Intersting report from OECD. I focussed here in diffusion, i.e. the spread and systemic adoption of AI. With an eye on the eventual improvements in productivity effective diffusion is essential. The full report can be found at end of post..
For Bankers, consider this quote. This is the beginnings of the holly grail to address the elephant in the room of Banks data worries.
Applications can transform customer relationships and marketing, by improving market segmentation and raising capacity to “Know Your Customer” (KYC), by automating basic customer services and enabling a 24/7 attendance, or by supporting content creation at lower cost, such for communication or advertising.
First the OECD view of AI:
What AI technological change entails
Exploring trends in AI diffusion first requires to understand the fundamental basics of AI technology change. AI is data-driven, and essential assets include “soft” assets, i.e data, algorithms and skills, and “hard” assets, i.e. AI compute infrastructure (Figure 2).
• AI is trained by processing large amounts of data to identify patterns and infer responses. In general, the larger the volume of higher quality and relevant data, the better, although in some cases lean models can perform just as well as those trained on large amounts of data. There is a great variety of data, often categorised by ownership (e.g. personal, public or proprietary data) or by sourcing context (e.g. user, consumer, supplier or taxpayer data, process, industrial or government data, or product, technical, multimedia or communication data) (OECD, 2022[25])
(OECD, 2022[26]).
• Algorithms are step-by-step instructions that guide AI, with different degrees of complexity and supervision, on how to process this information. Once trained, the AI model can perform based on new data, and can keep improving with more training, better data and human feedback for reinforcement learning (E.U. and U.S Trade and Technology Council, 2023[27]) (OECD, 2024[28]).
• Advanced statistical, mathematical or programming skills are required for developing AI models. But non-AI, expert and soft skills, such as domain knowledge, business acumen, critical thinking and communication, are important to train the machine, interpret results and ensure a trustworthy and ethical use of AI. In addition, AI diffusion requires public acceptance, a collective data culture and digital security literacy (OECD, 2022[25]) (OECD, 2024[28]).
• AI infrastructure (“AI compute”) is thecomputational power required to train and run AI models. AI computing resources include one or more stacks (layers) of hardware and software used to support specialised AI workloads and applications, i.e. computing power (e.g. processors), data pipelines and storage facilities (e.g. data lakes, warehouses), high-speed networks and supporting software and platforms. AI compute requirements can vary significantly, depending on the application, AI system lifecycle stage, or size of the system (OECD, 2023[29]).
AI technological change remains complex to implement, because of the diverse technologies andapplications it entails, the diversity of soft and hard assets it requires to perform, and because the frontieris in constant and fast evolution.
Scandinavia is a stand out
Lagging places, sectors and firms may not be able to catch up
Technology diffusion is not homogeneous and universal at inception. Recent research shows that the adoption of general purpose technologies (GPTs) – such as computers, the internet and electricity – follows an initial period of acceleration and slows when a saturation point is reached in demand (Filippucci, Gal and Schief, 2024[54]) (Filippucci et al., 2024[55]). AI presents the features of a GPT for its pervasive impact across all segments of the economy, capacity for continuous improvement and potential to steer innovation and innovation complementarities at large scale (Calvino, Haerle and Liu, 2025 forthcoming[31]).For AI, firms who started sooner their journey are therefore likely to be on this accelerating path, which would lead to a natural rise of disparities in diffusion.
Delayed technology adoption is however hard to overcome. Early adopters of a new technology capture benefits that increase exponentially as the technology becomes mainstreamed. First (or second) mover advantages arise from the fact that they have already set industry standards, built reputation, and consolidated their markets and networks, when late adopters come in. For digital technologies in particular, network effects can be at play that increase the value and return on a technology as its number of users increases (e.g. matchmaking platforms). For later adopters and laggards, the potential benefits and returns
on technology investment decrease rapidly. This is due to market saturation by incumbent competitors, higher costs to switch away from obsolescente technologies, or the unbreakable advantages gained by first and second adopters.….
Although it is still too early to say how competition conditions will evolve with the rapid deployment of AI, and even if few network effects have materialised so far, risks of algorithmic collusion and the strong AI asset concentration raise concerns for future business environment (OECD, 2024[56]) (Kergroach, 2025 forthcoming[1]). Risks for competition could in particular be amplified in specific sectors, e.g. sectors where market concentration is already high, or where early AI adopters can increase and consolidate their market shares, but at the same time bring new competition in currently highly concentrated markets, such as the ICT sector (OECD, 2025 forthcoming[57]). Indeed, an early majority of AI adopters is emerging across regions, sectors, and firms, with likely decreasing returns on investment for next adopters (Figure 7) (Figure 8) (Figure 9).
AI can unlock substantial corporate efficiency gains
AI can transform businesses, their internal processes and cost structure, ultimately unlocking productivity and quality gains (Box 5). Table 1 shows how AI can be applied along the internal value chain of the firm, and improve operations from pre-production to post-production. Predictive and generative AI increase capacity for business decision making and corporate planning, budgeting and management, by reinforcing knowledge management and business intelligence systems, supporting scenario analysis and forecasting, and by automating central corporate functions, such as accounting, financial affairs, or human resources (HR). New AI applications are also found for IT security.
Face recognition based on computer vision is used for authentication of ICT users, or machine learning for better and faster detection and prevention of cyber-attacks. Performance and optimisation gains can be achieved in supply chain management, logistics and inventories, or production lines and workforce planning. AI helps predict and anticipate disruptions, shortages or maintenance, and reduce losses, uncertainty and volatility (e.g. in price and procurement).
Applications can transform customer relationships and marketing, by improving market segmentation and raising capacity to “Know Your Customer” (KYC), by automating basic customer services and enabling a 24/7 attendance, or by supporting content creation at lower cost, such for communication or advertising.
AI is finding promising uses in R&D and design by raising significantly corporate capacity for data analytics and experimentation, and by lowering the costs and risks of innovation. Simulations, prototyping and a parallel improvement in market knowledge and corporate business intelligence could indeed reduce the time to market and facilitate the organisational changes needed (e.g. in production or workers training).
Bankwatch 