How Stripe plans to make AI agents the web's new customers

AI agents are becoming a new type of customer online. Last week, Stripe launched Link wallets for agents, stablecoin-based streaming payments and expanded fraud protection — the package meant to make agentic commerce work in practice.

Image: Stripe CEO and co-founder Patrick Collison opened the Sessions developer conference.

The launches Stripe presented at its annual developer conference Sessions last week are not isolated product announcements, but pieces of a larger roadmap for how AI agents will transact online. Traffic from large language models to Stripe's developer documentation has grown tenfold in a year, while human use is flat. Machines now read the documentation — and increasingly act on it. For a payments provider that processes around two percent of global GDP, this is not a marginal trend but a rebuilding of the foundation of the internet's economy. What Stripe is rolling out — wallets for agents, stablecoin settlement and fraud protection across the entire customer journey — provides a useful map of where agentic commerce is actually headed, and which problems must be solved before it works in practice.

A new type of customer online

Stripe's Head of Data and AI, Emily Glassberg Sands, describes the development as the internet having gained «a new kind of actor», and says agents will over time become the dominant one. Sands divides agentic commerce into four levels: AI-assisted purchases where the human still decides, natural language search, delegated purchases within user-defined parameters, and ambient commerce where the system acts proactively. Most real-world use cases today sit at the two lower levels. It is only when delegation becomes real — when the agent actually transacts on your behalf — that the infrastructure Stripe is now launching becomes essential.

Wallets for agents — programmable authorization

The most visible announcement is Link wallet for agents, launched last week. In a post on X, Stripe writes that the wallet lets users «securely empower agents to spend on your behalf», without exposing payment credentials and with approval required for every purchase.

Behind it sits the Agent Commerce Protocol, developed with OpenAI and now in use at Microsoft Copilot and Meta. The protocol provides a common technical language between AI systems and merchants, where the merchant remains the formal seller and retains the customer relationship. The critical component is the Shared Payment Token, which lets payment details flow from agent to merchant without the agent ever seeing the card number, and with fraud assessment tied to the token itself.

Stripe's existing Link wallet, used by around a quarter of a billion consumers, is now being rebuilt for agent use: the user can define which agents are permitted to request payment, under what conditions, with what spending limits, and whether purchases require manual approval. The model is programmable delegation — a step up from «hand a random agent your card and hope for the best».

Stablecoins as settlement for machine-to-machine payments

When agents transact in real time and at scale, traditional payment infrastructure cannot keep up. Stripe therefore launched streaming payments aimed at AI companies last week, where consumption is metered continuously and settled in stablecoin on the company's own blockchain, Tempo. In practice, the customer pays per token at the moment the model consumes them.

The logic is straightforward: 30-cent card fees make no sense when the transaction is one tenth of a cent. Stablecoin-based payment networks make micropayments economically viable, and an own-chain gives Stripe control over speed and cost. This is not crypto as speculation, but crypto as infrastructure — and it is likely here that it will become clearest that agentic commerce requires a settlement layer that does not exist in today's payment networks.

Fraud becomes a full-stack problem

Agentic commerce also moves the fraud problem, and Stripe therefore extended its fraud product Radar last week to protect AI companies from token theft and trial abuse. Sands says fraud «used to be something that happened in a transaction, now it's something that happens across the customer journey». The logic is that free compute has become the new customer acquisition cost — and therefore the new attack target.

At one major AI company on Stripe, only four percent of trial users converted to paying customers, while each trial cost around 25 dollars in compute. That is 625 dollars in pure cost per paying customer, before anything is earned. Trial abuse has quadrupled in six months, and one company blocks a quarter of a million fraud attempts per week. Stripe states that one in six signups for AI services on its platform comes from malicious actors. Radar has been rebuilt to screen across the entire funnel — at signup, payment, overconsumption and refund — a quiet but radical change in what fraud protection must be when the customer is not a human clicking, but an agent scaling.

Pricing models follow

Business logic shifts with the infrastructure. Sands expects the classic SaaS model — fixed price per user per month — to largely disappear from the enterprise market. Three models are emerging: token-based pricing, where the price tracks underlying model cost with a markup; outcome-based pricing, where customers pay per resolved case or completed task; and fixed monthly pricing with overage charges for the consumer market.

The change is not cosmetic. If one AI agent can replace five customer service representatives, pricing the product per «seat» makes no sense. Outcome-based pricing shifts risk from customer to vendor and changes who profits from the efficiency gains. For investors and operators, it means that the margin structure of the software industry is being renegotiated at the same time as distribution moves into AI interfaces.

Who owns the infrastructure

Stripe is not alone in positioning itself. Microsoft is building universal protocols to become the foundational infrastructure of the agentic economy. Ant International has launched its Agentic Mobile Protocol to connect AI agents to mobile wallets in markets where digital wallets have surpassed card networks. Google is investing in the Gemini Enterprise Agent Platform with governance tools designed to handle thousands of agents simultaneously — while at the same time partnering with Stripe to let businesses sell directly in the Gemini app.

The global market for agentic commerce is estimated at 28 billion dollars by 2030, with annual growth of around 46 percent. The number itself is less interesting than what it signals: a new layer of the economy is being built, and the question is not whether it is coming, but which players will set the standards. Stripe's launches last week are one clear signal of where one of the biggest players believes the direction is heading — and which pieces of the infrastructure must be in place before agentic commerce can work in practice.

Sources: Stripe, Crypto Briefing, BigGo Finance