For decades, competition law has been guided by a simple idea: Collusion requires contact. Firms need to communicate, coordinate, or at least signal to each other in some identifiable way. Agreements might be explicit or tacit – but they always assume interaction between the companies themselves. The Danish Dinero case suggests that this premise is starting to break down.
Visma Dinero, an accounting software provider, was preparing to launch an AI assistant marketed as a “virtual CFO”. The tool analysed a firm’s own data and benchmarked it against aggregated data from around 99,000 other companies using the platform. At first glance, this looked like a familiar product: Smarter benchmarking, better insights, incremental efficiency.
But the tool did more than aggregate data. It translated the data into concrete recommendations. Firms could see how their wages, margins, and costs compared to their industry and region – and, crucially, receive suggestions on how to adjust. If a company was pricing below the average, the AI might, e.g., recommend raising prices.
The Danish Competition Authority conducted a dawn raid and intervened already before the product was launched. It warned that the tool could give firms “excessively detailed” and near real-time insight into competitors’ cost and price levels, even if only in an aggregated form. As a consequence of the dawn raids, Dinero disabled certain features. The case ended without a formal decision, but not without a clear signal: This kind of system is edging into antitrust territory.
The tool as the infringement
What is striking about the Dinero case is not only what happened (dawn raid before product roll-out), but also what did not. Based on the regulator’s press release, there was no agreement between competitors, no evidence of communication, no classic exchange of competitively sensitive information. And yet the concerns of the Danish regulator were real enough to stop a product rollout. The issue lies in the structure of the tool.
The AI assistant could have created a continuous feedback loop. Firms would have contributed their own data; the system would have aggregated it across the user base and the output would have returned as benchmarks and recommendations. Each firm would have acted independently, but within a shared information environment that would have made the market increasingly transparent – and, potentially, increasingly predictable.
As such, the regulator was no longer only concerned with whether firms would coordinate. It was starting to look at whether they would be placed in a position where coordination becomes almost frictionless. It is also noteworthy that the data to be shared would have been aggregated to a certain extent. The case indicates, however, that even aggregated data can become problematic if they are sufficiently granular, timely, and actionable.
The concern is not disclosure in the classic sense. It is normalisation. If firms anchor their behaviour around shared benchmarks, especially when those benchmarks are dynamically updated and tied to strategic recommendations, the incentive to deviate shrinks. Markets do not need explicit agreements to stabilise; they may simply converge.
Check the system
Seen through that lens, Dinero begins to look less like an “AI case” but more like an example of a broader shift. The traditional questions, did firms agree or signal, start to lose their grip. The more relevant question becomes: What does the system enable? That shift matters because it moves antitrust away from conduct and towards architecture. The risk is no longer tied solely to what firms decide, but to the informational environment in which those decisions are made.
What makes the case also particularly significant is how the Danish regulator frames responsibility. It emphasises that companies themselves may be liable if they use tools that provide a benchmark for where their prices or costs “should be” relative to competitors. This suggests that exposure can arise not just from agreements or exchanges, but from reliance on systems that embed a form of coordination into their outputs.
That, in turn, raises difficult questions. How much does a firm need to understand about the underlying data and logic of an AI tool? When does benchmarking cross the line into coordination? And how granular, timely, or prescriptive can such tools become before competition law steps in? There are no clear answers yet, but the direction of travel is hard to miss.
Not a one-off
The Dinero case is unlikely to remain an isolated episode. The CMA in the UK also launched an investigation into suspected sharing of information among competing hotel chains using the same hotel data analytics tool.
As AI and data analytic tools become more sophisticated and more widely deployed, similar issues will arise across sectors – from pricing software in e-commerce to analytics platforms in professional services. In each case, the underlying question will be the same: When does data-driven optimisation begin to look like coordination?
The uneasy part is that the tools in question are not designed to collude. They are designed to optimise. But optimisation in a highly transparent, shared data environment may produce outcomes that are not so different from alignment.
Competition law has always drawn a line between independent adaptation and coordinated behaviour but this line might (have) become harder to locate. Not because companies are suddenly more prone to collude, but because technology is changing the conditions under which they compete. If that is right, the real challenge for antitrust is not detecting agreements. It is understanding systems.
Photo from Ant Rozetsky on Unsplash
Antitrustpolitics.com 