Product1 publisher3 min readPublished
Personal AI agents return to a market where 2.2% of consumers pay
PNC figures put paying AI consumers at 2.2% in May, averaging $31 a month, while Meta's Muse, OpenAI's Dots and a $10bn Instinct revive personal assistants. Model upgrades have barely moved either figure, so the teams shipping these agents need revenue that does not depend on a monthly fee.
The Product Desk · Product desk
What happened
- OpenAI's enterprise bookings have reportedly doubled since July, following its widely reported pivot toward business customers.
- Instinct plans to take a cut of purchases made through its agent, and TechCrunch expects it to avoid the cost of training a frontier model.
- Muse has Meta's personalized ad targeting behind it, and Meta is already exploring an enterprise angle for the assistant.
Compiled by The Product DeskSomething wrong?How this is made
Why it matters
- decision A pricing plan that waits for a smarter model to lift paid conversion is betting against the only trend line on record, so the revenue model has to work on the model a team ships now.
- contradiction Menlo's numbers imply roughly one adult in eight pays, four to six times the PNC and Bank of America shares, so a team's conversion target depends on which survey it trusts.
- cost Inference bills land on whoever serves the model, and because AI costs more to run than social apps or cloud software did, even hundreds of millions of subscribers may not cover them.
- constraint Meta's ad business buys Muse time before monetization turns urgent; independent agent makers such as Instinct have to replace that cushion with a plan like a purchase cut.
One of the errands Instinct runs for people is cancelling subscriptions [3]. It is a sensible job to hand an agent. Most consumers have made the same call about paid AI on their own, by not paying for it [5].
Product teams tend to assume users will pay more once the model gets better. Andreessen Horowitz charted the PNC figures in its semiannual State of Markets report [5]. The paying share and the average spend both climb slowly, in a roughly straight line, and the jump from GPT-5.2 to Astra barely shows, according to TechCrunch's reading of the charts [7]. TechCrunch links that flat response to an industry move toward enterprise contracts [4]. Andreessen Horowitz said "it's still so early when it comes to mature AI adoption and utilization" [6].
The surveys disagree on how many people pay, and none of them shows a hard ceiling. Bank of America's count rose 40% in a year, to roughly 3% of U.S. consumers [9]. Menlo's September figures put about 12.5% of adults in the paying column [1], roughly four to six times the PNC and Bank of America shares [2]. Menlo's lead number is daily use, and half of those daily users do not pay [10].
TechCrunch's scale test multiplies Netflix's 325 million subscribers by $34 per customer to get $11 billion a year, less than a third of OpenAI's operating costs [8]. That product holds only if $34 is a yearly figure [3]. At the $31 monthly average in the PNC data, 325 million payers would spend about $121 billion a year [4]. The weak points are how few people pay and how much it costs to serve them. TechCrunch says as much: AI is unusually expensive to operate compared with social networking or cloud computing, and even hundreds of millions of paying customers do not guarantee a break-even [11].
A team shipping its own assistant can sort its options with two tests. The first is whether the task the agent completes most often ends in a purchase the agent handles. The second is whether the user does that task for an employer.
An errand that ends in a personal purchase supports a cut of the transaction, the route Instinct plans to take [15]. Work tasks with no purchase attached can be sold to the employer as a seat; OpenAI pitched Dots at software engineers and agency creatives [13]. When a task is both paid and done for work, a team has both levers. The hard cell is personal use with nothing to buy. There the options are a subscription at the conversion rates above [5], or ads, and Meta's targeting business is what lets Muse wait before choosing [14].
I'd push the product toward the purchase cells wherever the task allows it. The cost is trust. Once the agent earns on what it books, users have reason to ask whether it picked the best table or the best-paying one. A team whose main task sits in the last cell should budget the next model upgrade as an expense, since the charts so far show upgrades have not changed who pays [7].
What to watch
- The next Andreessen Horowitz State of Markets update, to see whether the paying share breaks from its straight-line trend after Muse and Dots.
- Whether Instinct discloses its cut on purchases and what share of its errands end in a paid booking.
- Whether Meta turns its enterprise exploration for Muse into a paid business product.