Artificial intelligence and competitive dynamics in downstream markets

2. The impact of AI adoption on market dynamics

As a general-purpose technology, GenAI exhibits defining characteristics such as pervasiveness, continuous improvement over time and innovation spawning, with implications for future productivity growth (Eloundou et al., 2024[12]; Calvino, Haerle and Liu, 2025[7]; Aghion and Bunel, 2024[13]). Estimates of future gains to total factor productivity (TFP) from the use of AI range from a conservative annual increase of 0.07 percentage points (Calvino, Haerle and Liu, 2025[7]), to an estimate of 0.68 pps of additional TFP growth per year (Aghion and Bunel, 2024[13]), to a recent OECD study which finds a range of 0.4-1.3 percentage points in annual TFG growth in countries with high AI exposure, such as the United States (Filippucci et al., 2025[14]). Most of these studies look at expected AI adoption rates by sector and their estimated potential effect on sector productivity, then aggregate to the wider economy. Each study comes with many caveats, not least the expected rate of adoption, and the likely effect on firm or sector productivity.

Taking a starting point in this discussion, this chapter focuses on the mechanisms through which AI adoption enables firms to lower costs and increase productivity, thereby reducing barriers to entry and enhancing market contestability. The chapter also explores sector-specific examples and emerging evidence on AI uptake, while noting that the competitive effects of AI depend on its diffusion and integration across industries. The chapter first considers in which ways AI may substitute for labour, lower costs for firms and increase labour productivity, using an emerging empirical literature. The chapter then looks at some of the more negative aspects associated with AI adoption, before moving sector-exposure to AI, to try to gauge its likely adoption. Finally, some surveys on actual AI use are discussed.

2.1. What are the mechanics by which AI can foster competition?

Downstream firms that adopt AI may be able to produce outputs and services at lower costs, fewer staff, and higher productivity, while also offering new services, all of which contribute to lowering entry barriers. This section takes a look at these aspects.

2.1.1. Labour substitution and augmentation

GenAI systems are by now able to perform many tasks that used to be reserved for cognitive workers, such as writing, coding, summarising articles, brainstorming ideas, organising plans, translating other languages, writing complex emails, generating images, and much more. These capabilities allow firms to automate or accelerate labour-intensive processes, often without requiring specialised technical skills. Natural Language Processing (NLP) interfaces such as chatbots and voice assistants, enable employees with limited training or qualifications to interact with AI tools to generate code, perform data analysis, or produce professional writing outputs, potentially lowering the skills threshold for market participation (OECD, 2024[2]).1 This could potentially allow SMEs or start-ups to enter markets with fewer staff and lower fixed costs.

GenAI helps structure and guide less experienced users while complementing the expertise of advanced workers. When used collaboratively in teams to perform group tasks, GenAI improves team performance and supports personalised learning. By automating routine tasks, AI also frees individuals for more complex or creative work, enabling leaner, more agile business models that can challenge established players OECD research finds that the strongest gains from using AI come when GenAI is used to augment human capacities by automating well-defined tasks, as opposed to tasks performed solely by the AI (Calvino, Reijerink and Samek, 2025[5]).

In particular, AI systems can substitute for more experienced workers for some tasks, and can also reduce the need for large teams in areas such as customer support, software development, and marketing. These dynamics could lower barriers to entry in knowledge-intensive sectors, where expertise and team size have traditionally conferred competitive advantage (Brynjolfsson, Li and Raymond, 2023[15]; Li et al., 2024[16]; Noy and Zhang, 2023[17]; Peng et al., 2023[18]). Field experiments confirm that less experienced developers achieve disproportionate productivity gains when supported by AI assistants (Cui et al., 2024[19]) (see also Box 1).

Box 1. AI and knowledge worker substitution

Experimental evidence suggests that generative AI systems could potentially alter the landscape of knowledge work by substituting for certain forms of expertise and narrowing performance gaps between workers. In a large-scale field experiment with over 750 consultants at Boston Consulting Group, Dell’Acqua et al. found that access to GPT-4 led to a 12% increase in task completion, a 25% reduction in time spent, and a more than 40% improvement in output quality. However, the largest gains were observed among lower-performing individuals, whose performance improved by 43%, compared to 17% for top performers. These findings suggest that generative AI could potentially substitute for experience, or specialised training, enabling less experienced workers to perform at levels previously associated with more senior staff.

This form of knowledge substitution could have important competitive implications: if deployed at scale it would reduce the need for large, highly skilled teams, lower the cost of entry into knowledge-intensive markets, and allow smaller firms to compete more effectively with incumbents.

The study also identifies emerging collaboration models, the so-called “Centaurs,” who strategically divide tasks between humans and AI, and “Cyborgs,” who fully integrate AI into their workflows, highlighting how AI can complement rather than replace human capabilities. Within knowledge workers, the use of AI, if fully integrated, could then potentially redistribute competitive advantage in favour of agile, tech-enabled challengers.

Source: Dell’Acqua et al (2023[20]), Navigating the Jagged Technological Frontier: Field Experimental Evidence of the Effects of AI on Knowledge Worker Productivity and Quality, http://dx.doi.org/10.2139/ssrn.4573321.

AI-enabled tools reduce the minimum efficient scale by allowing smaller teams to perform functions such as customer support, compliance monitoring and demand forecasting that previously required large workforces, thereby lowering entry barriers and enabling incremental scaling with limited upfront investment (Gupta, 2025[21]).2 In software development, GenAI models capable of generating and debugging code similarly reduce labour intensity and team size requirements. In software development, using coding copilot systems has enabled developers to complete coding tasks up to 55% faster (Peng et al., 2023[18]; Chui et al., 2023[22]). This reduces both the total hours of engineering labour required and the need for large quality-assurance teams. For example, an early-stage software-as-a-service (SaaS) company can release a minimum viable product with only a handful of engineers, whereas historically larger teams would have been required to build and maintain reliable software (InData Labs, 2024[23]). By cutting operational and engineering costs, AI lowers capital and labour thresholds for market entry, enabling micro-vendors to contest markets traditionally dominated by larger incumbents (Calvino, Reijerink and Samek, 2025[5]). These factors all point to GenAI as an instrument that can increase market competition by reducing the minimum efficient scale of firms, by substituting the AI systems for some of the required labour force.

The underlying assumption that GenAI is cheaper than human labour will depend on cost structures related to usage models and billing arrangements, as well as the sector and tasks where the AI systems are deployed. From a competition perspective, therefore, although the substitution of labour through AI may lower entry barriers by reducing the need for specialised staff and enabling leaner operations, these benefits may not be evenly distributed. The financial and organisational costs of adopting AI may constitute significant barriers, especially for smaller firms and less digitalised sectors. These include high upfront investment requirements (for data infrastructure, computing resources and personnel to handle the AI), integration costs linked to adapting legacy systems, and recurring expenses for cloud services, model fine-tuning and cybersecurity. These costs are not limited to technology acquisition but extend to complementary investments, such as staff training, change management and data governance frameworks (Kergroach and Héritier, 2025[24]). As such, the competitive impact of labour substitution is likely to be heterogeneous, depending on firm capabilities and sectoral conditions.

2.1.2. Product improvement and innovation

AI systems may also contribute to a reshaping of competition by enabling new forms of product and service differentiation, and improve service quality. GenAI allows firms to personalise offerings at scale, tailoring content, interfaces, and recommendations to individual users. This works by the AI learning detailed representations of individual users from their behaviour and context, which allows the system to predict what any given user is likely to find useful or engaging. Firms can then dynamically adjust the content, recommendations or interface shown to each user, scaling this personalised adaptation across millions of users at once (Rafieian and Yoganarasimhan, 2023[25]). Firms integrating AI into decision-making report better forecasting and personalised services, improving customer experience, which creating opportunities for challengers to differentiate rather than to compete on scale alone (Krakowski, Luger and Raisch, 2022[26]). This capacity for mass customisation is increasingly accessible to smaller firms using prompt-based tools and multimodal models.

In sectors where customer experience and brand identity are key competitive levers, such democratisation of design and communication tools can intensify rivalry and erode incumbents’ traditional advantages. It enables new forms of user engagement, product differentiation, and service delivery, with implications for market structure, innovation, and accessibility in downstream markets. AI-generated marketing assets, product imagery, and voice content allow small firms to present professionally branded services without external agencies (Calvino, Haerle and Liu, 2025[7]; OECD, 2024[2]). Over time, AI-generated outputs – cheaper, faster, and often indistinguishable from human-created content – may reshape market dynamics in sectors such as imagery, music, film animation, coding, marketing, translation, and customer services. In these fields, consumer preference for low-cost, high-quality AI content could increase rivalry between AI-enabled entrants and established firms.

Integrating AI can also stimulate innovation and research and development (R&D): More than 70% of surveyed enterprises conduct R&D on AI for their own use, and over half collaborate with universities or public research organisations, supporting both innovation and talent acquisition (OECD/BCG/INSEAD, 2025[27]). In pharmaceuticals, AI-driven drug discovery is expected to reduce R&D times, opening opportunities for biotech entrants (OECD/BCG/INSEAD, 2025[27]). In transport, studies of autonomous vehicle fleets in Singapore suggest that AI can expand modal choice and reduce consumer costs when integrated with public transport (Mo et al., 2021[28]).

By lowering the cost of prediction, data analysis, optimisation and creative production, AI may enhance competition by enabling new entrants to challenge incumbents with novel offerings and shifting the sources of competitive advantage towards firms that combine domain expertise with AI capabilities (Agrawal, Gans and Goldfarb, 2018[29]; Krakowski, Luger and Raisch, 2022[26]; Gupta, 2025[21]). These capabilities intensify competition and raise incentives for incumbents to innovate (Gupta, 2025[21]).

2.1.3. Cost reduction, efficiency and productivity gains

AI adoption can support competition by lowering entry barriers thanks to a reduction of operating costs. This happens through improving information efficiency, automation, task complementarity, and improved error control (Acemoglu, 2024[30]). When firms incorporate AI into their production processes, they incur variable costs for acquiring the necessary software as well as fixed costs for installing AI infrastructure (Maydell, 2024[31]). Costs involved in deploying AI systems include model training, infrastructure, and data integration. However, AI-systems enable task-level cost savings for downstream users, implying that additional tasks can be handled at relatively low marginal cost once systems are deployed (teneo.ai, 2025[32]; Acemoglu, 2024[30]; Korinek, 2024[33]). For example, predictive maintenance systems can monitor thousands of machines simultaneously without proportional increases in labour or analytics costs, enabling firms to expand output and profitability with relatively low incremental cost. (Aggarwal et al., 2021[34]). In health diagnostics, AI models used in clinics can interpret additional scans or tests at negligible extra cost compared to human specialists, whose time and attention scale linearly with workload (McGenity et al., 2024[35]; Aggarwal et al., 2021[34]).

AI systems can replicate certain expert tasks (e.g. fault detection, anomaly classification, image interpretation) which traditionally require highly trained personnel. Unlike human experts, automated inference allows scaling through compute rather than labour, reducing the cost of handling additional cases. In manufacturing, AI-based predictive maintenance systems can anticipate failures ahead of time, and thereby sharply lower unplanned downtime and emergency repair costs (e.g. unplanned downtime reductions of up to 70%, lifecycle extension of 25–30%) (Zeb and Lodhi, 2025[36]; Benhanifia et al., 2025[37]). Compared to reactive or schedule-based regimes, these systems also lower maintenance costs (Florian, Sgarbossa and Zennaro, 2021[38]) (see Box 2 for examples of AI use in manufacturing). These efficiencies allow firms to expand production without equivalent increases in staff or cost, enhancing competitiveness (Gupta, 2025[21]).

Box 2. Use of AI in the manufacturing sector

• Predictive maintenance and MRO (maintenance, repair, operations). Generative AI can help forecast equipment failures by analysing sensor, log and maintenance history data; generate diagnostic suggestions or “next-step” interventions; and optimise spare parts inventory (Santamaria, 2024[39]). In the aerospace sector, airlines and MRO providers are piloting AI “copilot” assistants that technicians can query (e.g. “What might cause a leak in compressor X?”), drawing on maintenance manuals, historical repairs and sensor data. AI tools can also streamline part planning, align procurement with expected failures, and reduce unplanned downtime. (McKinsey & Company, 2024[40]; MarketsAndMarkets, 2025[41])

• Digital twin and data integration. Some MRO or manufacturing firms use digital twin models, i.e. virtual replicas of physical machines or systems, that are continuously updated with sensor data. Generative AI may use these digital twins for simulation, anomaly detection, and “what-if” scenarios, enabling more advanced diagnostics or maintenance planning (MarketsAndMarkets, 2025[41]; Apostolidis and Stamoulis, 2021[42]).

• Quality inspection and defect detection. In manufacturing lines, AI models (often combining convolutional neural networks, generative adversarial networks or autoencoder architectures) are used to detect microscopic defects, anomalies or deviations beyond human inspection thresholds. Some systems incorporate generative modules to synthesise rare defect variants for training, increase robustness to edge cases (Kumbhar et al., 2023[43]). In trimming-die design, for example, one system automates inspection by integrating AI modules with CAD, reducing inspection time by ~80 % relative to expert manual inspection (Lee et al., 2023[44]).

• Design automation and generative manufacturing. More recently, generative manufacturing systems have been proposed whereby AI (e.g. diffusion models, prompt-based systems) can assist in layout design, production sequencing, and component configuration. In one prototype, a generative manufacturing system reduced decision latency and improved responsiveness to changing constraints, allowing more flexible production (Li et al., 2024[16]).

Sources: Apostolidis and Stamoulis (2021[42]), An AI-based Digital Twin Case Study in the MRO Sector, https://doi.org/10.1016/j.trpro.2021.09.007; MarketsAndMarkets (2025[41]) AI Impact Analysis on Digital MRO Industry, https://www.marketsandmarkets.com/ResearchInsight/ai-impact-analysis-on-digital-mro-industry.asp; McKinsey & Company (2024[40]) The generative AI opportunity in airline maintenance, https://www.mckinsey.com/industries/aerospace-and-defense/our-insights/the-generative-ai-opportunity-in-airline-maintenance; Kumbhar et al. (2023[43]) DeepInspect: An AI-Powered Defect Detection for Manufacturing Industries, https://doi.org/10.48550/arXiv.2311.03725; Lee et al. (2023[44]) Automation of Trimming Die Design Inspection by Zigzag Process Between AI and CAD Domains, https://arxiv.org/abs/2305.16866v1; Santamaria (2024[39]) Blog, Industrial Parts, Industrial Technology, https://www.advancedtech.com/blog/ai-and-industrial-mro; Li et al. (2024[16]) Generative manufacturing systems using diffusion models and ChatGPT, https://arxiv.org/abs/2405.00958v2.

In high-volume customer support services (in sectors such as large e-commerce platforms, major telecoms, large financial institutions and global SaaS/IT service providers), AI systems may benefit from economies of scale, as per-interaction costs decline with rising volumes, potentially offering efficiency gains compared to human operations which require additional staffing, benefits, facilities and training. Although enterprise-grade deployment can entail significant upfront expenditure, estimates suggest that these may be recovered within a relatively short implementation period, depending on usage intensity.3 Once in place, AI agents can operate at around USD 0.25–0.50 per interaction – compared with USD 3–6 for human agents and up to USD 15 per interaction for complex tasks, or working overtime – suggesting substantial scope for variable cost reduction (teneo.ai, 2025[32]).

AI adoption may also reduce reliance on external business process services (e.g. customer support, data entry, payroll processing, translation, or IT services)4 and related agency fees, by enabling in-house execution of tasks that were previously outsourced. The best performing organisations report outsourcing cost reductions of USD 2 million-10 million annually and a 30% decrease in the costs of external creative or content generation, including translations, from switching to GenAI for tasks previously outsourced (MIT, 2025[45]; CMA, 2023[46]).5,6 In the audit sector, AI adoption has been linked to lower audit fees and reduced staffing needs, attributed to improved accounting quality and reduced information asymmetry. The authors interpret this as AI enabling more efficient audit and accounting processes, displacing routine labour and lowering variable internal costs (Lai, 2025[47]; Fedyk et al., 2022[48]). In knowledge-intensive activities, AI systems can accelerate design and development cycles and automate administrative functions, helping smaller firms to reduce unit costs and compete more effectively in digitally intensive sectors (Babina et al., 2024[49]; Gupta, 2025[21]).

AI capabilities can also be used across different domains, improving the return on upfront investments. For instance, the GenAI used in customer service can also be adapted for compliance monitoring, sales forecasting, or internal documentation. Access to third-party foundation models via APIs or open-source tools allows firms to deploy “plug-and-play” capabilities (e.g. chatbots, image generation, code synthesis) that can be adapted to specific use cases without training their own models (OECD, 2024[2]; Chui et al., 2023[22]). Such modularity may reduce fixed costs, shorten development cycles and lower the minimum efficient scale required to compete in AI-enabled service markets (Brynjolfsson, Li and Raymond, 2023[15]; OECD, 2024[50]; Gupta, 2025[21]). For instance, a start-up offering financial analysis tools can integrate a third-party generative model to provide natural-language query functionality (chatbots) to clients, without having first to train its own large-scale model (see (Yang et al., 2023[51])).

Such efficiencies may lower the cost of innovation and support entry into adjacent or previously unattainable markets, with indications of improved performance outcomes such as reduced defect rates, lower energy consumption and greater supply chain resilience (Calvino, Haerle and Liu, 2025[7]; Calvino, Reijerink and Samek, 2025[5]).

Modularity is not unique to AI. Many other, new general-purpose technologies, such as cloud computing, APIs, or Services-as-a-Service (SaaS) platforms have also enabled scalable deployment. Accordingly, the pro-competitive effect of AI may depend on broad, affordable and non-discriminatory access to these tools (OECD, 2025[3]; CMA, 2023[46]; Agrawal, Gans and Goldfarb, 2018[29]). Where such access is open, AI can enhance market contestability by enabling smaller firms to enter and scale in markets that were previously dominated by incumbents with proprietary infrastructure and large datasets. In the financial sector, Fintechs can use third-party AI models to assess creditworthiness using alternative data, bypassing the need for traditional credit scoring infrastructure (CGAP, 2024[52]) (see Box 3.). This may reduce scale advantages and allow firms to compete more on ideas, speed and customer insight rather than size.

Box 3. AI-driven credit scoring can boost alternative finance

In the domain of financial services, AI is being leveraged to expand access to credit for underserved populations across Africa. The Consultative Group to Assist the Poor (CGAP) has highlighted how financial institutions and fintech startups are using AI to analyse alternative data sources – such as mobile phone usage, transaction histories, and social media behaviour – to assess creditworthiness.

This approach is particularly relevant in contexts where formal credit histories are rare or non-existent. By using AI to interpret behavioural and transactional data, lenders can offer microloans and tailored financial products to individuals and small businesses that would otherwise be excluded from formal financial systems. Moreover, AI enables automated customer service and fraud detection, reducing operational costs and improving service delivery.

The competitive impact of these innovations is twofold. First, they enable a broader range of actors to access capital, thereby increasing the number of market participants. Second, they allow smaller financial institutions and fintechs to compete with established banks by offering more agile and inclusive services. This contributes to a more diversified and competitive financial ecosystem.

Source: CGAP (2024[52]) “Data and AI for Inclusive Finance”, https://www.cgap.org/topics/collections/data-and-ai-for-inclusive-finance.

Productivity gains

Several studies have also tried to measure increases in productivity from using AI tools. Some of the tasks included customer support in the form of resolving queries, (Brynjolfsson, Li and Raymond, 2023[15]); mid-level writing such as press-releases, reports and emails (Noy and Zhang, 2023[17]); economic research (ideation, data analysis and mathematical derivations) (Korinek, 2023[53]); professional services tasks (text and coding, summarising, translating, drafting) (Eloundou et al., 2024[12]) as well as cognitive tasks in writing, coding, planning, translating and summarising for cognitive workers (doctors, lawyers managers, researchers) (Baily, Brynjolfsson and Korinek, 2023[54]). Most of the above studies find positive productivity gains, of between 10%-20%, reaching up to 35% for more low-skilled agents (Brynjolfsson, Li and Raymond, 2023[15]). Baily, Brynjolfsson and Korinek find a potential increase in the level of GDP of 18% over ten years (Baily, Brynjolfsson and Korinek, 2023[54]). Experimental evidence shows significant productivity gains from using AI chatbots for professional use (see Box 4).

Box 4. Experimental evidence on the productivity effects of generative AI

In a controlled study measuring the productivity effects of generative AI in professional tasks, Noy and Zhang (2023) conducted a randomised controlled trial involving 444 university-educated professionals across various sectors, including marketing, human resources, consulting, and data analysis. Participants were recruited via an online labour platform and asked to complete a series of business-writing assignments simulating real-world professional outputs such as press releases, analytical summaries, and policy briefs.

Participants were randomly assigned to either a treatment group with access to ChatGPT or a control group without access. Each participant completed two writing tasks: one before and one after ChatGPT was introduced to the treatment group, allowing the authors to capture both cross-sectional and within-subject changes in output quality and efficiency. Task quality was evaluated by an independent panel blind to treatment status, based on criteria such as coherence, accuracy, tone, and persuasiveness.

The results show a marked increase in productivity among users of generative AI tools. Access to ChatGPT reduced task completion time by around 40% and increased output quality ratings by approximately 18%, leading to an overall productivity gain of about 37%. The productivity improvements were most pronounced among lower-skilled participants, who closed much of the gap with higher-skilled peers. This suggests that generative AI may act as an “equalising technology” in written communication tasks, potentially levelling skill disparities in cognitive and creative work.

The study also found that workers tended to integrate AI into their workflow strategically, using it to structure drafts, refine tone, and correct grammar, rather than fully delegating the writing task. However, follow-up qualitative analysis indicated that repeated reliance on ChatGPT led to increasing stylistic convergence, with outputs becoming more formulaic and less distinctive over time. The authors warn that this “homogenisation effect” could limit creativity, diversity of thought, and differentiation in markets relying heavily on written content.

Finally, the authors emphasise that their findings may overstate short-run productivity gains and do not capture potential long-term trade-offs such as skill erosion, diminished originality, or shifts in professional norms. The study concludes that while generative AI can substantially enhance individual task performance in the short term, the aggregate and sustained productivity impacts across firms and sectors remain uncertain.

Source: Noy, S. and Zhang, W. (2023[17]) Experimental Evidence on the Productivity Effects of Generative Artificial Intelligence, https://ssrn.com/abstract=4375283.

What seems to be common in these studies is that using AI frees staff from routine tasks, enabling them to focus on higher-value activities, which supports in-firm productivity gains. A LinkedIn/Ipsos survey found that 84% of salespeople saved at least 30 minutes per day through AI, and 65% said it helped them exceed quotas. More than 70% of executives expressed confidence in AI managing core functions such as forecasting and account research. Firms also report improved customer outcomes: AI-powered outreach systems have boosted customer retention by 10% in some cases and increased the identification of sales “leads” by up to 40%7 (LinkedIn/Ipsos, 2025[55]). In some sectors, AI adoption may fundamentally change business models and raise productivity at scale.8 In agriculture, for example, AI-enabled systems for crop monitoring and predictive management have been shown to boost output and resilience, offering particular benefits for smallholders facing the challenges of climate change (see Box 5.).

Box 5. Generative AI for agricultural support in Malawi

In Malawi, the deployment of generative AI technologies has begun to transform the agricultural landscape for smallholder farmers, particularly as climate-related disruptions are rising. One initiative involves the use of a locally adapted AI chatbot named Ulangizi, which means “advisor” in Chichewa. Developed by Opportunity International, this tool operates via WhatsApp and is capable of delivering agronomic advice in local languages, including audio responses for users with limited literacy.

The chatbot was introduced in response to the devastation caused by Cyclone Freddy, which displaced thousands of farmers and destroyed crops across southern Malawi. One such farmer, Alex Maere, lost his maize fields and was forced to reconsider his farming strategy as the hurricane altered the soil quality. With guidance from the AI chatbot, Mr Maere planted potatoes in addition to his staples, corn and cassava, ultimately earning over USD 800 from his potato harvest alone. This income enabled him to pay school fees for his children and reinvest in his farm, demonstrating the potential of AI to support rapid recovery and resilience among small-scale producers. The United Nation’s International Fund for Agricultural Development believes that AI has the potential to uplift agriculture in sub-Saharan Africa, where an estimated 33-50 million smallholder farms such as Mr Maere's produce up to 70-80% of the food supply. Private investment in agriculture-related tech in sub-Saharan Africa went from USD 10 million in 2014 to USD 600 million in 2022, according to the World Bank.

From a competition perspective, using such AI-powered advice can reduce barriers to agronomic expertise, which have traditionally favoured larger, better-resourced farms and better educated farmers. By providing tailored, real-time advice to smallholders, AI may enable more equitable participation in agricultural markets and enhances the productivity of previously disadvantaged producers.

Source: Gondwe (2025[56]) “In Malawi, small-scale farmers turn to AI technology for advice”, https://phys.org/news/2025-09-malawi-ai-technology-small-scale.html.

These productivity gains may not always materialise. At the economy-wide level, productivity growth may be hindered by the fact that certain tasks cannot be fundamentally altered even with AI, however, wages will still increase, thereby lowering productivity. This is the so-called Baumol effect. Some tasks (for instance, brick-laying) are still executed more or less as they have been for centuries, or a symphonic orchestra will still take the same time to perform Beethoven’s 5th Symphony today as when it was first performed, but the workers or musicians will be receiving today’s wages. Extended across a large number of sectors that benefit less from technology-driven output gains, this Baumol effect will undermine economy-wide productivity gains (Filippucci et al., 2025[14]). This is also related to the fact that some sectors are naturally more prone to AI adoption and exposure than others (see Section ‎2.3).

Some evidence also indicates that AI use may even reduce efficiency in certain contexts. A randomised controlled trial found that experienced developers using early-2025 AI coding assistants completed tasks around 19% more slowly, suggesting that automation may interfere with expert workflows (Becker et al., 2025[57]). AI-assisted code writing has also been associated with additional security vulnerabilities, as some coding assistants are trained on erroneous public code, potentially generating rework that offset short-term gains (Sandoval et al., 2023[58]). A systematic review of AI in software engineering likewise concluded that, while AI can accelerate development and support proficiency, hidden vulnerabilities in generated code may increase cybersecurity risks and reduce net productivity gains (Negri-Ribalta et al., 2024[59]). Further, several studies suggest that efficiency improvements accrue mainly to lower-skilled workers, implying that for some expert tasks, AI may offer more limited benefits (Baily, Brynjolfsson and Korinek, 2023[54]). Taken together, these findings point to a more nuanced assessment of AI’s impact on performance, where potential time savings may be offset by quality control and security demands.

Overall, this aligns with findings that the productivity and innovation effects of AI appear heterogeneous and highly context-dependent, with potential gains sometimes constrained by quality, security and skill-related challenges (Filippucci et al., 2024[60]). While AI adoption can reduce marginal costs and improve scalability, its competitive implications depend on how such efficiencies are distributed across firms. In theory, lower costs should enable more firms to enter and compete. In practice, however, firms with established scale, data assets and cloud infrastructure may be better positioned to benefit, as they can more readily absorb high upfront investment in AI systems. This raises the possibility of asymmetric gains, where incumbents strengthen their position by leveraging AI to reinforce economies of scale and scope, potentially widening an “AI-divide” (Acemoglu, 2024[30]; Korinek, 2024[33]; Kergroach and Héritier, 2025[24]). In other words, AI may support more competition in some markets, but may also entrench market power in others.

2.1.4. Data access and downstream contestability

Access to high-quality data plays a central role in shaping both upstream and downstream dynamics in AI markets. At the upstream level, proprietary datasets and cloud infrastructure are widely recognised as essential inputs for training foundation models, in addition to compute and skills (OECD, 2024[2]; European Commission, 2023[61]). These models require vast amounts of high-quality data and significant computational resources, often only available through large-scale cloud platforms (OECD, 2025[3]; 2024[2]; Vipra and Korinek, 2023[62]). There is emerging evidence to suggest that access to these inputs is increasingly concentrated among a small number of firms, including major cloud providers and model developers, the so-called hyperscalers, who control both the infrastructure and the data pipelines necessary for large-scale training (OECD, 2025[3]; FTC, 2023[63]). This has been discussed extensively by other OECD work, but concern is emerging that this concentration may limit the ability of smaller firms or new entrants to develop competitive alternatives, particularly where access to proprietary datasets or compute is restricted or bundled with other services (OECD, 2025[3]; OECD, 2025[4]; OECD, 2024[50]; Villalobos et al., 2024[64]). That said, while access to proprietary datasets may confer advantages, recent analysis suggests that many data sources used for training AI models are non-unique and substitutable, limiting the potential for exclusionary control (Hagiu and Wright, 2025[65]). Moreover, many data licensing agreements are non-exclusive, which may help preserve contestability in downstream markets. The implications of upstream concentration for downstream markets depend on whether access to trained models and enabling infrastructure is broad-based, interoperable, and competitively priced.

Downstream, firms deploying AI systems often require access to operational or user-generated data to adapt and fine-tune models bought off-the-shelf to specific tasks, improve performance, and personalise services. In sectors such as finance, health, or retail, where such data are likely to be proprietary, incumbents may benefit from large volumes of interaction data, which can be used to fine-tune models and reinforce service quality (André et al., 2025[66]; Calvino, Reijerink and Samek, 2025[5]). Smaller firms or new entrants may face challenges in accessing comparable data, which can affect their ability to compete on accuracy, efficiency, or customer experience. These dynamics are particularly relevant where AI systems rely on continuous feedback to improve, and where data access determines the scope for innovation and differentiation.

The extent to which upstream concentration cascades into downstream markets then depends on several factors, including the availability of interoperable models, the portability of data, and the contractual terms governing access. Open data initiatives and public datasets can help reduce entry barriers, but the availability of such resources is uneven across sectors and jurisdictions. Technical and legal barriers to switching or multi-homing may reinforce dependence on dominant providers, especially where compute and data services are bundled (Vipra and Korinek, 2023[62]). Ensuring contestability in downstream markets may therefore require attention to data governance frameworks, interoperability standards, and the practical enforceability of access rights. A different point is the fact that AI might make data accessible and useful, both to those who hold the data, and also to those who might be given access to it. Making information more accessible reduces information asymmetries, which may in turn facilitate a level playing field, and reduce potential barriers to entry.

While further analysis is needed to assess the magnitude of these effects, data access likely remains a relevant parameter in evaluating the competitive impact of AI adoption across the value chain (OECD, 2024[2]; CMA, 2024[67]). From a competition perspective, relevant parameters to observe are likely to include the availability of interoperable data sources, the terms of access to proprietary datasets, the enforceability of portability rights, and the extent to which data access enables meaningful adaptation or differentiation of AI applications. Monitoring these factors may help assess whether upstream concentration risks translate into reduced contestability in downstream markets (OECD, 2024[2]; CMA, 2024[67]; Gupta, 2025[21]).

2.1.5. Model access and downstream differentiation

While foundation model development requires substantial data and compute resources, most downstream firms do not build their own models. Instead, they rely on access provided by a small number of upstream providers through APIs, plug-ins, or hosted platforms. The degree to which these models can be adapted or fine-tuned – sometimes referred to as “model openness” – influences how firms can integrate AI into their operations. However, it is not openness alone that matters for competition, but rather the practical ability of firms to access and tailor models to their specific needs. “API-gated” systems, i.e. closed models that interact through API endpoints, may offer convenience and reliability, but they can also limit visibility, control, and scope for differentiation. Conversely, semi-open or open-weight models may allow firms to fine-tune systems using proprietary data, enabling more targeted and differentiated applications and potentially supporting innovation in niche or sector-specific contexts (OECD, 2025[68]; Calvino, Haerle and Liu, 2025[7]; CMA, 2024[67]). The diversity of downstream use cases and modalities at present suggests that foundation models are unlikely to exhibit winner-takes-all dynamics across all sectors (Hagiu and Wright, 2025[65]).

From a competition perspective, the relevant parameters would include the extent to which downstream firms can customise models, the availability of interoperable alternatives, and the contractual or technical constraints on switching. The Portuguese Competition Authority has stressed that model openness is a critical determinant of contestability (AdC, 2024[11]). Where models remain closed and accessible only via APIs, downstream firms are likely to face dependence on provider terms (AdC, 2024[11]). This could potentially raise risks of refusal to supply, tying, self-preferencing or exploitative pricing, similar in effect as what has been observed in jurisprudence under Article 102 TFEU in Europe (European Commission, 2018[69]). Such risks could arise in cloud marketplaces (Model-as-a-Service), where hyperscalers act as both suppliers and gatekeepers (OECD, 2025[3]). The French Autorité de la concurrence notably has emphasised these risks, identifying cloud lock-in, exclusive data partnerships, and vertical integration as potential foreclosure channels in its Opinion (Autorité de la concurrence, 2024[70]).

While widespread access to general-purpose tools may compress productivity gaps, it may also standardise business processes and limit differentiation. Therefore, the ability to fine-tune models using firm-specific data could become a source of competitive advantage, particularly in sectors where domain knowledge or regulatory adaptation is important. However, such strategies often require complementary assets – such as compute, talent, and integration capacity – which may not be evenly distributed across firms. As such, competition authorities may wish to monitor whether access conditions, licensing terms, or ecosystem dependencies restrict downstream contestability, especially where switching costs or interoperability limitations arise (Krakowski, Luger and Raisch, 2022[26]; OECD, 2025[68]; Gupta, 2025[21]).

For SMEs and firms in less represented sectors or regions, downstream AI tools such as plug-ins, APIs, and pre-trained applications can lower barriers to adopting digital technologies and improve efficiency (OECD, 2025[68]; Calvino, Reijerink and Samek, 2025[5]). However, as will be discussed in section ‎2.3., uptake remains uneven as SMEs often lack the technical skills, infrastructure, or complementary assets to integrate and use these tools effectively, risking a widening AI capability divide (OECD, 2024[50]; Kergroach and Héritier, 2025[24]).

2.1.6. Search and switching costs

A final way that AI can support competition in downstream markets is by lowering lower customers and consumers’ search and verification costs. AI-enabled search, recommender and conversational systems filter, rank and personalise consumer options, which increases matching efficiency and the incidence of comparison across offers. This mechanism broadens effective choice sets and can foster or intensify competitive pressure on price and quality. Reviews of algorithmic recommendations explicitly note that such systems “save time and search costs”, while also cautioning about design choices that might counteract these gains (Calvano et al., 2024[71]). Building on formal consumer-search models, recent theory shows that personalisation creates value precisely by reducing search costs, drawing out higher-value (previously inactive) consumers and changing search breadth at the intensive margin (Rafieian and Yoganarasimhan, 2023[25]). Empirical work on AI-assisted information search similarly highlights that AI tools can reduce the time and effort required to locate and assess alternatives, facilitating more efficient consumer decision-making (Pham, Pham Thi and Duong, 2024[72]).

However, from a competition perspective, these systems may also create new frictions: by shaping consumer exposure through ranking or recommendation biases, they can channel demand towards certain suppliers, reducing effective transparency and potentially reinforcing incumbency advantages. The opacity of recommender algorithms further limits contestability if rival firms cannot ensure visibility under fair conditions, while consumers’ reliance on AI intermediaries may heighten switching costs or enable discriminatory steering based on inferred willingness to pay (Calvano et al., 2024[71]).

2.2. Limits to the usefulness of AI systems

While most of the findings discussed above point to a degree of cost reductions and productivity gains from adopting AI, with noted caveats, there are other, less positive, effects of the use of AI systems.

Several studies indicate that AI adoption may affect the quality and diversity of creative and cognitive work. Experimental evidence shows that the use of LLMs tends to homogenise creative outputs, reducing idea diversity and originality (Anderson, Shah and Kreminski, 2024[73]). In the context of marketing, AI-generated content has been found to be more repetitive and less engaging for consumers. Specifically, content generated by popular LLMs may appear more similar to each other compared to human created content, known as content homogenisation. This suggests diminishing returns on creative differentiation from using LLMs without human input (Liu, Wang and Yang, 2025[74]).

In scientific communication, LLMs have been shown to oversimplify and misrepresent research abstracts, thereby compromising accuracy and interpretative depth. This is especially problematic for scientific studies, since scientists must frequently include qualifications, context and limitations in their research results. Scientists discovered that some versions of popular chatbots were five times more likely to oversimplify scientific findings than human experts in an analysis of 4 900 summaries of research papers (Peters and Chin-Yee, 2025[75]). Moreover, AI-powered chatbots, including LLMs, are prone to generating inaccurate or fabricated information – a phenomenon commonly referred to as "hallucination". These errors can arise even when the model appears confident and fluent, undermining its reliability in high-stakes or information-sensitive contexts. While ongoing research aims to reduce such occurrences, the persistence of hallucinations poses challenges for trust, accountability, and safe deployment in both public and private sector applications (OECD.AI, 2023[76]).

Cognitive research further shows that reliance on AI tools can weaken users’ analytical and critical-thinking capacities over time: a recent study revealed a significant negative correlation between frequent AI tool usage and critical thinking abilities, owing to increased cognitive offloading (i.e. outsourcing parts of one’s thinking or memorising). Younger participants exhibited higher dependence on AI tools and lower critical thinking scores compared to older participants (Gerlich, 2025[77]). While deep integration with AI (i.e. actively interacting with the AI to generate content) has been found to foster new skill acquisition, passive use, such as delegating and automating tasks, might erode knowledge. Similar studies have found that delegating writing tasks fully to an AI system can reduce neural, linguistic and behavioural performance (MIT Media Lab, 2025[78]). Indeed, an overreliance on AI could erode human knowledge over time (Baily, Brynjolfsson and Korinek, 2023[54]). On the other hand, productivity gains are maximised when AI is deployed to well-defined and bounded tasks, such as coding, writing and summarising, and integrating AI as a complement rather than a replacement for human skills (Calvino, Reijerink and Samek, 2025[5]).

The next section takes a brief look at the way sectors and tasks are exposed to AI, with potential ramifications for AI-use.

2.3. Sector exposure and uptake of AI vary significantly

To help identifying sectors that are more likely to see the largest impact from AI adoption, this section looks at potential sector exposure to AI based on types of occupation, followed by a discussion on observed AI uptake.

2.3.1. Sector exposure to AI is high for professional services

As discussed above, exposure to AI automation or substitution varies significantly across sectors, one of the reasons for the Baumol effect mentioned earlier. A large study by (Eloundou et al.[12]) investigates the potential labour market impact of GenAI, by assessing how exposed various occupations and tasks are to LLM capabilities.9 The authors identify exposure as the extent to which LLMs can reduce the time required to complete a task by at least 50% without compromising quality; and automation as the extent to which LLMs (and LLM-powered software) can fully or mostly perform a task without human oversight. This is relevant as a potential gauge of where AI-adoption may lead to disruption or structural changes that affect the competitiveness of market players.

The Eloundou study finds that a large share of United States workers (80%) are in occupations with at least 10% of tasks exposed to LLMs, and 18.5% of workers are in occupations where over 50% of their tasks are exposed to LLMs. The most exposed occupations include professional and business services, including mathematicians, writers and authors, tax preparers and auditors, financial quantitative analysts, web and digital interface designers and coders, whereas exposure is lowest in manual occupations, including cement masons, roofers, meat packers, bus mechanics and dishwashers.10

Concurrent with the findings in Section ‎2.2, the tasks with high exposure also correspond to those listed above where the use of GenAI tends to compress the duration for routine tasks, reduce the need for some types of labour and lead to cost reductions or higher productivity, potentially therefore lowering barriers to entry. Such effects are already visible in areas like software development, legal and accounting services (professional services), and in customer service and business process outsourcing, including supply-chain management (Peng et al., 2023[18]; Cui et al., 2024[19]; McKinsey, 2025[79]; André et al., 2025[66]; Calvino, Reijerink and Samek, 2025[5]; Brynjolfsson, Li and Raymond, 2023[15]; Rajendran, Balaraman and Viswanathan, 2025[80]). 11

Sectors relying on physical or on-site tasks, such as maintenance or logistics operations, remain less affected by AI. While predictive-maintenance systems in manufacturing use AI to forecast failures and optimise parts procurement, the machines still depend on human technicians for diagnosis and repair (Santamaria, 2024[39]; McKinsey & Company, 2024[40]). In logistics and retail, AI modules can assist small firms in demand forecasting, stock management and routing, significantly increasing throughput and effectiveness and boost competitiveness, even though staff are still required for handling and delivery functions (Dicey, 2025[81]; Nishar, 2024[82]; Shen et al., 2025[83]).12

Differences in task exposure may shape the potential for GenAI to alter market structures and support competitive outcomes. As such one may argue that sectoral exposure to GenAI is a key determinant of the likely competitive impact from AI adoption. In sectors where a large share of tasks can be automated or augmented, competing firms may in principle replicate the capabilities of incumbents by adopting generative-AI tools, provided that access costs for data, models and computing resources remain manageable (Kergroach and Héritier, 2025[24]). The ability to translate automation potential into genuine contestability however also depends on whether new entrants can overcome other structural barriers such as brand recognition, distribution channels, or regulatory complexity.

Hence, AI’s potential to reshape competition is contingent on both technological exposure and institutional context. The creative industries illustrate this tension: although writing, editing, design and composition tasks are among the most exposed to GenAI, incumbents can retain advantages through control of intellectual-property assets and established reputations, while legal uncertainty surrounding copyright and text-and-data-mining frameworks might discourage entry by smaller creators (Lucchi, 2025[84]; Marcelin and Cassetti, 2025[85]).

2.3.2. AI uptake is rising

This section reviews some of the available evidence on the uptake of AI. As the use of AI (and the technology itself) are constantly evolving, the data are illustrative rather than authoritative.

Confirming the discussion above, knowledge-intensive services are consistently among the most advanced AI users: in 2024, 44% of firms in the information and communication sector in the OECD area reported using AI, with figures above 67% in Denmark, Sweden and Finland (Kergroach and Héritier, 2025[24]). Professional, ICT, finance and scientific services showed an average of 26% adoption across the OECD, while utilities reported similar levels, with much higher figures in northern Europe. Other sectors that have seen a widespread uptake of AI are sales, marketing and customer-facing functions (MIT, 2025[45]). By contrast, adoption remains low in food products, textiles, construction and hospitality (Kergroach and Héritier, 2025[24]; OECD/BCG/INSEAD, 2025[27]).13

Firm size remains one of the strongest predictors of AI use. In 2024, 39% of large firms in the OECD area reported using AI, compared with 12% of small firms; a 3.3-fold difference (Kergroach and Héritier, 2025[24]). Larger firms are also more likely to report redesigning workflows and embedding AI into multiple business functions, from IT and marketing to service operations (McKinsey, 2025[79]).

Adoption is often limited by a lack of digital and data maturity, and uncertainty about return on investment (OECD/BCG/INSEAD, 2025[27]). Skills shortages are cited for high-tech industries such as computer and electronics manufacturing, utilities, telecommunications and travel agencies. Lack of expertise, incompatibility or lock-in with existing legacy systems, and high costs are also cited as reasons for not adopting AI. Legal uncertainty is also increasing, with more firms reporting difficulties understanding AI-related obligations. MIT reports that over 80% of organisations have piloted general-purpose tools, but only 20% reached pilot stage, and just 5% of AI systems that are trialled and fully deployed. This gap has been described by MIT as the “GenAI Divide”, reflecting widespread experimentation but limited transformation within firms (MIT, 2025[45]). Difficulties in accessing quality data and compliance with privacy rules are also frequently cited (Kergroach and Héritier, 2025[24]).14 AI-related risks related to cybersecurity, intellectual property and data governance are also holding back uptake, even in AI-advanced sectors like professional services and publishing (Kergroach and Héritier, 2025[24]; McKinsey, 2025[79]).

AI-use may however be starting to alter the market structure in selected industries, with potential competitive ramifications. MIT (2025[45]) found that two of nine surveyed sectors (technology and media) have experienced structural change attributable to GenAI, such as shifts in recruitment, business models or market leadership. In other industries however, adoption remains incremental, focused on efficiency gains rather than disruption. Nonetheless, surveys suggest that early adopting firms that invest in customised, workflow-integrated AI systems are gaining a competitive advantage. These include revenue growth of 3%–15% and shorter times to market (OECD/BCG/INSEAD, 2025[27]; MarketsAndMarkets, 2025[41]). Furthermore, a survey carried out by LinkedIn/Ipsos amongst sales forces shows that 73% of sellers believe they will be left behind without an AI strategy, and 75% of top-performing sellers are eager to adopt AI agents (LinkedIn/Ipsos, 2025[55]).

Finally, new forms of so-called “shadow AI” use are emerging: While only 40% of companies report purchasing official LLM subscriptions, employees in more than 90% of surveyed firms report regular use of personal AI tools for work tasks (MIT, 2025[45]). This informal uptake suggests that adoption is often moving faster at the individual than at the organisational level, and may in time accelerate enterprise deployment. India seems to be following these trends (Box 6).

Box 6. AI uptake and diffusion in downstream sectors in India

India’s AI market has nearly doubled in value since 2020, growing from USD 3.2 billion (2020) to USD 6.1 billion in 2024. AI adoption has seen a rapid expansion across downstream sectors, driven by government initiatives, startup activity, and increasing private sector investment. The IndiaAI Mission, launched in 2024 with funding of around INR10 300 crore (USD 1.2 billion), aims to expand national compute capacity, create a shared dataset platform and promote indigenous AI models. These measures aim to reduce reliance on a few global providers and make advanced AI infrastructure accessible to smaller firms.

Adoption has been fastest in IT and digital services, financial technology, manufacturing and agriculture. Large Indian firms are scaling from pilot projects to production-level AI use in process automation, analytics and customer engagement, while public-private initiatives in agriculture are using AI to forecast weather, soil and pest conditions through mobile-based advisory systems. The Competition Commission of India (CCI), in its market study on AI and Competition (September 2025), notes that Startups dominate the downstream layers, especially in application development, with 67% of surveyed firms focused on building AI-powered solutions. They are helping democratise AI adoption across sectors while navigating challenges like infrastructure costs, data access, and ecosystem lock-in.

AI appears to be already producing pro-competitive effects in selected niches. By enabling lower-cost predictive analytics, demand forecasting or customer personalisation, AI tools allow smaller firms to enhance operational efficiency and differentiate their offerings. The IndiaAI Mission’s stated objective is precisely to “democratise computing access” and support indigenous innovation, thereby enabling smaller players to participate more fully in the AI ecosystem. Open or shared compute infrastructure and dataset platforms can serve as common inputs, helping to level the playing field if governed on fair, transparent and non-discriminatory terms.

However, AI-diffusion remains uneven. Most adoption is concentrated among large firms in major cities, while smaller enterprises face barriers related to skills, infrastructure and data access. These gaps risk widening productivity and competitiveness divides. At the same time, concentration in compute resources, cloud services and foundational models may allow a few firms to shape downstream access, creating dependencies and potential foreclosure risks. The CCI has highlighted issues such as preferential access to AI inputs, bundling and tying, as areas requiring close monitoring.

The CCI market study notes that competition risks in the AI ecosystem are multifaceted and evolving. It further highlights concerns about market concentration, particularly the dominance of large technology firms across multiple layers of the AI stack, data, infrastructure, and foundational models, which can create high entry barriers for startups. The market study identifies risks such as algorithmic collusion, price discrimination, self-preferencing, and opaque decision-making, which may distort market outcomes and reduce consumer choice. It also warns of ecosystem lock-in, where users become dependent on specific platforms due to high switching costs. These dynamics, if unchecked, could entrench incumbents, stifle innovation, and undermine fair competition.

Sources: Press Information Bureau (2024), IndiaAI Mission approved by Cabinet, https://www.pib.gov.in/PressReleasePage.aspx?PRID=2178092; IndiaAI, About the Mission, https://indiaai.gov.in/; ISAS (2025), AI Adoption in India: Moving the Needle Forward, https://www.isas.nus.edu.sg/papers/ai-adoption-in-india-moving-the-needle-forward/; Competition Commission of India (2025), Market Study on Artificial Intelligence and Competition, https://www.cci.gov.in/images/marketstudie/en/market-study-on-artificial-intelligence-and-competition1759752172.pdf

2.4. The emerging case of Agentic AI

Agentic AI is an emerging case, referring to a developing generation of systems that integrate memory, planning, and tool-use to pursue goals with a high degree of autonomy. Unlike traditional LLMs that merely respond to prompts or generate content, agentic systems can initiate actions, interact with external systems, and adapt their behaviour over time, potentially operating independently in real-world or digital environments. These applications act as agents for users, powered by foundation models (FMs), but with substantial decision-making autonomy. Though still nascent, agentic AI could significantly reshape competition in areas such as search, professional services, workflow automation, and assistant tasks (e.g. travel booking, agenda management, procurement, or sales). Agentic AI differs from traditional AI tools because it combines reasoning, planning, memory, and the ability to act autonomously across digital environments, often through multi-agent frameworks (Derouiche, Brami and Mazeni, 2025[86]; Nisa et al., 2025[87]). The OECD has highlighted the importance of monitoring these frameworks closely, given their potential to reshape market structures and regulatory frameworks (OECD, 2024[50]).

These advances could mark the beginning of a broader shift towards the “Agentic Web”, characterised by autonomous, goal-driven interactions where AI agents interact directly with one another to plan, co-ordinate, and execute complex tasks on behalf of users (Yang et al., 2025[88]). For instance, instead of generating text or images for a social media post, an AI Agent would post the content directly across a host of social media outlets (Desai and Riedl, 2025[89]). This transition from human-driven to machine-to-machine interaction would eventually allow users to delegate routine digital operations entirely to AI systems, potentially transforming online services and business models.

Although still in their infancy, Agentic systems may also be able to further reduce consumer search and transaction costs, discussed in 2.1.6. By aggregating information across multiple providers, these systems could potentially support multi-homing and switching, supporting competition by weakening incumbents’ ability to exploit behavioural biases and consumer inertia (OECD, 2023[90]). Agentic AI may enhance contestability by enabling smaller firms to deploy AI “agents” to co-ordinate workflows, search for information, and interact with customers without requiring extensive technical resources or in-house expertise (Yang et al., 2025[88]; Sapkota, Roumeliotis and Karkee, 2025[91]). Such usage could allow new firms to scale faster and challenge incumbents in areas such as professional services, digital commerce, and back-office operations. However, many agentic AI systems developers rely on cloud integration with the world’s three hyperscalers (OECD, 2025[3]), implying a vertical link for downstream users, which could potentially hamper competition over time.

2.5. Conclusion

While comprehensive data on indicators such as entry and exit or firm churn remains limited, the evidence reviewed in this chapter suggests that AI adoption in downstream sectors may support competition. In particular, AI can reduce average and fixed costs, substitute for labour in routine tasks, increase efficiency, and enable new forms of innovation. Firms may benefit where automation frees employees to focus on higher-value activities and reduces reliance on outsourced services.

However, these effects are not automatic and may vary substantially across sectors, firm sizes, and use cases. Smaller firms could gain from modular AI tools and task-level automation, but larger incumbents often benefit from structural advantages – such as greater access to data, infrastructure, and integration capacity – which may allow them to capture disproportionate gains. The competitive impact of AI therefore depends not only on technological capabilities but also on market conditions, regulatory frameworks, and access to enabling inputs. Competition conditions upstream, including in markets for compute, model design, and especially data access (CMA, 2023[92]; Autorité de la concurrence, 2024[70]; OECD, 2025[3]; ACCC, 2025[93]) also influence competition downstream.

Notes

← 1.  Field experiments at major firms likewise report that less experienced developers achieve disproportionate productivity gains when supported by AI assistants (Cui et al., 2024[19]).

← 2.  OECD surveys show that smaller firms in digitally intensive sectors have higher adoption rates of AI than less digitalised firms, and that adoption often correlates with entry and survival in competitive markets (Calvino, Haerle and Liu, 2025[7]).

← 3.  One estimate for enterprise-grade implementations of AI systems for customer support says they typically require between USD 100 000 and USD 2 million in initial expenditure, including platform setup and licensing fees. However, depending on the volume of customer queries (in this estimate, the volume was around 55 000/year), these costs may be recovered within a relatively short period of around six months.

← 4.  Business Process Outsourcing (BPO) refers to the practice of contracting specific business functions or processes to external service providers, often in another country, to reduce costs and improve efficiency. These outsourced processes are typically non-core but essential operational tasks, such as customer support, data entry, payroll processing, or IT services. For example, a telecommunications company might outsource its customer service operations to a call centre in the Philippines or India to benefit from lower labour costs and 24/7 coverage. Similarly, a bank might outsource its back-office functions, such as loan processing or compliance checks, to a specialised service provider that can perform these tasks more efficiently at scale.

← 5.  Randomised experiments show substantial time savings and quality improvements on mid-level professional writing tasks when workers use a general-purpose assistant, with the magnitude varying by task and baseline skill (Noy and Zhang, 2023[17]) (OECD/BCG/INSEAD, 2025[27]).

← 6.  In a natural experiment, GenAI reduced the average price of commissioned digital artwork by 64%, while increasing order volumes by 121%, indicating a downward shift in marginal and average production costs as automation replaced labour-intensive creative tasks (Zhang, Yuan and Xiong, 2023[160]). The authors conclude that using AI can shift the supply curve outward, reducing average operational costs.

← 7.  It should be noted that these findings are survey based. While AI has potential to improve a vast array of uses, including forecasting included, it still makes mistakes. Some evidence seems to suggest, for example, that it might be better at forecasting modal or common events, for instance related to the weather, but not so good at extremes. https://arxiv.org/abs/2508.15724

← 8.  Nubank, a digital financial services platform now covering Brazil, Mexico and Colombia, has used generative AI copilots to automate internal compliance and expand customer service capacity through rolling out a call-centre pilot, allowing it to compete with established banks despite limited physical infrastructure (Ashton, 2025[133]). Nubank’s AI-powered assistant can resolve 55% of initial enquiries without escalating to a human agent, handling over 2 million monthly chats and reducing chat response time by 70%, responding 24 hours a day (OpenAI, 2024[134]).

← 9.  The authors introduce a systematic methodology to evaluate exposure and automation potential across United States occupations. While the study focuses on the United States, the findings related to task automation and AI exposure is likely to be widely applicable across other economies, and relevant for the discussion about AI-adoption.

← 10.  (Eloundou et al., 2024[12]) find that tasks requiring programming or writing show the highest exposure to AI automation, whereas those requiring scientific reasoning, critical thinking and active learning remain less susceptible. Education-wise, higher exposure was found in jobs requiring bachelor’s or master’s degrees, whereas jobs requiring extensive on-the-job training continue to depend primarily on human skills: “Jobs with no on-the-job training or only internship or residency required appear to yield higher income but are more exposed to LLMs.” (Eloundou et al., 2024, p. 34[12]).

← 11.  Other occupations include education and training: generative AI tools are increasingly used to automate lesson preparation, grading support and content generation, illustrating how exposure does not necessarily imply full automation but rather task augmentation. In India, for instance, small education-technology firms draw on publicly available or firm-owned materials to generate course content and adaptive assessments, substantially reducing production costs and time-to-market (Digital Futures Lab, 2022[145]). These developments align with (Eloundou et al.[12])’s findings that teaching and writing-intensive occupations are highly exposed to LLM capabilities, yet the technology primarily complements rather than replaces skilled educators for now.

← 12.  In logistics and retail, operators integrate artificial intelligence into demand forecasting, stock management and routing through cloud modules, enabling small firms to replicate efficiencies once available only to incumbents with in-house analytics. ShopriteX in South Africa has scaled its grocery delivery services through AI-enabled logistics optimisation (Dicey, 2025[81]; Nishar, 2024[82]).

← 13.  The study was published in 2025, but the survey was conducted in 2022-23.

← 14.  MIT also identifies additional organisational bottlenecks, including poor workflow integration, lack of memory and learning in AI systems, and low tolerance from users for “static” tools compared to flexible consumer LLMs (MIT, 2025[45]).