How to Audit Your Pricing Algorithm for Antitrust Risk
Published 2026-09-19. Last updated 2026-09-19. Sources are listed and linked at the end of this page.
Audit the inputs before the model. What decides most algorithmic pricing matters is not whether the model is clever, it is whether a competitor's non-public price, cost or capacity figure reached it, directly or through a vendor serving your rivals. Before anyone reads code, freeze four records: every model version with its deployment window, the lineage of every training and runtime input, the parameter set in force on each date, and the human override log. The United States has told a federal court that "Sharing confidential pricing information with a common pricing agent can be equivalent to sharing that information directly with a competitor." So the input inventory is the first deliverable and the model review is the second.
This page is information about compliance practice. It is not legal advice and it does not create a lawyer client relationship.
Freeze four records before you start
The Antitrust Division's *Evaluation of Corporate Compliance Programs in Criminal Antitrust Investigations*, November 2024, puts two questions to prosecutors: "Does the company have clear document retention guidelines and does it educate employees on the ramifications of document destruction and obstruction of justice?" and "Are compliance personnel involved in the deployment of AI and other technologies to assess the risks they may pose?" An audit that starts after the model version was overwritten examines something that no longer exists.
The UK Competition and Markets Authority puts the documentation burden on the firm rather than the investigator: "in preparation for potential regulatory intervention, we suggest it is incumbent upon companies to keep records explaining their algorithmic systems, including ensuring that more complex algorithms are explainable."
| Record to preserve | What it has to show | Why it is asked for |
|---|---|---|
| Model version register | Every build, its tag, the dates it was live in each market, who approved deployment | Identifies the system that set a price named in a complaint |
| Training data lineage | Source of every field, its contract, the period it covers, whether that supplier serves competitors | Separates your own and public data from rival non-public data |
| Parameters in force by date | Objective function, floors, elasticity bands, competitor-derived features, geographic granularity | A rule that blocks price decreases is a design choice, readable later |
| Human override log | Who changed a recommended price, by how much, why, who approved | Shows whether the model was advisory or decisive |
| Vendor and forum record | Contracts, data schedules, certifications, attendance at vendor-hosted meetings | Vendor-hosted groups of competing customers feature in enforcement |
| Pricing committee material | Agendas, minutes, decision memoranda | Contemporaneous reasons are the only surviving evidence of intent |
Retention should run at least as long as the limitation period in your markets, and the register should be append only, because a registry that can be edited in place is an assertion rather than evidence.
The input that creates the sharpest exposure
A rival's non-public price changes the legal character of a pricing system. The European Commission's 2023 horizontal guidelines state that information may be exchanged "indirectly, by or through a third party (such as a service provider, platform, online tool or algorithm), via a common agency (for example, a trade association)", and a footnote settles the scope: "It includes data pools, where data holders group together to share data." A feed no employee reads is still an exchange.
The inventory needs separate lines for a direct contractual feed from a competitor, a vendor product that co-mingles your data with rival data, and a pool where reciprocal contribution buys access. In Agri Stats the Antitrust Division described a tell a pricing team can apply without a lawyer: "When companies decide certain information is too sensitive to share with the broader market, but not too sensitive to share with their closest competitors, that is a significant red flag that competition is being harmed".
Age is the most practical control, because it turns a legal test into a field in a data dictionary. The Commission gives the principle: "information can be considered historical if it is several times older than the average length of the pricing cycles or the average lengths of the contracts in the industry". The proposed RealPage consent judgment gives a number: the software provider would "Cease using active lease data for purposes of training the models underlying the software, limiting model training to historic or backward-looking nonpublic data that has been aged for at least 12 months", and would not "use models that determine geographic effects narrower than at a state level". Neither is a safe harbour for your market, and both are testable against your feature store.
Passive recipients are exposed too: an undertaking that receives commercially sensitive information from a competitor "will be presumed to take account of such information and to adapt its market conduct accordingly, unless it publicly distances itself". Disclosure in the other direction is treated the same way, since "Unilateral disclosure can occur, for example, via (chat) messages, emails, phone calls, input in a shared algorithmic tool, meetings, etc." Uploading your own numbers into a shared tool is an act, not a transfer.
Hub and spoke: one vendor, many competitors
When several competitors buy the same optimisation product, the vendor becomes a point through which sensitive information moves without anyone speaking to a rival. The United States answered the argument that this differs from direct contact: "Automating an anticompetitive scheme does not make it less anticompetitive." It also answered the defence most often raised in a pricing committee, that recommendations are only advisory: "It is per se illegal for competing landlords to jointly delegate key aspects of their pricing to a common algorithm, even if the landlords retain some authority to deviate from the algorithm's recommendations."
Overrides are evidence worth keeping and are not a defence on their own. Treat a low override rate as a fact to explain, not a result to report.
Settlement terms show what vendor diligence produces. The proposed Pinnacle consent decree would require the landlord to "Refrain from using any anticompetitive algorithm that generates pricing recommendations using its competitors' competitively sensitive data or that incorporates certain anticompetitive features", and to "Accept a court-appointed monitor if it uses a third-party pricing algorithm that is not certified pursuant to the terms of the consent decree". If your vendor will not put in writing which competitors contribute data, at what granularity and lag, and whether that moves your recommendation, the silence is your finding.
The CMA reached the same place in March 2026: "If you can reasonably expect that a pricing recommendation could be drawing on confidential information from a competitor (even if you have not been told this directly), you may still be breaking the law." Its instruction to buyers is plain: "Make sure that any pricing guidance or actions, generated by a pricing solution you use, are not influenced in any way by competitively sensitive information from rivals".
What to test for internally
The Antitrust Division puts two more questions to a compliance function: "Does the company use any type of screen, communications monitoring tool, or statistical testing designed to identify potential antitrust violations?" and "Does the compliance program monitor and detect decision-making by AI or other technology tools to ensure they are not violating antitrust laws?"
- Input census. Enumerate every field the model consumes at training and runtime, reconciled against the feature store rather than the design document. A field whose source cannot be named in a sentence is a finding.
- Provenance trace. Record supplier, contract, collection method, and whether that supplier serves named competitors. Web collection needs a dated capture record, since "public" is a claim about a moment.
- Reciprocity check. Flag every feed where access is conditioned on your own contribution.
- Re-identification test. Try to recover a single rival's price from an aggregate feature. If a rank, an index or a small-cell average allows it, the aggregation is cosmetic.
- Aging test. Compute the age distribution of each competitive input against your pricing cycle length.
- Granularity test. Record the narrowest geography, product and customer segment at which competitive data moves an output.
- Constraint read. List every rule limiting downward movement, from floors to recommendation bands, and capture each justification, dated when it was set.
- Counterfactual run. Re-run a representative period with competitor-sourced features removed and measure the price delta.
- Override analysis. Measure override frequency, direction, size and reason codes.
- Convergence screen. Test published prices against available rival series for synchronised movement, narrowing dispersion, weak response to your own cost shocks.
- Shared-vendor exposure. Estimate the share of the relevant market served by the same vendor model.
- Forum review. Cover vendor user groups, benchmarking panels and trade committees where pricing came up, plus the channels pricing staff actually message on.
For language models in pricing work the CMA adds that "In some cases, you may need to audit the input data and statistical approaches used, whether in-house or third-party. A traditional algorithmic audit might not be enough", suggesting stress testing prompts and explicit anti-collusion constraints.
Who should run the audit, and why privilege decides it
An internal audit produces candid written findings on the most sensitive subject in the company. In *Upjohn Co. v. United States* the Supreme Court protected communications with counsel made to secure legal advice, while limiting what that reaches: "The privilege only protects disclosure of communications; it does not protect disclosure of the underlying facts by those who communicated with the attorney". Log files, the feature store and the model registry are facts. The memorandum analysing them may be protected if it was created for legal advice, which means counsel commissions and directs the audit and engages the experts.
For any group with European exposure the position changes. In *Akzo Nobel Chemicals Ltd and Akcros Chemicals Ltd v European Commission* the Court of Justice held that "the requirement of independence means that there should exist no employment relationship between the lawyer and his client, so that legal professional privilege does not cover exchanges within a company or group with in-house lawyers". An audit memorandum written by your in-house competition lawyer can be read by the Commission during an inspection.
Scope the engagement in writing before data is pulled, keep factual artefacts separate from legal analysis, and decide who receives the findings in advance, because wide internal circulation is the most common way privilege is lost by accident.
What a defensible audit trail looks like afterwards
A defensible trail is contemporaneous and versioned, and it closes every finding out. It records what was tested, on which model version, against which data snapshot, on what date, by whom, and what became of every finding. Remediation entries carry an owner and a completion date, and a finding accepted rather than fixed carries a reason written at the time by someone with authority to accept it.
The CMA is blunt about the shelf life of this work: "Algorithmic systems are often updated regularly with new or evolving datasets and dynamic models, meaning that one-off audits may become quickly outdated." Bind the audit to the release process, so a new competitive feature or a changed objective function triggers a review rather than waiting for the annual cycle.
What an audit cannot tell you
An audit cannot establish the absence of collusion. It can establish the absence of specific detectable markers, in the data and documents examined, at a point in time, for the model versions in scope. Coordination reached outside the pricing stack never appears in a feature lineage report, and a model trained on data that already carries coordinated prices can reproduce the pattern with no prohibited input present today. Screens produce candidates, not verdicts.
State two limits to the board before the engagement letter is signed: a clean audit does not immunise conduct, and a failed test is not a finding of wrongdoing. The CMA reminds firms in the United Kingdom that businesses can be "fined up to 10% of their annual turnover" and that it offers "a reward of up to £250,000 to anyone who tells the CMA about illegal cartel activity, including algorithmic collusion", while its updated leniency policy "makes leniency available for conduct such as exchanging competitively sensitive information through a shared algorithm".
Frequently asked questions
Does using a pricing algorithm violate antitrust law?
Not by itself. The United States has told courts that "not every use of an algorithm to set price qualifies as a per se violation of Section 1 of the Sherman Act", but that it is per se unlawful where "competitors knowingly combine their sensitive, nonpublic pricing and supply information in an algorithm that they rely upon in making pricing decisions, with the knowledge and expectation that other competitors will do the same." Risk attaches to shared non-public inputs and joint delegation, not to automation.
Who is liable if the algorithm fixes prices?
The firm. The Commission's guidelines state that "firms involved in illegal pricing practices cannot avoid liability on the ground that their prices were determined by algorithms", comparing an algorithm to an employee or consultant under the firm's control. The vendor is also exposed: the CMA warns a pricing service provider "could be held to account for breaking competition law".
What is a hub and spoke agreement in antitrust?
Competitors pass information or pricing decisions through a common third party instead of dealing with each other. The CMA describes businesses using "the same algorithm or data hub to exchange competitively sensitive information indirectly", and the United States has argued that sharing with a common pricing agent "can be equivalent to sharing that information directly with a competitor."
Who should conduct an internal antitrust audit?
Outside counsel, with technical specialists engaged under counsel. The reason is evidentiary: in the European Union an in-house lawyer's audit memorandum falls outside legal professional privilege under the Akzo Nobel judgment.
Does attorney client privilege cover an internal audit report?
It may cover the report as a communication made for legal advice, but not the facts: "The privilege only protects disclosure of communications; it does not protect disclosure of the underlying facts". Data, logs and model artefacts stay producible whoever analyses them.
Sources
- Justice Department Requires RealPage to End the Sharing of Competitively Sensitive Information and Alignment of Pricing Among Competitors, DOJ Office of Public Affairs, 24 November 2025. https://www.justice.gov/opa/pr/justice-department-requires-realpage-end-sharing-competitively-sensitive-information-and
- Justice Department Reaches Proposed Consent Decree with Pinnacle, One of America's Largest Landlords, to Resolve Information Sharing and Algorithmic Coordination Claims, DOJ Office of Public Affairs, 4 September 2026. https://www.justice.gov/opa/pr/justice-department-reaches-proposed-consent-decree-pinnacle-one-americas-largest-landlords
- Justice Department Requires Agri Stats to End Exchange of Competitively Sensitive Information Among Nation's Largest Meat Processors that Suppressed Competition and Increased Prices for Decades, DOJ Office of Public Affairs, 7 May 2026. https://www.justice.gov/opa/pr/justice-department-requires-agri-stats-end-exchange-competitively-sensitive-information
- Statement of Interest of the United States of America, filed in Duffy v. Yardi Systems, Inc., with attachments from In re RealPage, Inc., Rental Software Antitrust Litigation. https://www.ftc.gov/system/files/ftc_gov/pdf/YardiSOI-filed%28withattachments%29_0.pdf
- Evaluation of Corporate Compliance Programs in Criminal Antitrust Investigations, DOJ Antitrust Division, November 2024. https://www.justice.gov/atr/media/1376686/dl
- Guidelines on the applicability of Article 101 of the Treaty on the Functioning of the European Union to horizontal co-operation agreements, European Commission, C(2023) 3445 final, 1 June 2023. https://competition-policy.ec.europa.eu/system/files/2023-07/2023_revised_horizontal_guidelines_en.pdf
- Algorithms: How they can reduce competition and harm consumers, Competition and Markets Authority, GOV.UK. https://www.gov.uk/government/publications/algorithms-how-they-can-reduce-competition-and-harm-consumers/algorithms-how-they-can-reduce-competition-and-harm-consumers
- AI and collusion: frontiers, opportunities and challenges, Competition and Markets Authority blog, 4 March 2026. https://competitionandmarkets.blog.gov.uk/2026/03/04/ai-and-collusion-frontiers-opportunities-and-challenges/
- UPJOHN COMPANY et al., Petitioners, v. UNITED STATES et al., 449 U.S. 383 (1981), Legal Information Institute. https://www.law.cornell.edu/supremecourt/text/449/383
- Judgment of the Court (Grand Chamber) of 14 September 2010. Akzo Nobel Chemicals Ltd and Akcros Chemicals Ltd v European Commission. Case C-550/07 P, EUR-Lex. https://eur-lex.europa.eu/legal-content/EN/TXT/?uri=celex:62007CJ0550