The Office for National Statistics (ONS) in the UK is preparing to undertake a task that may seem tedious, but could actually have an impact on investment decisions: to isolate artificial intelligence (AI) from the vast national economic accounts and create a set of “AI thematic accounts.” This is not about attaching another fancy label to the AI industry, nor is it about immediately releasing figures such as “AI accounts for GDP.” According to the methodological document released on September 21st, the ONS’s first steps are to address issues related to boundaries, data, and attribution, with the goal of establishing experimental accounts that can be put into use by March 2028. Officials have specifically noted that this work is still in the methodological development phase, and the document itself does not constitute official statistics.
Over the past two years, almost every company has been talking about AI, but macro data struggles to answer the most basic questions: How much AI product has the UK actually produced? How much money have companies invested in deploying AI? And in which industries has this investment created added value? The current system of national accounts is adept at tracking software, hardware, consulting, and research and development, but it cannot automatically distinguish the portion that belongs to AI just because a piece of software suddenly adds generative capabilities. As a result, while the market witnesses a boom in investments in large models, chips, and data centers, it is difficult to find a clear correlation between these investments and productivity, capital formation, and industrial added value.
From models to consulting services, the industry boundaries of AI are broader than imagined.
The Office for National Statistics in the UK adopts an approach that is aligned with the 'System of National Accounts 2025'. It plans to compile statistics not only on model companies but also on the hardware, software, and infrastructure that support the operation of AI, professional services related to strategy, deployment, and security, as well as activities in various industries that actually utilize AI. In other words, a chip designed for training, a set of cloud-based inference services, a consulting firm that helps banks transform their processes, and a predictive system developed internally by retail enterprises could all find a place in this statistical account.
Troubles also begin from here. Many products embed AI into their existing functions, and the price does not separately list the “AI component”; a multinational company may have models in the United States, train or sell services in the UK, with revenues, intellectual property, and employees distributed across different jurisdictions; companies also have different understandings of “adopting AI”, ranging from rule automation to generative models, which might all be included in the same answer. If the boundaries are too broad, accounts may include ordinary software and automation as well; if the boundaries are too narrow, a large number of internally developed products and industry-specific applications will be overlooked.
Therefore, the authorities have proposed a three-tier observation framework for production, usage, and the production chain. On the production side, it is necessary to identify entities that directly provide AI goods and services; on the usage side, it is essential to estimate the scale of AI purchases or self-builds in other industries; while the production chain tracks investments in chips, computing, data, and professional services. The data from these three layers must be consistent with the existing national accounts, and no value should be calculated repeatedly for the same amount. For investors, this is more useful than simply listing the number of AI companies: only by distinguishing between intermediate inputs and final outputs can one determine whether the boom has actually resulted in added value or is mainly reflected as procurement expenditures between enterprises.
The source of data will not be a one-size-fits-all questionnaire either. The UK Office for National Statistics is considering combining corporate surveys, administrative records, recruitment and job data, company accounts, and other business data, and then developing estimation methods for the parts that cannot be observed directly. Corporate confidentiality restrictions can affect the granularity of the data by industry, small sample sizes and extreme values can cause fluctuations, and the speed at which product classifications are updated may not keep up with technological changes. The final figures released will likely require ranges, revision notes, and methodological explanations, rather than a seemingly precise total that cannot be verified.
It will only be in 2028 that experimental accounts are handed over, which precisely indicates that this matter cannot be achieved by mere slogans.
According to the roadmap, in the first quarter of 2027, the main focus will be on scope definition, user requirements, and method design; from the second to the fourth quarter of 2027, prototypes will be developed and tested; by the first quarter of 2028, an experimental thematic account ready for production will be established, with a target date of March 2028. The key word here is "experimental." This indicates that the data and methods may still be adjusted, and it does not mean that the UK has already fully calculated the contribution of AI to GDP.
Why is it nearly a year and a half needed? Because the economy of AI is not something that exists naturally, just waiting for statisticians to discover. Models, computing power, software, and services are highly interrelated; there is no market price for internal corporate development, and cross-border intellectual property rights can change the ownership of value. If statistical agencies pursue speed, the most common mistake they make is to directly consider financing amounts, company valuations, or software revenues as economic contributions. The significance of thematic accounts is precisely to separate these metrics, allowing for comparisons between different years, industries, and countries.
This set of accounts may also change the discussion about productivity in the future. Whether the AI tools truly save working hours, whether the costs saved by enterprises are translated into higher output, and whether the related benefits are concentrated among technology suppliers or spread to the industries that use them cannot be answered solely by case studies. If thematic accounts can link investment, employment, added value, and import dependence, they could help the market discern whether there is a time lag between the speed of adoption and economic returns.
Public attitudes also explain why officials cannot simply focus on measuring scale alone. A survey cited by the UK National Statistics Office shows that people have significantly more trust in the use of AI for scientific research, customer support, or creative tasks than in government decision-making, healthcare, or major corporate decisions. Economic value and social acceptance do not follow the same trajectory. Even if an industry rapidly deploys AI, it may incur additional costs due to requirements related to responsibility, privacy, and reliability.
For businesses, the most practical preparation at this moment is not to guess a national AI output value, but to keep clear records of their investments: what computing power and software has been purchased, which personnel are involved in model development, how much revenue or cost changes have been brought about by AI functions, and who owns the intellectual property rights. Once the future statistical criteria are stabilized, these records will determine whether a company's activities will be noticed. This time, the UK did not rush to present a shocking figure; instead, they first acknowledged why such figures are difficult to obtain. For the AI industry that is transitioning from storytelling to infrastructure, this is actually a more mature sign.
