Competition Law and Competition Governance of Autonomous Firms

1. Introduction

Autonomous firms are enterprises in which significant commercial decisions are made, recommended, executed, or continuously optimized by autonomous software systems, artificial intelligence (AI), machine-learning models, algorithmic agents, or interconnected automated systems, with limited real-time human intervention.

An autonomous firm may use AI to determine:

• prices and discounts;
• production levels;
• inventory;
• advertising expenditure;
• supplier selection;
• customer segmentation;
• credit and risk decisions;
• hiring and resource allocation;
• contractual terms;
• market entry or exit;
• responses to competitors;
• mergers or strategic investments.

Competition law traditionally assumes that a human-controlled undertaking decides how to compete. Autonomous firms challenge that assumption because the immediate decision-maker may be an algorithm rather than an individual executive.

The OECD has specifically identified both algorithmic collusion and algorithmic unilateral conduct, including exclusionary and exploitative conduct, as competition concerns. It also observes that algorithms can facilitate anticompetitive conduct at greater scale and speed.

The central legal principle is therefore:

Delegating a competitive decision to an autonomous system does not, by itself, remove the undertaking from competition-law responsibility.

2. Meaning of an Autonomous Firm

An autonomous firm can be understood through five levels of automation:

Level 1 – Decision-support firm

Humans make decisions after receiving AI recommendations.

Example: AI recommends a price of ₹999, but the manager approves it.

Level 2 – Human-supervised automation

The system automatically makes decisions within parameters established by humans.

Example: an algorithm automatically changes prices within a prescribed range.

Level 3 – Adaptive firm

The system learns from market conditions and modifies its own strategies.

Example: a pricing model learns that competitors respond aggressively to discounts and changes its pricing behaviour.

Level 4 – Highly autonomous firm

AI independently determines substantial commercial strategies.

Example: an AI agent determines prices, advertising allocation, inventory and supplier selection without prior approval.

Level 5 – Multi-agent autonomous firm

Different AI agents interact with suppliers, customers, competitors, platforms and other AI agents.

This produces a particularly difficult competition-law environment because competitive interaction may increasingly occur machine-to-machine rather than human-to-human.

3. Legal Status of Autonomous Firms

Competition law generally regulates the undertaking, not the individual human who physically presses the button.

Accordingly, the fact that an AI system:

• designed the strategy;
• selected the price;
• negotiated automatically;
• learned the behaviour;
• communicated through an API; or
• acted without immediate human approval

does not necessarily prevent liability of the company operating the system.

The critical questions become:

1. Who deployed the system?
2. Who established its objectives?
3. What data was supplied to it?
4. What constraints were imposed?
5. Could the company monitor its conduct?
6. Did the company know or reasonably foresee the competitive consequences?
7. Did the company continue using the system after discovering anticompetitive effects?
8. Was the system supplied by a common third-party provider to competing firms?

The OECD has noted that existing competition law may still apply to algorithmically implemented conduct, while the more difficult question concerns situations where independent algorithms reach collusive outcomes without an identifiable agreement.

4. Competition-Governance Framework

Competition governance of autonomous firms should operate at four levels.

A. Ex ante governance

Before deployment:

• competition-risk assessment;
• algorithmic impact assessment;
• testing for discriminatory outcomes;
• testing for coordinated pricing;
• identification of sensitive information;
• restrictions on competitor-data ingestion;
• documentation of objectives and parameters.

B. Continuous monitoring

During operation:

• price monitoring;
• market-share monitoring;
• competitor-response monitoring;
• detection of unusual parallel pricing;
• monitoring of exclusionary recommendations;
• logging of significant model changes.

C. Human accountability

There should be identifiable officers responsible for:

• algorithm deployment;
• competition compliance;
• data governance;
• model-risk management;
• incident reporting.

D. Post-incident remediation

Where an autonomous system produces potentially anticompetitive effects:

• suspend the relevant functionality;
• preserve logs;
• investigate the cause;
• notify appropriate authorities where legally required;
• modify the model;
• compensate affected parties where applicable;
• establish safeguards against recurrence.

5. Major Competition-Law Risks

I. Algorithmic Collusion

This is perhaps the most important problem.

Two competing firms may independently use autonomous pricing algorithms. The algorithms repeatedly observe each other's prices and may learn that aggressive price competition reduces profits.

They may consequently converge on stable high prices.

The difficult question is whether this constitutes:

• an agreement;
• a concerted practice;
• tacit coordination; or
• merely parallel independent conduct.

The OECD distinguishes coordinated algorithmic conduct from unilateral algorithmic conduct and recognizes that existing legal concepts can become difficult to apply where machine-learning systems independently reach coordinated outcomes.

Example

Firm A's AI observes Firm B's price.

Firm B's AI observes Firm A.

Both systems independently learn:

"If I reduce price, the competitor immediately reduces its price."

Eventually both systems maintain a high price.

There may be economic coordination without conventional human communication.

6. Hub-and-Spoke Algorithmic Coordination

A particularly significant model occurs where several competitors use the same third-party algorithm.

For example:

Competitor A → common pricing platform
Competitor B → common pricing platform
Competitor C → common pricing platform

The common algorithm may receive market information and determine prices for all three firms.

This can create a technological version of hub-and-spoke coordination.

The legal concern becomes greater if:

• competitors know that rivals use the same system;
• the provider facilitates exchange of competitively sensitive information;
• the system is designed to maintain market prices;
• firms deliberately adopt the system for coordination.

The OECD has specifically identified common third-party algorithms as a potential mechanism for hub-and-spoke coordination.

7. Algorithmic Price Fixing

Traditional price fixing does not become lawful merely because the agreement is executed by software.

Case 1 – United States v. David Topkins

This is one of the most important early algorithmic competition cases.

Online sellers of posters agreed to coordinate prices on Amazon Marketplace and used pricing software to implement the arrangement.

Topkins pleaded guilty to price-fixing.

The important principle is:

An algorithm can be the instrument through which an ordinary cartel is implemented.

The technology did not immunize the underlying agreement from antitrust liability. The U.S. Department of Justice classified the case as horizontal price fixing.

Significance for autonomous firms

An autonomous firm cannot argue:

"The computer fixed the price, not the company."

If the company deliberately configures or deploys the system to implement a cartel, ordinary competition-law principles remain relevant.

8. Case 2 – Eturas UAB v Lietuvos Respublikos konkurencijos taryba

Court of Justice of the European Union, Case C-74/14

Eturas operated an online travel-booking platform.

A software message effectively restricted discounts that travel agencies could offer through the platform.

The case concerned whether participating undertakings could be held responsible where the restrictive mechanism was implemented through the platform's electronic system.

Principle

Electronic communication and automated technological mechanisms can form part of the evidentiary framework for establishing coordinated conduct.

Autonomous-firm relevance

The case is important because competition law does not require coordination to occur through:

• meetings;
• letters;
• telephone calls; or
• traditional corporate communications.

Digital systems can constitute important evidence of coordination.

9. Case 3 – European Commission: Google Shopping

Google Search (Shopping), Case AT.39740

The European Commission found that Google had used its dominant position to favour its comparison-shopping service in search results.

Although this was not an "autonomous firm" case in the modern AI-agent sense, it is highly relevant to autonomous firms because the conduct involved algorithmic ranking and search mechanisms.

Competition concern

A vertically integrated firm may operate:

1. a platform;
2. the algorithm controlling access to consumers; and
3. a competing downstream service.

The platform can potentially use algorithmic ranking to disadvantage rivals.

Autonomous-firm relevance

If an autonomous system independently changes rankings in favour of the firm's own products, the legal issue does not disappear merely because a human did not manually rank each result.

The relevant questions become:

• Who designed the objective?
• What parameters governed ranking?
• What effects did the system produce?
• Was the firm dominant?
• Did the conduct foreclose equally efficient competitors?

10. Case 4 – Amazon Marketplace / E-Commerce Algorithmic Conduct

Amazon-related competition investigations and litigation illustrate another major issue: platform algorithms can simultaneously govern access to customers while serving the platform's own commercial interests.

The potential theories of harm include:

• self-preferencing;
• discriminatory ranking;
• use of seller data;
• exclusionary marketplace rules;
• tying;
• parity obligations;
• manipulation of visibility.

Autonomous firms intensify these concerns because an AI may continuously modify rankings, recommendations, commissions or access conditions.

Principle

The competition analysis should examine the economic effect of the algorithm, rather than simply asking whether a human employee manually made each decision.

11. Case 5 – ASUS Vertical Restraints

European Commission, AT.40465 – ASUS

The Commission examined restrictions concerning online pricing practices and found that ASUS had engaged in conduct involving monitoring and pressure concerning retailers' online prices.

The case is relevant to autonomous firms because online pricing tools and automated price-monitoring mechanisms can substantially increase the speed and scope of vertical price coordination.

The OECD identifies ASUS among examples where firms used online tools and pricing robots to monitor market prices.

Autonomous-firm lesson

A firm's use of automated monitoring does not change the underlying competition-law question:

Is the technology being used to improve legitimate market efficiency, or to restrict independent pricing decisions?

12. Case 6 – Online Hotel Booking / Booking.com Parity Cases

European competition authorities have examined price-parity clauses used in online hotel-booking markets.

These cases are important for autonomous firms because online platforms can automatically enforce contractual restrictions across thousands of transactions.

Competition concern

A platform may automatically prevent hotels from offering:

• lower prices elsewhere;
• different conditions through competing platforms; or
• better offers directly to consumers.

The autonomous enforcement mechanism can amplify the competitive effect.

Principle

Automation can transform a contractual restriction from an individual contractual practice into a market-wide technological constraint.

13. Case 7 – Apple Inc. v. Pepper

U.S. Supreme Court, 2019

This case concerned alleged anticompetitive conduct involving Apple's App Store and the distribution of apps.

Although not an autonomous-AI case, it is significant for autonomous firms because it demonstrates how a digital intermediary can occupy a powerful position between suppliers and consumers.

An autonomous platform may:

• control access;
• determine commissions;
• rank products;
• impose payment rules;
• collect transaction data;
• determine visibility.

Competition analysis therefore increasingly needs to consider the architecture of the platform, not merely conventional product pricing.

14. Case 8 – United States v. Apple Inc. – E-books

The Apple e-books litigation demonstrates the continuing relevance of traditional antitrust rules when conduct is implemented through sophisticated digital commercial structures.

The broader lesson for autonomous firms is that technological sophistication does not eliminate the need to establish whether firms have coordinated commercially sensitive behaviour.

An AI system may execute the coordination, but competition law can still focus on:

• the underlying arrangement;
• communications;
• incentives;
• implementation mechanisms;
• market effects.

15. Autonomous Exclusion

Not every competition problem involves collusion.

An autonomous firm may independently exclude competitors.

For example, an AI-powered marketplace may decide that:

"Products from rival suppliers should receive lower visibility."

If the firm is dominant, this may raise questions concerning:

• refusal to deal;
• discriminatory access;
• self-preferencing;
• margin squeeze;
• tying;
• leveraging;
• exclusionary ranking.

The OECD expressly identifies algorithmic exclusionary and exploitative conduct as a category of competition concern.

16. Autonomous Self-Preferencing

Consider a dominant platform operating:

• its marketplace;
• its logistics service;
• its payment service;
• its advertising system; and
• its own competing products.

An autonomous ranking algorithm may continually optimize for the platform's own commercial objectives.

The result could be:

Competitor → lower ranking → fewer consumers → lower sales → weaker competitor → greater platform dominance.

This creates a feedback loop.

Competition governance therefore needs to examine not merely whether the algorithm is "neutral" in technical terms, but whether its objective function systematically disadvantages competing undertakings.

17. Data Advantages and Autonomous Firms

Autonomous firms can create enormous competitive advantages through continuous data accumulation.

A platform may collect:

• consumer behaviour;
• prices;
• search histories;
• transaction data;
• competitor performance;
• inventory information;
• conversion rates;
• advertising data.

AI can transform this data into competitive intelligence.

Potential competition problems include:

1. Data foreclosure

Competitors cannot obtain essential data.

2. Data leveraging

The dominant firm uses data obtained in one market to dominate another.

3. Discriminatory access

The platform gives itself superior access to information.

4. Predatory learning

The autonomous system continuously learns from smaller competitors and adjusts its strategy against them.

5. Network effects

More users generate more data → more data improves AI → better service → more users.

This can create a self-reinforcing competitive advantage.

18. Autonomous Mergers

Autonomous firms also create challenges for merger control.

Traditional merger analysis examines:

• market shares;
• concentration;
• entry barriers;
• efficiencies;
• unilateral effects;
• coordinated effects.

But AI firms may have relatively small current revenues while controlling strategically important:

• datasets;
• models;
• computing infrastructure;
• APIs;
• talent;
• algorithms;
• distribution channels.

Consequently, competition authorities may need to examine innovation competition and potential competition, not merely current turnover.

19. Algorithmic Coordinated Effects in Mergers

Suppose a merger reduces four autonomous pricing firms to three.

Each firm operates a sophisticated machine-learning pricing system.

The merger could potentially make coordination easier because:

• fewer firms remain;
• market information becomes easier to process;
• price responses become more predictable;
• algorithms can monitor competitors continuously.

Thus merger assessment may need to consider whether the transaction changes the algorithmic conditions for coordination.

20. Autonomous Firms and Abuse of Dominance

A dominant autonomous firm may potentially engage in:

A. Predatory pricing

AI deliberately prices below an appropriate cost benchmark to eliminate rivals.

B. Excessive pricing

AI continuously maximizes prices in markets where consumers have limited alternatives.

C. Discriminatory pricing

Different consumers receive different prices based on data-driven classifications.

D. Self-preferencing

The AI systematically favours the firm's own services.

E. Refusal of access

AI automatically denies API, data or infrastructure access.

F. Tying

The system makes access to one service conditional upon adoption of another.

G. Loyalty restrictions

AI automatically rewards customers for remaining exclusively within an ecosystem.

21. Autonomous Firms and Predatory Pricing

Autonomous systems may be particularly effective at predatory strategies because they can:

• identify vulnerable competitors;
• monitor their cash position indirectly through market behaviour;
• reduce prices locally;
• increase prices after competitors exit;
• differentiate strategies across geographic markets.

The competition authority would need to distinguish:

legitimate dynamic pricing

from

strategic exclusionary pricing.

The fact that the price was produced by machine learning should neither establish nor eliminate liability by itself.

22. Autonomous Discriminatory Pricing

AI can produce individualized prices based on:

• location;
• purchasing history;
• device;
• browsing activity;
• willingness to pay;
• income proxies;
• urgency;
• previous purchases.

This creates potential concerns where a dominant undertaking uses personalized pricing to exploit consumers or disadvantage competitors.

The analysis must distinguish ordinary price discrimination from conduct prohibited under applicable competition law.

23. The "Black Box" Problem

One of the greatest governance challenges is explainability.

An AI system may generate the following outcome:

Price = ₹1,499.

But even its developers may not be able to explain precisely why the model selected that price.

This creates evidentiary difficulties.

A competition authority may need to determine:

• input variables;
• training data;
• objective function;
• model architecture;
• reward function;
• constraints;
• historical outputs;
• model updates;
• interactions with competitor systems.

Therefore, autonomous firms require strong algorithmic recordkeeping.

24. Evidence in Autonomous-Firm Investigations

Competition authorities may increasingly examine:

Technical evidence

• source code;
• model architecture;
• APIs;
• system logs;
• prompts;
• model weights where legally obtainable;
• training datasets;
• deployment records.

Commercial evidence

• pricing strategies;
• contracts;
• internal communications;
• market shares;
• customer data.

Behavioural evidence

• parallel pricing;
• exclusionary rankings;
• sudden market responses;
• systematic discrimination.

Governance evidence

• compliance policies;
• risk assessments;
• board decisions;
• model approval procedures.

The OECD has emphasized that authorities are developing technical capabilities to investigate algorithmic competition issues, while noting that sophisticated technical methods are not always necessary; the appropriate evidentiary method depends on the case.

25. Corporate Liability for Autonomous Decisions

A central legal question is:

Who is responsible when the AI makes the decision?

A useful responsibility chain is:

Developers → Enterprise → Managers → Algorithm → Market outcome

Competition law should ordinarily investigate the enterprise's role rather than treating the algorithm as an independent legal person.

Relevant questions include:

1. Who commissioned the system?
2. Who deployed it?
3. Who controlled its objectives?
4. Who supplied the data?
5. Who monitored it?
6. Who benefited from the conduct?
7. Who knew about the conduct?
8. Who could have stopped it?

26. The "AI Did It" Defence

An undertaking should generally not be able to avoid competition-law scrutiny simply by stating:

"The algorithm acted independently."

The stronger question is:

What degree of organizational control, foreseeability, monitoring and intervention existed?

This is especially important where:

• the firm chose the algorithm;
• the firm defined the objective;
• the firm supplied the data;
• the firm ignored warning signals;
• the firm benefited from the resulting conduct.

27. Autonomous Firms and Compliance Programmes

Traditional compliance programmes must evolve.

A modern autonomous-firm competition programme should contain:

Algorithmic Competition Compliance Policy

Every AI system affecting market conduct should undergo competition-risk assessment.

Competition-by-design

Competition safeguards should be incorporated into the system before deployment.

Algorithmic audits

Periodic testing should identify:

• collusion;
• discrimination;
• exclusion;
• self-preferencing;
• abnormal pricing.

Human override

There should be mechanisms for stopping or modifying problematic behaviour.

Audit trails

Important decisions should be reconstructable.

Model-change controls

Major changes to objectives or training data should undergo competition review.

28. Competition-by-Design

The concept is analogous to privacy-by-design, but applied to competition.

For example, a pricing AI could be designed so that it:

• does not receive competitors' confidential information;
• does not communicate with competitors' systems;
• avoids prohibited price-signalling mechanisms;
• maintains independent pricing parameters;
• records material price changes;
• flags unusual parallel pricing.

The objective is not to prohibit automation.

Rather:

Automation should be designed to compete, not coordinate.

29. Autonomous Agents as Market Participants

Future markets may involve autonomous agents negotiating directly with one another.

For example:

AI buyer → AI supplier → AI logistics agent → AI payment agent

No human may participate in an individual transaction.

This creates new questions:

• Can two AI agents form an agreement?
• What constitutes communication?
• Is machine-to-machine negotiation legally attributable to their firms?
• What happens when agents independently converge on a restrictive outcome?
• Who bears responsibility for emergent behaviour?

Existing competition law can address many traditional forms of conduct, but these questions may require further doctrinal development.

30. Autonomous Firms and Tacit Collusion

This is one of the hardest problems.

Suppose:

• Firm A independently develops AI;
• Firm B independently develops AI;
• neither communicates with the other;
• both systems learn that high prices maximize profits;
• prices converge.

There may be no conventional agreement.

This creates the distinction:

Explicit algorithmic collusion

Humans or firms agree to use algorithms to coordinate.

Facilitated algorithmic collusion

A third-party system facilitates coordination.

Autonomous tacit coordination

Independent algorithms learn coordinated strategies without an identifiable agreement.

The first two categories fit more comfortably within traditional competition-law frameworks. The third presents a much more difficult legal problem. The OECD has specifically highlighted this distinction.

31. Autonomous Firms and Third-Party AI Providers

Suppose ten competing firms purchase an AI pricing system from one provider.

The provider's system receives:

• market prices;
• demand data;
• inventory information;
• competitor information.

It then recommends prices to all ten firms.

This creates a potential AI hub-and-spoke structure.

The competition authority should examine:

• information flows;
• provider contracts;
• algorithm design;
• whether firms knew competitors used the same system;
• whether the provider encouraged coordination;
• whether competitively sensitive information was shared.

32. Governance Model

A practical regulatory model can be represented as:

Autonomous Firm

↓

Competition Risk Classification

↓

Algorithmic Audit

↓

Data & Information-Flow Review

↓

Collusion Testing

↓

Dominance / Exclusion Testing

↓

Human Accountability

↓

Continuous Monitoring

↓

Regulatory Audit / Investigation

↓

Remedy

This converts competition law from a purely ex post enforcement model into a combination of ex ante governance + continuous supervision + ex post enforcement.

33. Remedies Against Autonomous Firms

Competition authorities may employ:

Structural remedies

• divestiture;
• separation of business units;
• interoperability obligations.

Behavioural remedies

• prohibit discriminatory ranking;
• prohibit competitor-data use;
• require non-discriminatory access.

Algorithmic remedies

• independent algorithm audits;
• modification of objectives;
• restrictions on sensitive data;
• logging requirements;
• model transparency requirements.

Governance remedies

• compliance monitors;
• designated competition officers;
• board-level reporting;
• periodic certification.

Interim remedies

Where serious harm may occur rapidly, authorities may need to intervene before a lengthy investigation is completed.

34. Six Core Case-Law Lessons

35. Emerging Doctrine: Algorithmic Responsibility

A useful emerging principle is:

Algorithmic autonomy should not automatically equal legal autonomy.

The company remains relevant because it:

• selects the system;
• defines objectives;
• provides data;
• determines deployment;
• receives economic benefits;
• controls the surrounding business structure.

Therefore, competition governance should focus on attribution, foreseeability, control and effects.

36. Difference Between Human Firms and Autonomous Firms

37. Key Legal Challenges

1. Attribution

Who legally made the decision?

2. Intent

Can an algorithm possess the equivalent of anticompetitive intent?

3. Agreement

Can machine-to-machine interaction constitute coordination?

4. Evidence

How can regulators prove what occurred inside a black-box system?

5. Foreseeability

How much autonomous behaviour should a firm reasonably anticipate?

6. Causation

Did the algorithm actually cause the competitive harm?

7. Remedy

How can an authority remedy a constantly learning system?

8. Innovation

How can regulation prevent anticompetitive conduct without unnecessarily restricting beneficial AI innovation?

38. Indian Competition-Law Perspective

For India, autonomous-firm issues can be examined principally through the Competition Act, 2002, particularly:

• Section 3 – anti-competitive agreements;
• Section 4 – abuse of dominant position;
• Sections 5 and 6 – combinations;
• Section 19 – inquiry powers;
• Section 26 – investigation procedure;
• Section 27 – orders after inquiry;
• Section 28 – division of dominant enterprises;
• Section 36 – procedural and evidentiary powers.

The technological form of conduct should not automatically determine its competition-law character.

For example:

Human cartel → Section 3 issue

AI-executed cartel → potentially the same Section 3 issue

Similarly:

Human discriminatory ranking → possible Section 4 issue

AI-generated discriminatory ranking → potentially the same Section 4 issue

The difficult question is not whether AI exists, but how the AI's conduct fits within the statutory elements.

The CCI's work on digital markets and algorithmic risks is particularly relevant because algorithmic pricing and self-learning systems can create novel coordination and discrimination concerns.

39. Future Competition-Governance Principles

A mature autonomous-firm framework should incorporate the following principles:

Principle 1 – Human accountability

Every commercially consequential autonomous system should have an accountable enterprise owner.

Principle 2 – Competition by design

Competition risks should be evaluated before deployment.

Principle 3 – Data separation

Competitively sensitive information should not unnecessarily flow between rivals.

Principle 4 – Auditability

Important decisions should be reconstructable.

Principle 5 – Continuous monitoring

A system should not be treated as safe merely because it was compliant when first deployed.

Principle 6 – Model-change governance

Changes in objectives or learning architecture should trigger renewed competition assessment.

Principle 7 – Independent auditing

High-risk systems should be capable of independent review.

Principle 8 – Rapid intervention

Authorities should have appropriate tools where autonomous systems can cause rapid market-wide harm.

40. Conclusion

Autonomous firms represent a fundamental evolution in the structure of competition.

The principal problem is not simply that AI makes decisions. It is that AI can make millions of economically significant decisions continuously, adaptively and simultaneously.

Traditional competition law remains relevant to many of these situations. The cases involving Topkins, Eturas, Google Shopping, ASUS, online-platform parity restrictions, Apple and digital-platform conduct demonstrate that the use of technology does not by itself remove conduct from competition-law scrutiny.

The most difficult frontier is autonomous tacit coordination—where independent algorithms learn competitive strategies that produce coordinated outcomes without an identifiable human agreement. The OECD has recognized this as an important unresolved issue and has also emphasized that algorithmic harms can be both coordinated and unilateral.

Thus, the future of competition governance should move from a purely "human decision → legal consequence" model toward:

Data → Algorithm → Autonomous Decision → Market Effect → Continuous Audit → Corporate Accountability.