How Stripe Is Building for an Agent-native World

AI Agents Are Becoming Predominant Actors on the Internet

The internet economy is undergoing a fundamental shift from being human-centric to agent-driven. For a long time, the internet was built on the simple assumption that the main actor was a person browsing, filling out forms, clicking through checkout, or writing code. This assumption is now breaking, as AI interfaces, agents acting on behalf of humans, and software interacting directly with other software become more common.

This emergence of AI agents as new actors on the internet requires a complete rethinking of nearly every layer of the internet stack. It impacts how products are discovered and bought, what developer tools should look like, and critically, how underlying economic infrastructure like payments, billing, fraud detection, and identity layers function when the actors are no longer solely humans.

Stripe, observing this macro trend, sees AI companies experiencing unprecedented revenue growth. A tangible signal of this shift is a 10x year-over-year increase in LLM traffic to Stripe's developer documentation, indicating that machines are increasingly becoming users of developer infrastructure. This highlights the urgency for companies to get 'agent-ready' and adapt to this evolving landscape.

The rise of these autonomous actors fundamentally changes the nature of the internet, moving beyond just improving existing functions like search or coding. It's about a new class of participants that will eventually become the predominant actors, necessitating a re-evaluation of established economic and technical frameworks.

The internet has this new kind of actor on it. Over time, this actor, these agents, will become the predominant actors on the internet.

Compute Theft Becomes an Existential Threat for AI Businesses

Traditional fraud largely focused on payment fraud and stolen credentials, but for AI companies, the problem has evolved into compute theft. Unlike older SaaS models where free tiers had minimal cost, giving out credits, freemium offerings, or free trials in AI incurs a very real cost with every prompt, image generated, or API request. This makes compute theft an existential risk, as intelligence is still far from free.

Free compute serves as the new customer acquisition cost (CAC) for many AI companies, replacing traditional paid media spending. This reliance on free trials and credits for growth creates a significant vector for abuse. Fraudsters exploit these systems by engaging in multi-account abuse, where bad actors sign up repeatedly with new identities and email addresses to claim new user credits, staying ahead of detection.

Across AI companies on Stripe, about 7% of new sign-ups are from multi-account abusers. Another major issue is free trial abuse, which rapidly breaks unit economics. One large AI company saw only 4% of free trials convert to paid, with each trial costing $25 in LLM spend. This meant they were spending $625 per payer before earning any revenue, primarily due to abusers with no intent to pay.

Companies have responded by either dropping free trials, which throttles growth, or by blocking virtual cards. However, blocking virtual cards also harms legitimate users, as about 15% of legitimate card transactions for AI companies on Stripe utilize them. Free trial abuse has quadrupled in the last six months, highlighting the urgent need for effective solutions that don't impede legitimate growth or users.

every prompt, every image that gets generated, every API request has a very real cost attached to it.

AI and Valuable Services Transform Fraud into a Full-Funnel Challenge

Fraud is escalating rapidly, driven by two key factors: bad actors are now leveraging AI, and the services they steal, such as compute or inference, hold significantly higher value than traditional SaaS offerings. This allows fraudsters to resell stolen resources or use them for other illicit activities, making the incentive for fraud much greater.

One major AI user on Stripe, for example, experiences 250,000 fraudulent free trials blocked weekly, highlighting the scale of the problem. Beyond direct payment fraud, companies face "non-payment abuse," where users consume thousands or tens of thousands of dollars in compute time or other services before their invoice fails, leaving the service provider with substantial unrecoverable losses.

This evolution means fraud can no longer be addressed solely at the point of transaction. It has become a "full-funnel" customer problem, requiring vigilance from the moment of sign-up, including detection of multi-account abuse or inappropriate free trial access. Companies must also manage credit allocation, enforce throttling for overages, or require top-ups to prevent abuse.

The nature of what can be stolen and its increased value has made the impact of fraud more existential for businesses, necessitating a comprehensive, lifecycle-based approach to security and abuse prevention.

fraud used to be a transaction thing, now it is a customer thing, it is a full funnel thing.

Stripe Radar Adapts for Full-Funnel AI Fraud Protection

Stripe Radar, traditionally used for fraud detection at the point of checkout, has evolved to address new fraud risks in the AI industry. Due to significant 'up-funnel' fraud originating earlier in the customer journey, AI companies are now integrating Stripe Radar at signup to monitor metadata and assign scores from the very beginning.

This shift acknowledges that fraud is now a full-funnel problem, not confined to transactions alone. For AI companies, Stripe Radar focuses on protecting critical customer lifecycle moments, including signup, initial payments, and overages. These protections are implemented through calls to the Radar API at each stage.

Businesses unsure about their fraud rate, especially regarding free trial abuse, can use the Radar Assistant in the Stripe dashboard. Users can describe their business model, like having high marginal costs, to tailor insights. Integrating Radar's up-funnel capabilities provides a clearer understanding of potential issues through back-testing.

For example, one user discovered their business had a low total fraud rate of 0.2% with only 0.02% early fraud warnings after checking their Stripe Radar dashboard. This immediate feedback helps businesses assess their risk posture and decide if further integration is necessary.

fraud is now a full funnel problem, not a transaction problem alone.

Stripe combats AI-powered fraud with comprehensive data and AI defenses

The internet is experiencing an escalating "AI-on-AI" arms race in fraud detection. Fraudsters are leveraging artificial intelligence to bypass security measures, indifferent to payment boundaries such as processing platform, currency type, or payment method (e.g., cards, crypto, buy now, pay later). Their goal is simply to find workarounds.

Stripe addresses this challenge by providing comprehensive fraud protection. Initially, their Radar system only covered card transactions, but it has expanded to include all payment methods with disputes, like ACH, SEPA, and crypto. Stripe also offers the Radar API, allowing businesses to screen transactions for fraud even if they are processed by other providers.

Furthermore, Stripe treats fraud mitigation as a public good, investing significantly in its defense. By processing 2% of global GDP and growing 34% year-over-year, Stripe gains access to an immense dataset. This data, combined with their AI, enables them to identify and quickly neutralize new fraud vectors emerging across various processors, payment methods, and merchants.

While acknowledging the constant creativity of fraudsters, Stripe remains bullish yet vigilant, using its extensive network and AI capabilities to stay ahead in this dynamic battle.

fraudsters don't really care about boundaries. They don't care about whether this transaction is processed on Stripe or off Stripe. They don't care about whether this transaction is on fiat or crypto, whether it's on a card network or a buy now, pay later. They're just gonna figure out how to work around the system, to get through.

AI companies achieve revenue milestones three times faster than top SaaS companies, driving rapid monetization model evolution

AI companies are demonstrating unprecedented revenue growth, outpacing previous cohorts significantly. Data from Stripe indicates that the top 100 AI companies reach $30 million in annual recurring revenue (ARR) in approximately 18 months. This growth rate is three times faster than that observed for the top 100 SaaS companies from 2018, with similar accelerated scaling evident even at lower ARR milestones like $1 million and $5 million.

This hypergrowth is accompanied by a swift evolution in monetization strategies. Traditional SaaS models often relied on fixed, seat-based subscriptions, which made sense for human-centric usage and minimal marginal costs. However, AI solutions face substantial and very real inference costs, necessitating different approaches.

As a result, usage-based billing has rapidly gained importance. Companies now meter a variety of units, including tokens, API calls, workflows, and even outcomes, choosing units that best align with both customer value and their underlying cost structure. This requires high-precision charging, tracking every event and its associated metadata.

The industry is also seeing a rise in hybrid monetization models. While subscriptions remain relevant, they are now frequently augmented with usage overages, prepaid credits that burn down, or real-time top-ups, reflecting a dynamic adjustment to the unique economics of AI.

these AI companies are just growing from a revenue perspective faster than any previous cohort we've seen.

Outcome-Based Billing Emerges as the New Standard for Vertical AI Solutions

As AI-powered products evolve, pricing models are shifting from traditional subscriptions to more dynamic, usage-based approaches. Companies like Lovable initially used simple subscriptions but moved to token consumption-based billing for their AI offerings. This hybrid model helps align revenue with actual usage, value, and the operational costs of running AI models.

Stripe's 'token billing' solution addresses the volatility of underlying LLM costs. It allows businesses to track and price their services in real-time based on the cost of tokens, with a business-defined markup. This prevents margin erosion when the expenses for core AI models fluctuate rapidly, ensuring profitability for companies that build on top of other LLMs.

For core AI model providers, such as OpenAI or Claude, the emerging standard for monetization is metering tokens consumed via their APIs. However, for vertical AI solutions designed to solve specific business problems, the future of pricing is moving towards outcome-based billing. Customers pay for achieved results, rather than raw usage or seat licenses.

Examples like Fin and Intercom already implement outcome-based metering for services such as resolved support tickets. This model ensures that end users hold vertical AI solutions accountable for delivering tangible outcomes and a clear positive return on investment, as the onus for demonstrating ROI falls on the specialized AI provider.

Because I think end users are gonna wanna hold those vertical solutions accountable for outcomes, and they're gonna wanna know that they have positive ROI on their spend.

AI is ending seat-based enterprise software pricing and fueling net new spending

The traditional enterprise software model of seat-based pricing, often with overages, is rapidly becoming obsolete. Historically, software was free to run, but with the advent of AI, new operational costs emerge, requiring providers to add overages to protect margins. However, tying revenue to the number of users becomes illogical when AI agents begin to automate and enhance individual productivity.

This shift is most apparent in sectors like customer service or software development. If an AI tool significantly reduces the need for customer service representatives, charging per rep makes little sense. Similarly, if AI agents boost a developer's productivity by tenfold, pegging software revenue to the headcount of developers becomes counterproductive, as the total number of developers needed might decrease.

A substantial decline in seat-based licenses for enterprise software is expected in the near future, potentially by half within six months. Furthermore, the rapid growth seen in new AI companies, with some reaching $30 million in annual recurring revenue in only 18 months, indicates that much of this expansion is driven by net new spending entering the economy, rather than simply a reallocation of existing budgets from other providers.

I would be super surprised if six months from now, we have half of the seat-based licenses that we have today.

AI spending will increasingly substitute traditional SaaS and headcount, with overall high retention for categories but lower for individual providers, while unique vertical solutions support valuations.

Initial AI spending was largely experimental and additive, but it is now shifting to substitute traditional SaaS licenses and headcount operational expenses. Organizations are beginning to factor AI costs into their existing budgets, for example, by adjusting the perceived cost of an engineer to include LLM expenses. This marks a transition from purely new expenditures to a re-allocation of funds.

An interesting dynamic exists within AI retention: while the overall retention rate for AI tool categories (like AI dev tools or coding assistants) is higher than traditional SaaS, the retention for individual providers is slightly lower. This means users are committed to using AI tools but frequently iterate and "hop" between different providers within a category to try new models or find better quality solutions.

Despite some crowded areas, AI company valuations are largely supported by "blue ocean" vertical solutions that address previously unsolved problems without direct competitors. These pioneering companies can initially charge sticker prices. However, in more competitive segments, companies are starting to see complex sales processes, nuanced billing, and reactions to competitor monetization models, indicating a maturing market.

Once you start using an AI dev tool, like a coding assistant, you love it, you're not going to stop using it, but you very well may iterate across providers as models vary in their quality or anytime a new model comes out, you're just like, "I gotta try this."

Stripe Adapts Developer Experience for AI Agents and Automated Stack Provisioning

The traditional view of developer experience, focused on human engineers at a keyboard, is rapidly evolving to encompass AI agents, coding assistants, and non-technical founders. This shift means products must now cater to a broader range of "developers," including scenarios where humans work through agents or agents act autonomously. The goal is to provide a coherent and trustworthy product experience across all these actors.

A clear signal of this change is the tenfold increase in LLM traffic to Stripe Docs year-over-year. This demonstrates that machines are actively consuming developer documentation and infrastructure, acting as direct users. While human usage of Stripe Docs remains steady or grows, AI agents represent a significant and expanding portion of the overall developer activity, rather than a straight substitution.

Stripe is also addressing the persistent challenge of manual software stack provisioning, which often involves creating accounts, managing credentials, and navigating multiple dashboards, even as coding gets easier. With the recent launch of Stripe Projects, users or their agents can now create and manage parts of their software stack directly from the command line, automating resource provisioning in owned accounts and syncing credentials back to their environment.

I think it's, it's less about just like, okay, how do we help a developer, human developer, write code, and more about how do we have a coherent and trustworthy product experience sort of end to end that acknowledges that at some moments the actor is a human, at some moments the actor is an agent, and at some moments the actor is a human working through an agent.

Exploring the Agentic Commerce Spectrum and its Payment Implications

Agentic commerce describes a range of AI-driven purchasing scenarios, often misunderstood as only the most extreme form where a system autonomously buys things. Instead, it operates on a spectrum, beginning with AI simply reducing friction in current online experiences, such as aiding research or filling forms while the human still makes the final decision.

The spectrum advances to more sophisticated stages like descriptive search, where users specify complex needs (e.g., a summer camp for kids with specific budget and dates) instead of blunt keywords. Further along is real delegation, where a system makes purchases based on user-defined constraints. The most futuristic stage is ambient commerce, where the system autonomously anticipates and fulfills needs without explicit prompts.

Regardless of where on this spectrum a system operates, the fundamental economic infrastructure, particularly payments, must evolve. The traditional checkout process, which relies on humans creating accounts, entering details, and clicking purchase buttons, becomes increasingly obsolete as AI agents take over or assist with these steps.

This shift necessitates a redesign of payment flows, especially for agent-assisted buying, where discovery and checkout occur within AI interfaces like Gemini or ChatGPT rather than directly on a merchant's website. A key challenge is ensuring AI agents can fully comprehend a merchant's product catalogs, pricing, and checkout sequences to act effectively on behalf of consumers.

Even the earlier stages force a redesign of payments infrastructure in particular, because, the today model, the old model, again, humans sitting in front of a browser, creating account, choosing plan, filling out the forms, clicking purchase, entering card details not all those steps are happening anymore.

Stripe Introduces Agentic Commerce Protocol and Shared Payment Tokens for Secure AI-Assisted Transactions

The rise of AI agents in commerce brings trust challenges: consumers are hesitant to share payment credentials with agents, and merchants need assurances that bots are legitimate. To solve this, Stripe collaborated with OpenAI to create the Agentic Commerce Protocol, a shared technical language between AI systems and businesses.

Merchants integrate their product catalog, prices, and checkout flows with Stripe once. From their dashboard, they can then expose their offerings to a range of AI agents, including those used by Microsoft Copilot and Meta's in-app shopping. Crucially, merchants remain the merchant of record, preserving customer relationships and control over fraud, even as transactions are facilitated by AI agents.

A core component of this system is the Shared Payment Token (SPT). This primitive securely passes tokenized payment credentials from the AI agent to the merchant, ensuring the agent never views sensitive payment information. Additionally, the SPT transmits fraud scores to the merchant, providing essential data to assess the legitimacy of the transaction and maintain trust.

Stripe developed these solutions to simplify participation in the evolving AI commerce landscape. Merchants no longer need to build custom integrations for every new AI storefront. Stripe aims to abstract this complexity, enabling businesses to easily sell to customers who are increasingly making purchases through AI tools and agentic flows, mirroring their past efforts in enabling sales across websites, apps, and marketplaces.

But now, like, you know, where are the consumers? Where are they wanting to buy? Increasingly through sort of AI tools and agentic flows.

Agentic Commerce Volume and Early Trends

Agentic commerce volume is currently a relatively small percentage of all commerce but is growing quickly. This trend is particularly noticeable for items considered commodities, as consumers begin to trust agents with simple purchasing decisions.

Early agent-driven purchases focus on products that are low-stakes, reasonably known, and not high-priced. This pattern is similar to the initial cautious adoption of online shopping, where consumers first bought less expensive, easily understood items rather than high-ticket goods like a $2,000 couch or a mattress.

While complex, higher-value purchases, such as planning an entire summer trip for a family of four, are not yet satisfactorily handled by agents, it is expected that the underlying models, interfaces, and user experiences will continue to improve. The foundational elements of agentic commerce, including catalog exposure, shared payment tokens, and fraud protections, are designed to be adaptable to these future advancements.

What is the first thing that people are comfortable buying through agents? It's like things that are reasonably known, reasonably observable, not super high priced.

Stripe's Link Wallet Evolves for Delegated Agent Purchases

Consumers face a challenge when wanting AI agents to make purchases on their behalf: how to do so safely without giving away full payment credentials. Businesses need a way to safely expose their products and checkout processes while retaining control and managing fraud, even with agents involved.

Stripe's Link is a consumer wallet widely used by approximately a quarter of a billion consumers for quick checkout. It's becoming particularly relevant for the AI economy due to its dense network of AI-savvy consumers.

Link is evolving to support secure delegated agent purchases. Instead of handing over card details, consumers can use Link to define which agents are allowed to request credentials, under what specific conditions, and with what spending limits. It also allows for purchases to require explicit approval before they go through.

This new model shifts from blind trust to delegated authority with guardrails, providing a much more sensible and secure framework for consumers to leverage agents for buying things like Halloween costumes.

the model isn't, you know, give a random agent your card and hope for the best. Instead, it's delegated authority with guardrails.