The CCI Study on AI and Competition, Explained

Quite recently, the Competition Commission of India in partnership with the Management Development Institute, Gurugram, came up with a Market Study on Artificial Intelligence and Competition.

The study is extensive, and to that end, this explainer explores the most important estimates that have been provided by India's antitrust regulator.

---

The report is divided into six chapters and one annexure. However, this explainer intends to keep things specific, to examine the context set by the authors of the report, and evaluate the initial relevance of the suggestions made by them.

Defining the "AI Ecosystem"

Figure 1: Figure 3 on Artificial Intelligence (AI) Stack for Market Study on AI and Competition of CCI and MDI, Gurugram.

Market Growth Dynamics

• India's AI market expanded from USD 3.20 billion in 2020 to USD 6.05 billion in 2024

• Market projected to reach USD 31.94 billion by 2031 with 39-43% CAGR between 2025-2032

• Global AI market expected to reach USD 1 trillion by 2031

Startup Distribution Patterns

• 67% of AI startups concentrate on AI application model layer using existing foundation models

• Only 3% develop foundation models independently

• 20% operate in AI data layer functions

• 10% provide compute and AI infrastructure services

• 76% of startups build solutions using open-source frameworks

Technology Adoption Characteristics

• Machine learning dominates usage at 88% of startups

• Natural language processing adopted by 78%

• Generative AI/LLMs utilized by 66%

• Computer vision implemented by 27%

• 63% use pre-trained proprietary algorithms

• 66% use pre-trained open-source algorithms

Infrastructure Market Concentration

• AWS holds 32.6% market share in cloud computing

• Microsoft Azure commands 20.8% share

• Google Cloud Platform maintains 11.5% share

• Semiconductor dependency on NVIDIA, Intel, AMD specialized AI chips

Foundation Model Ecosystem Dominance

• Google leads with 18 foundation models

• Meta operates 11 foundation models

• Microsoft maintains 9 foundation models

• OpenAI provides 7 foundation models

Data Layer Market Structure

• Appen dominates with 23.4% market share in data services

• AWS holds 19.2% of data provision services

• Google commands 15.3% market share

• Microsoft Azure maintains 11.3% share

• Scale AI operates with 9.4% market presence

Entry Barrier Analysis

• 68% identify data availability/quality as primary barrier

• 61% cite cost of cloud services as significant challenge

• 61% report talent availability constraints

• 59% face high computing facility costs

• 56% encounter limited funding access

• 39% struggle with high data acquisition costs

  
  

Talent Scarcity Indicators

• 66% report talent not easily available

• 15% state talent rarely available

• 12% find talent somewhat accessible

• 52% of companies resort to in-house talent development programs

Funding Landscape Constraints

• 83% of startups rely on in-house funding

• 44% receive angel investor support

• 15% access government funding schemes

• 50% report next-level funding not easily available

• 13% find funding rarely available

Competitive Impact Measurements

• 52% report AI substantially improves competitiveness

• 34% note moderate competitiveness improvements

• 79% observe increased customer interactions through AI

• 21% report improvements in customer loyalty

• 69% use AI for demand prediction

• 24% utilize AI for pricing trend forecasting

Competition Issues in AI Industry with a User-Specific Focus

This section examines pages 47-71 of the market study, which depict the competition issues and advantages arising from AI adoption, revealing sophisticated dynamics specific to different AI classes, each with unique competitive implications.

Algorithmic Collusion in Large Language Models

Self-learning algorithms, particularly those utilising deep reinforcement learning, demonstrate the most concerning competitive risks in foundation model deployment. Unlike traditional collusion requiring human coordination, Q-learning algorithms can independently develop cooperative pricing strategies that maximise profits across market cycles. Consider a hypothetical scenario where multiple e-commerce platforms deploy foundation models such as GPT-4 or Claude for dynamic pricing: these models might independently learn that maintaining higher price levels generates superior collective outcomes, effectively achieving supra-competitive pricing without explicit communication between companies.

The study reveals that monitoring algorithms represent the most basic form of algorithmic coordination, essentially serving as sophisticated data collection tools that enable traditional collusion through enhanced market transparency.

In contrast, signalling algorithms utilising sophisticated statistical models can achieve above-equilibrium pricing without requiring explicit collusion amongst market participants. The most sophisticated self-learning algorithms demonstrate autonomous capability for deep reinforcement learning, dynamically responding to market changes whilst maximising profit through iterative learning cycles.

Machine Learning-Driven Market Concentration

Machine learning applications, utilised by 88% of surveyed startups, create distinctive competitive advantages through data network effects. The concentration of high-quality training data amongst established players creates insurmountable barriers for new entrants attempting to develop competing ML systems.

The study identifies that enterprises controlling vast datasets can train more effective AI models, creating substantial competitive advantages that entrench market power and establish high entry barriers for smaller players. This concentration effect becomes particularly pronounced in computer vision applications, where companies like Google and Meta leverage billions of labelled images from their platforms to maintain competitive moats that independent developers cannot overcome.

Natural Language Processing and Self-Preferencing

Natural Language Processing systems, adopted by 78% of startups, enable sophisticated self-preferencing strategies that traditional competition analysis struggles to detect.

Search ranking algorithms powered by NLP can systematically favour affiliated products through subtle modifications in relevance scoring that appear technically justified but substantially distort competitive dynamics.

Consider a hypothetical scenario where a major platform's NLP-powered search algorithm learns that promoting internally affiliated services generates higher user engagement metrics: the algorithm might organically develop preferential treatment patterns that effectively exclude competitors without explicit programming instructions.

Computer Vision and Predatory Pricing Precision

Computer vision technologies, implemented by 27% of startups, enable unprecedented precision in predatory pricing strategies. AI systems can monitor competitor behaviour in real-time, identify vulnerable market players, and implement targeted below-cost pricing exclusively for price-sensitive customers whilst maintaining standard pricing for less elastic demand segments.

Here's a hypothetical scenario.

Consider "RideMax," a dominant ride-sharing platform competing against smaller rival "LocalCab" in metropolitan areas. RideMax deploys sophisticated machine learning algorithms that continuously monitor LocalCab's driver availability, pricing patterns, and customer demand in real-time across hundreds of geographic zones.

When RideMax's algorithm detects that LocalCab has only three drivers active in the business district during peak hours, it immediately implements surgical predatory pricing—offering rides at 40% below cost exclusively in that zone. However, customers in suburban areas where LocalCab has no presence continue paying standard rates. The algorithm identifies price-sensitive customers through purchasing history analysis, targeting them with push notifications for the discounted rides, whilst maintaining regular pricing for business travellers who demonstrate lower price elasticity.

Within six months, LocalCab's drivers in the business district abandon the platform due to insufficient earnings, and the company exits that market segment. RideMax then gradually increases prices to 20% above pre-competition levels, using the same algorithmic precision to optimise revenue recovery whilst avoiding customer defection thresholds.

Competitive Advantages Through AI Classes

The study reveals that 52% of companies report substantial competitiveness improvements from AI adoption, with 79% observing increased customer interactions. Machine learning applications provide competitive advantages through demand prediction capabilities, utilised by 69% of respondents, enabling superior inventory management and pricing optimisation. Natural language processing systems enhance customer engagement through improved chatbot interactions and sentiment analysis, creating customer loyalty advantages that compound over time.

Computer vision applications deliver competitive advantages through automated quality control, enhanced security systems, and improved logistics optimisation that reduce operational costs whilst improving service quality. Generative AI provides competitive advantages through content creation efficiency, personalised customer communications, and automated report generation that enables smaller firms to compete with larger entities in content-intensive markets.

Technical Considerations for AI Competitive Advantage Assessment

Study Findings vs Emerging Technical Understanding

• Positive Market Indicators: The CCI study effectively documents that 52% of companies report substantial competitiveness improvements through AI adoption, with 79% observing increased customer interactions

• Technical Reality Gap: Recent academic research suggests a potential disconnect between perceived benefits and underlying technical capabilities, particularly regarding Large Language Model reliability in business-critical applications

• Generative AI Usage: While 66% of surveyed startups utilize generative AI for competitive advantages, emerging studies indicate these systems may exhibit "fluent but not faithful" behaviour patterns that could impact long-term strategic value

Critical Assessment Areas for Policy Consideration

• Accuracy in Forecasting Applications: Given that 69% of respondents rely on AI for demand forecasting and 24% for pricing predictions, policymakers may wish to consider frameworks for validating AI system reliability in financial decision-making contexts

• Structural Limitations: Technical analysis suggests that hallucinations in AI systems may be architectural features rather than engineering problems, potentially affecting the sustainability of claimed competitive advantages

• Training Data Dependencies: Research indicates that AI systems may perform poorly when encountering scenarios significantly different from their training environments, raising questions about the robustness of transfer learning benefits highlighted in the study

Market Structure Implications

• Concentration Effects: The study's finding that 67% of Indian AI startups operate at the application layer using existing foundation models may create systemic vulnerabilities if underlying technologies have inherent limitations

• Investment Strategy Considerations: Policymakers may benefit from developing frameworks to assess the technical sustainability of AI-driven competitive strategies before making large-scale infrastructure investments

  
  

Regulatory Framework Enhancement Opportunities

• Transparency Initiatives: Enhanced technical assessment capabilities within regulatory bodies could help identify genuine competitive advantages versus temporary market perceptions

• Balanced Innovation Support: Future policy frameworks might benefit from incorporating both market dynamics and technical sustainability assessments to ensure long-term competitive ecosystem health

Feedback on Annexure 1 on Guidance Note on Self-Audit of AI Systems for Competition Compliance

The checklist exhibits characteristics that strongly suggest AI-generated content, lacking the precision and specificity required for meaningful competition law compliance.

Definitional Vagueness and Threshold Absence

• The guidance note fundamentally fails to establish any meaningful thresholds for compliance obligations.

• Questions like "Is there a documented AI governance framework in place?" and "Are roles and responsibilities for AI competition compliance clearly assigned?" provide no guidance on what constitutes adequate documentation or sufficient assignment of responsibilities.

• This absence of specific criteria renders the checklist essentially meaningless as a compliance tool, as enterprises could theoretically satisfy these requirements with minimal, superficial documentation while completely failing to address substantive competition risks.

• The framework's reliance on subjective assessments without quantitative benchmarks creates a regulatory vacuum where compliance becomes a matter of interpretation rather than adherence to clear standards.

For instance, the question "Are safeguards against collusion built into the algorithm design?" provides no definition of what constitutes effective safeguards, how they should be implemented, or what level of protection is considered adequate.

Mechanistic Structure Without Strategic Focus

• The six-pillar framework (governance, algorithm design, testing, monitoring, transparency, and compliance integration) appears systematically comprehensive but lacks strategic prioritization based on actual competition risks. The checklist treats all AI systems uniformly, regardless of their market impact or competitive significance.

A recommendation algorithm for a small e-commerce platform receives the same scrutiny framework as pricing algorithms deployed by dominant market players, demonstrating a fundamental misunderstanding of how competition risks scale with market power and system influence.

• This one-size-fits-all approach contradicts established competition law principles that recognize the differential impact of practices based on market position and competitive context.

• The guidance note's failure to differentiate between high-risk and low-risk deployments suggests either inadequate understanding of competition dynamics or reliance on generic AI governance templates rather than competition-specific analysis.

Technical Superficiality Masking Legal Complexity

• The technical questions reveal a concerning disconnect between the complexity of modern AI systems and the simplistic approach adopted in the checklist.

• Questions like "Has the algorithm been tested across various market scenarios?" and "Has the algorithm been evaluated for potential collusive outcomes?" demonstrate fundamental misunderstanding of how algorithmic systems actually operate.

• Modern machine learning systems, particularly those utilising deep neural networks, cannot be meaningfully "tested" for collusive outcomes through traditional scenario analysis, as their behaviour emerges through training processes that may produce unpredictable market interactions.

• The guidance note's treatment of algorithmic transparency through questions like "Is the algorithm's decision-making process explainable?" ignores the inherent opacity of many effective AI systems, particularly deep learning models that operate as "black boxes" even to their developers.

• This creates an impossible compliance burden where enterprises must choose between deploying effective but opaque systems or explainable but potentially inferior alternatives.

Enforcement Gap and Regulatory Weakness

Perhaps most critically, the guidance note provides no connection between self-audit findings and regulatory consequences.

• The disclaimer explicitly states that "any issue relating to alleged anti-competitive conduct would be examined by the Commission on a case-by-case basis within the provisions of the Act". This renders the entire self-audit framework essentially advisory, with no clear pathway from compliance failures to enforcement action.

• The absence of mandatory disclosure requirements, standardized reporting formats, or regulatory review mechanisms means enterprises could conduct perfunctory self-audits that satisfy the letter of the guidance while completely failing to identify or address genuine competition risks.

• The framework's voluntary nature, combined with its vague requirements, creates perverse incentives where superficial compliance efforts provide legal protection without meaningful risk mitigation.

AI-Generated Content Characteristics

• The checklist exhibits several hallmarks of AI-generated content that compromise its credibility as a serious regulatory instrument.

• The repetitive structure, generic phrasing, and comprehensive-but-shallow coverage pattern suggests automated content generation rather than expert legal drafting.

• Questions like "Does AI/algorithmic pricing strategies of the firm align with competitive fairness?" and "Does AI/algorithmic pricing strategies of the firm align with regulatory compliance?" demonstrate the redundant, template-like quality typical of AI-generated compliance materials.

The guidance note's inability to address edge cases, provide specific examples, or acknowledge the inherent tensions between different compliance objectives further supports this assessment. Human-drafted competition law guidance typically includes detailed scenario analysis, practical examples, and acknowledgment of implementation challenges—all absent from this document.