, it influences their choices in terms of which AI algorithms are used in the bidding process and with what effects on final prices.
So in conclusion, the jury is still out on how AI will impact pricing and consumer welfare.