AI Antitrust: Agencies Shift to Computational Screening

The old antitrust file room had boxes, coffee, and someone junior discovering the word privilege at 2 a.m. That picture is now out of date. The more useful mental model is a regulator with public procurement data, public price signals, uploaded documents, and enough tooling to ask patterns first and read memos second. This is not a claim that agencies have bought magical cartel detectors. McCann FitzGerald says competition authorities in the EU and beyond are using AI to gather market intelligence, increase efficiency, and improve antitrust enforcement. The practical point is narrower and more annoying: if conduct leaves structured traces, someone may screen those traces before your response letter is polished.

What computational antitrust actually means Isabella Lorenzoni’s Jean Monnet

Network working paper gives the least theatrical inventory of the new toolkit. It describes text mining systems based on natural language processing, document management software with machine learning pattern recognition features, and digital screening tools in public procurement. Translation: agencies are not only reading emails and submissions, they are building workflows that can cluster documents, detect recurring language, and flag procurement patterns. Your vendor contract does not need poetry here, it needs auditability, retention controls, and a defensible explanation of how pricing and bidding decisions are made. The Hellenic Competition Commission’s Computational Competition Law and Economics inception report frames the same development as part of a broader move toward screening tools and algorithms to detect collusive conduct. That matters because screening is a triage function, not a final judgment. A model can point an agency toward a tender, a market, or a set of firms that deserve human attention. The compliance risk is therefore not that an algorithm replaces due process; it is that weak records and messy communications make a company look more interesting than it needed to be.

The enforcement workflow is changing before

the law does McCann FitzGerald points to the European Commission’s tyre manufacturer investigation as a high profile example of AI-assisted antitrust enforcement. The Commission suspected that several tyre companies, including Michelin, may have used public communications to signal and coordinate future sales prices, particularly wholesale prices for replacement tyres for cars and trucks in the EEA. The legal theory is old enough to have opinions about fax machines. The evidence workflow is what has changed. For companies, this means public statements are not only communications risk, they are data risk. Pricing announcements, investor materials, website updates, tender submissions, and trade association material may be compared across time and counterparties. McCann FitzGerald also warns that preemptive monitoring and faster analysis of large information sets could increase market oversight, investigations, and dawn raids. That is not a reason to stop communicating; it is a reason to make sure public communications review includes competition counsel before the sentence about future prices becomes everyone’s problem.

The UK and EU are not waiting for perfect tools

Dentons says the European Commission and the UK Competition and Markets Authority are stepping up enforcement in AI, building on years of work on algorithms and digital markets. That should be read plainly. Agencies are looking both at how companies use algorithms and at how regulators can use computational methods to investigate them. Builders stuck between product velocity and legal review should assume both sides of that sentence are now live operational concerns. The policy explainer version is simple. If your product affects pricing, ranking, procurement, matching, allocation, or market access, document the human decision points and the data sources used. If your sales team participates in tenders, preserve bid rationale in a way that can be reconstructed without interpretive dance. If your legaltech product sells discovery, procurement analytics, or case search into regulated sectors, design for provenance, access controls, and review trails, because the buyer may be an authority or a company expecting authority-grade scrutiny.

Agencies also need governance, not just models Richard May’s Stanford Law paper

on generative AI use by competition authorities is useful because it does not pretend adoption is costless. It argues that authorities need a top-down governance strategy, risk guardrails, scaling plans, and cooperation with other authorities. That is the regulator-side version of the memo companies already receive from outside counsel. Models used for intelligence gathering still need scope limits, validation, human review, and records showing why a lead became an investigation. Stanford Law’s CodeX account of the Computational Antitrust Project at the OECD shows why this is becoming institutional rather than experimental. It says the project launched its fourth cross-agency report at the OECD in Paris on June 18, 2025, bringing together regulators, scholars, practitioners, and technologists to discuss whether antitrust law can absorb computational methods without losing legal guarantees. Separately, the Schrepel and Groza adoption report says the Stanford Computational Antitrust project team invited partnering agencies in the first quarter of 2022 to share advances in implementing computational tools. Network Law Review’s evidence from 25 antitrust agencies points in the same direction: computational antitrust is now an agency capacity question, not a conference panel curiosity. The next compliance deadline is not printed in an official journal. It arrives when a regulator can compare your bids, documents, public statements, and historic cases faster than your team can assemble the chronology. Companies do not need to panic, which is rarely a useful governance mode. They do need to treat competition compliance as data governance with legal consequences, because the people reading the file may now search, cluster, and screen it first.