What the Agent Economy Looks Like From Inside Stripe
TL;DR
Stripe sees agents becoming the internet’s main users — Emily Glassberg Sands says the old assumption of “a person sitting in front of a screen” is breaking, and now payments, billing, fraud, identity, docs, and developer tooling all have to be rebuilt for software acting on behalf of humans or other software.
AI fraud is no longer just stolen cards — it’s stolen compute — Across AI companies on Stripe, about 7% of signups are multi-account abusers, free-trial abuse has grown 4x in six months, and one large customer is blocking 250,000 fraudulent free trials per week.
Free compute has become the new CAC, and it can wreck unit economics fast — Emily gives an example of a large AI company whose free trials cost $25 each, converted at only 4%, and were effectively burning $625 per paying customer before Stripe identified that most “non-converters” were actually abusers.
AI companies are scaling much faster than prior SaaS cohorts — Stripe’s top 100 AI companies that hit $30 million ARR are getting there in roughly 18 months, about 3x faster than the top 100 SaaS companies from 2018.
The pricing model is shifting from seats to usage and eventually outcomes — For model providers, token-based pricing makes sense; for vertical AI apps, Emily thinks outcome-based pricing will win over time, citing examples like Intercom’s Fin charging per support case resolved and Lovable layering usage on top of subscriptions.
Stripe is building for both agent builders and agent buyers — LLM traffic to Stripe docs is up 10x year over year, Stripe launched Projects to let developers or agents provision tools from the command line, and its agentic commerce stack with OpenAI uses shared payment tokens so agents can buy without exposing raw card details.
The Breakdown
The internet’s new actor: agents, not just people
Emily opens with the big frame: the internet used to assume a human was always the one browsing, coding, and checking out. That assumption is now breaking in multiple ways — sometimes humans act through AI interfaces, sometimes agents act on their behalf, and sometimes software just talks directly to software. Her point is simple but huge: every layer of infrastructure now needs to become “agent ready.”
Fraud has mutated from payments fraud into compute theft
Dan asks whether AI changes what even counts as fraud, and Emily says yes: increasingly, fraudsters aren’t stealing money first — they’re stealing expensive model usage. She describes three major abuse patterns: multi-account signups, free-trial abuse, and postpaid non-payment, with striking numbers like 7% of AI signups being repeat abusers and one company blocking 250,000 fraudulent free trials every week. The killer line is that “free compute is the new CAC,” which makes abuse an existential growth problem, not just a payments ops issue.
Stripe moved Radar upfunnel because AI fraud starts before checkout
Historically, Stripe Radar mostly lived at the transaction layer, but that no longer works if the expensive thing gets stolen at signup. Emily says AI companies are increasingly integrating Radar when users first create accounts, then again at payment and at overage moments, because fraud is now a “customer thing” and a “full-funnel thing.” Dan literally checks his own dashboard mid-conversation, which gives the whole segment a nice live-wire feel.
The fraud arms race is real, but Stripe thinks breadth is its advantage
Emily says fraudsters don’t care about product boundaries: they’ll move across cards, crypto, BNPL, processors, and whatever else works. Stripe’s answer is to make defenses more comprehensive, including Radar support beyond cards, across dispute-enabled methods, into crypto, and even on transactions processed off Stripe via the Radar API. She frames fraud defense as a public good, and argues Stripe’s scale — it sees 2% of global GDP and is growing 34% year over year — gives it the data edge to keep spotting new attack vectors quickly.
The Stripe view of the AI economy: absurd growth and messy monetization
From Stripe’s vantage point, the most obvious pattern is speed: top AI companies are reaching $30 million ARR in about 18 months, far faster than prior SaaS cohorts. The second pattern is pricing chaos in the best sense — subscriptions, token metering, prepaid credits, overages, top-ups, hybrid plans, and even real-time token billing so wrappers don’t get crushed if underlying model costs move. Emily uses Lovable as a concrete example of a company that started with simple subscriptions and evolved into hybrid usage-based billing as the product surface expanded.
Why seats may be dying, especially for AI software
Dan pushes on what the new default pricing model will be, and Emily gives a sharp answer: tokens for models, outcomes for vertical apps. She says seat-based pricing starts to look silly once software is automating the very humans you used to charge per headcount, especially in categories like developer tools or support tools. Her prediction is bold and memorable: six months from now, she’d be surprised if we still had half as many seat-based licenses as we do today.
Today’s AI growth is mostly net-new spend — but not forever
Emily says most of the AI revenue explosion so far looks additive, not merely shifted from legacy SaaS or headcount budgets, because enterprises are still experimenting and still locked into old software contracts. But she expects that to change: eventually companies will reason about LLM spend as part of total labor and software cost, not a side experiment. She also notes a subtle retention pattern: people stick with AI categories like coding tools, but often switch providers inside the category as new models and products keep appearing.
Stripe’s own product shift: from developer experience to agent experience and agentic commerce
Inside Stripe, “developer experience” now means serving humans, nontechnical founders using tools like Replit or Vercel, coding assistants, and autonomous agents. Emily says LLM traffic to Stripe docs is up 10x year over year, and points to Stripe Projects as a response to the fact that coding is getting easier faster than infrastructure setup is. On the commerce side, she sketches a spectrum from AI-assisted checkout to full delegated buying, then explains Stripe’s protocol with OpenAI: merchants integrate once, expose catalog and checkout flows to multiple AI shopping surfaces, and use shared payment tokens so agents can transact without ever seeing a user’s raw card credentials.