John Collison: How AI Agents Will Rewire Commerce and the Internet

Stripe co-founder John Collison says AI agents will do a large share of online buying, that computer use is the technology that makes it possible, and that the web’s defenses against bots will have to be rebuilt as a result. In a 28-minute conversation with Molly O’Shea of Sourcery, recorded on stage at a Browserbase event, Collison covers agentic commerce, why Stripe bought OpenRouter and Metronome, a blunt forecast on security breaches, and what 17 years of building Stripe taught.

TLDW

Stripe processed $1.9 trillion in 2025, up 34%, and Collison says a lot of the growth now comes from AI. Personal agents that know your address and card and can operate a browser will take over form-filling and much of shopping, and AI product research will become the new place where buyers discover sellers. Collison expects more breaches over the next five years than the last five, calls January 1, 2026 “day one of the singularity” with more optimism than dread, and says Stripe’s own craft is changing because coding agents, not people, now integrate its API. The open question nobody has answered is how sites tell good bots from bad ones.

Thoughts

Asked what question is not being asked, Collison gives the answer that quietly undercuts the rest of the talk: sites built bot detection for good reasons, and now “it’s increasingly all bots.” That matters because the thing Collison is most excited about, computer use, works precisely by behaving like a person on sites that never agreed to serve software. The line about not waiting for the DMV to add an MCP server gets a laugh, but the DMV did not consent either. A web where agents click through pages on our behalf needs some way for a site to know whose agent it is, what it is allowed to do, and who pays when it goes wrong. Stripe is unusually well placed here, since it already sits between buyers and sellers deciding which transactions to trust. Collison says plainly that nobody knows what the new equilibrium looks like. I would watch that space more closely than any model release.

The most candid admission is about Stripe itself. For its whole history the company competed on developer experience: careful documentation, painless upgrades, an API people enjoyed using. Collison now says that no one really integrates the Stripe API by hand anymore. They send a coding agent. What is pleasant for a person and what works for an agent are different things, and an agent will happily integrate the wrong API version. So a company whose edge was taste in developer tools has to learn a new skill. Collison’s image is a chessboard that got jiggled, with the pieces in slightly different places for everyone. If your product’s advantage is that humans find it easy, it is worth asking who or what is using it now.

The acquisitions say more about where Stripe thinks the money is going than any forecast would. Metronome is usage-based billing, because every AI product has inference costs and so almost every product now has a metered component. OpenRouter routes requests between models, and Collison’s case for it is specific: a business will use several models at once, switch as better ones arrive, and send a high-value customer to a frontier model and a low-value one to something cheaper. Put those together and Stripe is positioning itself to meter what AI companies charge and to route what they spend. Collison says inference is becoming a huge share of business spending. A payments company wanting to sit on that flow is not a surprise, but it is a clear sign of how large the flow has become.

On commerce, Collison argues that AI discovery is the new point of aggregation and that this is good news for small brands with products people like. The comparison is to Facebook and Instagram ad targeting, which produced a generation of companies native to it while big-box retailers with the same tools lagged. I find the first half convincing and the second half hopeful. Anything that ranks products becomes a target, and reviews and “what people like on the internet” are exactly the signals that get manufactured. The earlier platforms rewarded whoever learned the system fastest, which was not always whoever made the best thing. It is fair to note that Stripe earns from transactions whoever wins.

The closing lessons are the part worth keeping. Stripe, Collison says, never had a breakout moment. Each year was bigger than the last, growth sometimes slowed, and after enough of those years the total reached $1.9 trillion. Most new products did not look like the future of the company when they started. Collison’s word is “stick-with-itness,” and the example is Nvidia, which made graphics chips through long stretches when nobody cared. In a talk mostly about how fast everything is changing, the advice at the end is to stay put: do not quit, do not sell, be a good counterparty, and begin with what the customer wants. That holds up at any speed.

Key Takeaways

Filling out a web form is not a valuable activity for anyone. Agents that hold your shipping and payment details and can operate a computer will do it for you.
Computer use is “the ultimate backwards compatibility layer for the real world,” because it works on sites that will not build agent interfaces for years.
Collison expects good computer use to be seen as the next discrete step in AI, after transformers, large language models and reinforcement learning.
Agentic commerce changes two things: agents do much of the buying, and AI product research surfaces niche brands, which puts pressure on existing aggregators.
On security, new models have advanced cyber capabilities available to attackers and defenders alike. Collison expects more breaches in the next five years than in the last five, because many organizations will adopt defenses too slowly.
Startups are better placed on security than older firms: smaller, more uniform systems, and faster adoption of new models.
Stripe half-jokingly dates the singularity from January 1, 2026, after coding models crossed a threshold late last year. Collison points to AI now speeding up AI research.
Stripe is seeing the highest rate of new company creation in recent memory, which Collison attributes to cheap intelligence making it easier to start and run a business.
Stripe will not say what share of its volume is AI, partly because the line has blurred. Companies founded before AI are becoming AI companies.
Build or buy: Stripe builds what is close to its core and buys when the skill is new to it and the market is moving too fast for an internal effort to catch up.
Inside Stripe, an internal AI tool with access to company data, behind careful permission controls, has let sales and legal staff build their own applications.
Collison keeps up by building small apps for personal use and by spending time with the fastest-growing companies on Stripe, and recommends trying a personal agent to feel what is coming.

Chapters

00:02 Personal Agents and the End of the Web Form

O’Shea opens with Stripe’s $1.9 trillion in 2025 volume. Collison, in the company’s 17th year, says payments barely changed for a long time and are changing now. Personal agents work when they have context about you, enough compute, and the ability to operate a computer, so that “renew my driver’s license” replaces a session of form-filling.

02:03 Agentic Commerce as a Complete Rewiring

The first-order change is convenience. The second is a reshuffle of who sells. Collison compares it to the rise of Wish, Teespring, Temu and Shein on social ad platforms, and expects merchants who adapt their products and discovery to agents to climb the rankings.

05:07 Cybersecurity: More Breaches Are Coming

Bank robbers adopted the car before the police did, Collison notes, and offense and defense always chase each other. Stripe has long been paranoid about security and is spending more attention on it. In theory the two sides balance. In practice, organizations on decades-old systems assembled through acquisition will be slow.

07:09 Day One of the Singularity

Like the Unix epoch of January 1, 1970, Stripe picked January 1, 2026 as its singularity epoch. Collison cites coding models, a run of AI mathematics announcements, and labs using AI to build AI. The tone is upbeat: we are in a takeoff, it is hard to deny, and you should enjoy it.

09:10 Should You Be Scared? APIs Built for Agents

Collison’s advice to businesses is to pay attention and to worry about specific things, such as security, not to carry a general sense of dread. The example is Stripe’s own shift from developer ergonomics to what Collison calls AI ergonomics, now that coding agents do the integrating.

11:12 Why Stripe Bought Bridge, Privy, Metronome and OpenRouter

Bridge and Privy brought people who are native to crypto. Metronome brought usage-based billing, which AI cost structures make necessary. OpenRouter is a bet that businesses will run several models and route between them. Collison says AI accounts for a lot of Stripe’s volume but will not give a number.

15:15 Build Versus Buy

Collison recalls asking Aaron Levie of Box how to prioritize products and being told to list everything and do the most impactful things. Build versus buy is similar: the judgment is everything. OpenRouter’s market looked different a year ago and will look different again in a year, which favored buying.

17:17 Inside Stripe: Personalized Software and Computer Use

AI is uneven, strong where training data is plentiful and weak elsewhere, and the company version of that gap is business context. Stripe’s internal tool gives AI access to company data with access controls. Sales staff build their own tools for tailored pitches, and someone in legal built an agent to read new regulations, such as 600-page laws from India and Finland.

20:19 Staying Informed Without Losing Your Mind

Collison does not claim to read ten papers a night. The method is to build things and talk to customers, especially fast-growing ones. For staying sane: put the phone down and get outside, across the Golden Gate Bridge to Marin if you are in San Francisco.

23:22 The Bot Question and 17 Years of Lessons

Bot detection made sense when bots were the exception. Collison asks what replaces it when sites must sort good bots from bad. The closing lessons: no shortcuts, persistence, play a long game with partners, protect your reputation, and start from the user and work backwards.

Notable Quotes

“Computer use is the ultimate backwards compatibility layer for the real world.”
John Collison, on why agents do not need every site to build an interface for them

“AI discovery is the new point of aggregation for commerce.”
John Collison, on what agentic shopping does to existing marketplaces

“Over the next five years you’re going to see more breaches than you did over the past five years.”
John Collison, on AI and cybersecurity

“No one actually really integrates the Stripe API by hand anymore.”
John Collison, on coding agents replacing the developer as Stripe’s user

“If you’re not using a lot of AI in your business, you’re probably not paying attention.”
John Collison, on why Stripe cannot separate AI companies from the rest

“It’s increasingly all bots.”
John Collison, on why bot detection needs rethinking

“Stripe never had a South by Southwest breakout moment.”
John Collison, on growth that compounded without a single turning point

It is a quick and unusually direct conversation. Watch the full interview here.