Artificial intelligence and competitive dynamics in downstream markets

1. Introduction

This paper examines how the adoption of artificial intelligence (AI) by firms in traditional and emerging sectors may affect competition. While much attention has focused on competition in AI infrastructure and foundation model development, the emphasis here is on how AI is used as an input into production, service delivery, logistics, and customer engagement. The paper is part of the OECD Horizontal Project on ‘Thriving with AI: Empowering Economies, Societies and Citizens’ and also contributes to a workstream of OECD “Competition and AI” papers that discuss steps in the AI value-chain, as well as various aspects of competition within the field of AI itself (see notably (OECD, 2024[2]; OECD, 2025[3]; OECD, 2025[4])).

AI adoption is heterogeneous across firms and sectors. Productivity effects also vary, with significant gains in professional services and manufacturing, but less in markets that largely rely on manual work and in-person services (Calvino, Reijerink and Samek, 2025[5]). The central question is whether using AI technologies helps lower barriers to entry and promote competition, or conversely may reinforce incumbency through control of data, compute, and ecosystems. AI can allow firms to automate, scale, personalise, and optimise outputs and services, supporting entry and innovation. At the same time, it may facilitate concentration of capability, vertical leveraging by dominant platforms, or new forms of anti-competitive behaviour, including algorithmic collusion.

The analysis does not aim to provide an exhaustive account of all sectors. Rather, the aim is to shed light on the factors and mechanics that support or hinder competition in the markets that adopt AI based tools. Given that issues specifically related to AI infrastructure will be addressed in a separate OECD paper (OECD, 2025[4]), this paper refers to them only insofar as they have direct implications for downstream competition.

As the topic is as yet under-researched and the sector is incredibly fast-moving, with daily breakthroughs, start-up launches or new uses reported in the media, many examples rely on press-reports or Internet research. Examples cited herein should be taken as illustrative as opposed to authoritative. Taken together, the aim is that these perspectives will contribute to a body of work to help readers better understand in which circumstances the use of AI is likely to be pro-competitive and hence, where regulators should seek to promote or protect its entry and use from vested interests or regulation; or conversely, to understand when the use of AI might be anti-competitive, and hence national competition authorities (NCAs) should be vigilant to protect competition across a wide range of downstream markets.

1.1. Key concepts used

This section describes, for the purpose of this paper, several concepts in artificial intelligence (AI) that are central to the analysis of its impact on competition. The terminology is grounded in OECD frameworks, including analytical and policy-oriented publications. The OECD defines an AI system as a machine-based system that, for explicit or implicit objectives, infers, from the input it receives, how to generate outputs such as predictions, content, recommendations, or decisions that can influence physical or virtual environments. Different AI systems vary in their levels of autonomy and adaptiveness after deployment (OECD Council, 2024[6]). The AI stack (infrastructure, model development and deployment), its technical components as well as the key enablers of compute, data and talent, along with their competitive impact, have already been extensively analysed in previous OECD competition work (OECD, 2024[2]; 2025[3]), and are therefore taken as given here.

Foundation Models (FMs) refer to a class of very-large AI systems (often a deep neural network) trained on vast and diverse datasets (often known as corpora1) using self-supervised learning techniques. They are designed to perform a wide range of general tasks across domains such as language, vision, and speech, and to be general-purpose in nature. FMs can power a variety of downstream applications and underpin many of the most advanced AI systems currently in use, including those used in language processing, image generation, and scientific discovery (OECD, 2024[2]; Calvino, Haerle and Liu, 2025[7]; Zenner, 2023[8]).

Large Language Models (LLMs) are a prominent type of FMs. They are designed to understand and generate human language, and can perform a wide array of tasks such as summarisation, translation, question answering, and content generation. They are often accessed using chatbot interfaces such as ChatGPT (Open.AI) Claude (Anthropic) and LeChat (Mistral). These interfaces determine how end-users access and benefit from the capabilities of foundation models (OECD, 2023[9]; byby.dev, 2025[10]).

Generative AI (GenAI) refers to AI systems that are capable of producing novel content, such as text, images, audio, video, or code, based on patterns learned from training data. The GenAI sector can be broadly divided into two segments: upstream, where foundation models are developed, and downstream, where specialised models are built from FMs. Developers may decide (i) to vertically integrate their foundation models with downstream AI services, (ii) to provide access to third-party developers, (iii) and the level of access they give to their models. Access to FMs is often granted via cloud services, which may in turn vertically integrate with specialised models, or provide services to third-party developers. The extent to which a FM is vertically integrated or made accessible to cloud providers or to third-party developers is the result of a strategic decision made by the FM developer (AdC, 2024[11]).

GenAI has been described as a general-purpose technology. With characteristics such as pervasiveness, continuous improvement and innovation-spawning potential, it may reshape productivity, innovation, and market structures across a wide range of sectors (Calvino, Haerle and Liu, 2025[7]). As with previous GPTs, such as electricity or the internet, the diffusion of generative AI is expected to be uneven, with significant implications for competition policy, including the risk of market concentration and the advent of new bottlenecks in the AI value chain.

Applications built on FMs are typically fine-tuned to perform specific tasks or serve particular user groups. It is a distinct and critical phase in the GenAI lifecycle (OECD, 2024[2]). Fine-tuning involves training the model further on domain-specific or firm-specific data to improve performance. For example, enterprise copilots may be fine-tuned on proprietary business data, while consumer chatbots may be adjusted for tone or content moderation. Fine-tuning is key to how GenAI systems are deployed in downstream markets. It influences accessibility, differentiation, and the potential for innovation across sectors.

Understanding how AI systems function, is essential for analysing how using AI may affect market entry, firm behaviour, and competitive dynamics. As these technologies evolve, so too will their implications for competition enforcement and policy design. Throughout the paper, the terms “AI-systems”, “AI” and “Generative AI (GenAI)” will be used interchangeably and will also sometimes encompass the features of FMs and LLMs. If an example or use case relies on a particular form of AI, this will be specified.2

The paper is structured as follows: Section ‎2 examines the uptake of AI across sectors and firms, highlighting both the opportunities for greater efficiency and the barriers that can limit adoption, including access to data and model openness shape, with implications for downstream contestability. Section ‎3 explores traditional and emerging competition concerns associated with AI systems, including algorithmic collusion and vertical foreclosure, and outlines actions to support competitive dynamics in AI-enabled markets, including enforcement, advocacy, regulation, and co-operation. The paper concludes with reflections on future research directions.

Notes

← 1.  In the context of AI and machine learning, the term “corpora” (plural of corpus) refers to large and structured sets of data, typically text, that are used for training and evaluating models, especially in natural language processing (NLP). They are important because Corpora are used to train AI models to understand and generate human language; evaluate model performance on tasks like translation, summarisation, sentiment analysis, or question answering; and fine-tune model for specific domains (e.g., legal, medical, financial). The are difficult to replicate for new entrants and constitute a significant barrier to entry.

← 2.  Company, AI-system and sector examples used in this paper are only meant to be illustrative. They are all drawn from public sources and are included to provide context, and should not be interpreted as endorsements.