AI Subscribers Are Still Shopping Around, Harvard Study Finds

New research shows weak retention among paid AI services as customers experiment with rapidly changing technology, raising questions about how providers can build lasting subscription businesses.

People are paying for artificial intelligence tools, but many aren't sticking with one service for long.

A new Harvard Business School working paper found that 12-month retention among major paid AI services ranged from roughly 20% to 45%. OpenAI performed best, with approximately 45% of customers still paying a year after their first observed purchase. Anthropic, the company behind Claude, retained about 30%.

For comparison, Netflix and Spotify each retained approximately 70% of first-time subscribers in the same transaction data. The research examined 114 billion credit and debit card transactions associated with 174 million U.S. consumers, tracking payments to OpenAI, Anthropic, Google and xAI through April 2026.

AI Customers Are Still Exploring Their Options

By April 2026, approximately 2.9% of consumers in the study's data were paying for at least one of the four major AI providers. OpenAI had the largest share, with 2.4% making payments to the company.

The comparison with streaming needs some perspective. Netflix and Spotify are established businesses with years of customer relationships and familiar entertainment habits. Consumer AI subscriptions are much newer, and people are still discovering which tools are useful to them and worth paying for.

Other subscription research offers some context:

• Consumer subscription apps: RevenueCat's 2026 report, covering more than 115,000 apps using its platform, found median 12-month retention of 6.1% for monthly AI-app subscriptions, compared with 9.5% for non-AI apps. Annual plans showed a similar gap, at 21.1% versus 30.7%.
• Established business software: SaaS Capital's 2026 research reported 91% median annual gross revenue retention among bootstrapped business-to-business software companies with $3 million to $20 million in annual recurring revenue. This measures recurring dollars retained from existing customers, excluding additional sales, rather than the percentage of subscribers who stayed.
  
  

Better AI Models Can Send Customers Shopping

Harvard's researchers also found that customers respond when competing AI companies improve their products.

When OpenAI or Anthropic introduced a better-performing model, the rate at which customers began paying that provider after previously paying only its rival increased by approximately 25% compared with the usual level. Some customers added a second subscription rather than canceling the first.

The study also found that the share of Anthropic's paying customers who subscribed to OpenAI fell from 44% in June 2024 to 28% in April 2026, even as Anthropic's customer base grew.

The researchers suggest that as AI products gain similar capabilities, customers may have less reason to maintain several paid subscriptions. Someone who previously needed different tools for different tasks may now find that one service can do enough.

Can AI Companies Build Lasting Subscription Revenue?

The findings raise a difficult business question: How do AI providers turn growing interest and experimentation into recurring revenue they can count on?

Harvard's researchers found evidence that consumers increasingly view competing AI services as alternatives. Their economic model also suggests customers are highly sensitive to price, which could make increases difficult when similar capabilities are available elsewhere.

There is another side to the picture. RevenueCat found that AI-powered subscription apps generated 41% more first-year revenue per paying customer at the median than non-AI apps, despite weaker retention. The figures were $30.16 for AI apps, compared with $21.37 for other apps.

Higher revenue per customer is encouraging, but it doesn't necessarily mean a profitable business. AI companies still have to cover the costs of developing, improving and delivering their technology. Neither study establishes which providers have found a sustainable balance between revenue and operating costs.

The Harvard paper, released October 16, has not yet undergone peer review. Its main analysis used anonymized Fiserv payment data from approximately 82 million consumers with sufficient observed card activity. It measures payments rather than confirmed cancellations and excludes free AI users and enterprise subscriptions. The authors disclosed advisory relationships with Fiserv, including paid consulting work by one researcher.

Insider Take

The retention findings aren't particularly surprising. AI is still a young market, and customers are learning what these tools can do almost as quickly as the technology changes.

People are experimenting. They're subscribing to one service, trying another and figuring out which products are worth keeping. Some of that movement is understandable, especially when a competitor can introduce a better model and change the buying decision almost overnight.

For AI companies, the challenge is turning that experimentation into a business that can support itself. They need to build dependable subscription revenue while continuing to invest heavily in developing and delivering their products.

It's difficult to know what lasting customer relationships will look like when the technology is evolving so quickly. Today's most compelling feature may be available from several providers tomorrow, giving customers another reason to reconsider what they're paying for.

Will subscribers settle into more lasting relationships as the market matures, or will constant competition and product improvements keep them moving between services?

That's the question AI companies need to answer as they try to build viable subscription businesses.

Sources

• Harvard Business School, Competition in Generative AI: Evidence from Consumer Demand. Working Paper 27-037, October 16, 2026.
• RevenueCat, State of Subscription Apps 2026. AI and non-AI subscription retention and revenue benchmarks.
• SaaS Capital, 2026 Benchmarking Metrics for Bootstrapped SaaS Companies. B2B software revenue retention benchmarks.