How we measure who pays for AI
Our estimates come from de-identified UK bank and card records. This page explains what we observe, how we find AI payments, and what the numbers mean.
The data
We use de-identified UK consumer bank and card records from a financial data aggregator. The data cover 14.2 million accounts and about 8 million transactions a week between January 2022 and August 2026. For each payment, we see the date, amount, merchant category and statement text, alongside the account holder's age band, sex and postal area. We never see names or account numbers. To exclude dormant and secondary cards, we keep only active main accounts with at least 20 non-AI payments a month and median non-AI spending of at least £100 a month over the previous year.
Finding AI payments
We keep a registry of more than 300 standalone AI products, led by the nine main general-purpose assistants: ChatGPT, Claude, Gemini, Microsoft Copilot, Perplexity, DeepSeek, Mistral, Grok and Poe. A direct payment counts when the statement text, exact price and product launch date all match. We exclude software that added AI to an existing product, such as Microsoft 365, Canva or Adobe. Privacy filters sometimes replace the word Claude with a placeholder. When that happens, we match the web domain and the price. In hand-checked samples, 99.9% of the payments we flag are AI payments.
Payments through app stores
Subscriptions bought inside the ChatGPT or Claude mobile apps appear on statements as Apple or Google app store charges, without the product name. We estimate them in two steps. First, we measure the rise in recurring app store charges at the exact prices of AI plans after each product launch, compared with prices that no AI product uses. Second, we assign those extra charges to accounts by age, sex, region, income and spending, based on paid AI users who make direct payments to AI providers. The lower bound counts direct payments only. The best estimate adds the app store estimate.
Representing UK adults
The accounts in our data do not mirror the population exactly. Each month, we weight them to match official figures from the Office for National Statistics and HM Revenue and Customs on age, sex, region, income and neighbourhood type. We express rates per 1,000 of the 55.5 million UK adults aged 18 and over.
Earnings and occupations
For open-banking current accounts, we observe salary payments, which show take-home pay and, through the employer's name, the industry. Bank records do not state occupation. We identify it from payments to professional regulators, royal colleges, chartered institutes and occupation-specific trade unions, such as the General Medical Council or the University and College Union. We also use employers whose staff mostly do one job.
What the numbers mean
We measure subscriptions people pay for themselves. The data do not include licences that employers buy for staff. A rate of 50 paid AI users per 1,000 adults means that, after weighting, 5% of adults held an active paid AI subscription that month. We count accounts, and some people hold more than one. The paper's appendix examines this.
Privacy and disclosure
We work only with de-identified records on secure servers. We publish aggregates only, and no published figure rests on fewer than 10 accounts.
The paper's data section and appendices give the full method, the validation against official statistics, and robustness checks. Paper