Consumers paying for generative AI services often switch or cancel subscriptions, according to an October study of US card transactions. The pattern gives rival providers an opening to attract customers, while making recurring revenue less predictable.
Researchers Mark Egan, Dominic Russel, Claire Shi and Vincent Yao examined payments to four major providers: OpenAI, Anthropic, Google and xAI. In April 2026, 2.9% of consumers in their baseline data paid for at least one of the identified services.
The National Bureau of Economic Research working paper estimates annual customer retention between 20% and 45% across the providers. Switching increased around major model releases, and most paying consumers used entry-level plans.
The card records do not capture free users or enterprise accounts. The findings describe the observed US consumer-payment sample through April 2026, not current worldwide usage. The working paper has not been peer reviewed.
What card payments can show
The researchers used anonymized Fiserv credit and debit card data from January 2021 to April 2026. The original dataset covered 174 million consumers and 114 billion transactions. To make it more likely that a person’s primary card was observed, the baseline analysis restricted observations to individual months with at least 20 transactions, reducing the sample to 82 million consumers.
The 2.9% figure should be read against that filtered sample and its method of identifying purchases, not as a count of everyone who has tried an AI chatbot. It excludes people using free tiers and services paid for through enterprise arrangements.
Paid adoption was higher among younger and higher-income consumers in the data. OpenAI accounted for the largest observed share, while the paper also reports rapid growth in Anthropic payments after a model release.
Transactions can show that a charge occurred, but not precisely why a customer chose a tool, whether someone used a free account elsewhere, or how satisfied they were with an answer. The researchers combine the observed patterns with a model to estimate demand and substitution among services.
Retention and switching affect the business case
Retention measures the proportion of a group of paying customers still subscribed after a specified period. A low annual rate can make growth dependent on continually attracting new customers, although customers may return or use other payment channels.
Some consumers paid for more than one AI service, a practice called multihoming. The paper says this became less common as providers’ capabilities expanded. Its model suggests customers became more willing to substitute one service for another.
Those patterns are not proof that all AI products are interchangeable. A customer may need a particular feature, integration, privacy arrangement or model capability that is missing elsewhere. The paper’s conclusions concern the products and consumers it could observe.
Our earlier reporting examined how AI agents are changing business software pricing. Enterprise contracts can combine per-user fees and usage charges; the new research instead focuses on consumer subscriptions paid by card.
Price is only one cost of changing tools
For individual buyers, a lower monthly charge matters alongside the work needed to move saved conversations, files or workflows. Businesses evaluating AI products also have to consider security, administrative controls and the cost of training employees, none of which can be inferred from a consumer’s card payment.
The researchers’ estimated model points to substantial price sensitivity and competitive pressure. Their forecast about future pricing power depends on how products evolve and on the model’s assumptions, rather than guaranteeing future price cuts.
The paper discloses that Yao is a paid Fiserv consultant and both Yao and Egan serve on its advisory council without pay. Yao states that Fiserv did not direct the study’s design, analysis or publication.
For providers, new model launches may bring opportunities to win subscribers from competitors. Whether those customers stay will depend on the usefulness of the product, its price and the alternatives available after the initial launch.
