Stripe's AI Chief: How AI Agents Will Buy, Sell, and Pay

Summary

Emily Sands lays out Stripe’s view that AI agents will not just browse and buy, but eventually negotiate, transact, provision infrastructure, and run businesses end to end. To make that work, Stripe and partners are building agent-commerce primitives such as shared payment tokens, Link-based agent wallets, and protocols that let merchants expose catalogs, inventory, pricing, and authorization rules once across many AI surfaces. Sands says the economics of AI are forcing a move away from pure SaaS subscriptions toward usage-based, hybrid, and real-time billing because inference has real marginal cost and agents can spend at machine speed. She also highlights token theft and multi-account abuse as underappreciated fraud modes, and says the next bottleneck after “vibe coding” is “vibe deployment,” where apps can be created quickly but still need orchestration to ship and operate safely.

Notes

• Stripe frames agentic commerce as a spectrum, from fully autonomous agent purchases to human-led discovery with an AI buy button.
• Consumer behavior is mostly around level 2 autonomy today, with a hint of level 3.
• Merchants need deterministic exposure of product catalogs, inventory, and prices to agents.
• The protocol is designed so businesses publish product data once and do not have to register with every new agent surface.
• The Agentic Commerce Protocol includes shared payment tokens that hide underlying credentials from the agent.
• Stripe says the protocol is platform-agnostic and payment-processor-agnostic.
• The protocol is being built to work across OpenAI and other providers, not as a closed Stripe-only surface.
• Stripe’s Link wallet is being repurposed for agents with built-in controls and revocation.
• Link Wallet supports budgets across providers, per-transaction permissions, and aggregate limits.
• Shared payment tokens are a new primitive built for agentic commerce about six months ago.
• The token encodes merchant, amount, currency, and duration constraints.
• Microtransactions become viable when agents can transact without human credential entry or website navigation.
• Businesses must remain the merchant of record in agentic commerce.
• Radar scores for buyer and agent behavior are included in the shared payment token.
• Stripe says it sees about 2% of global GDP and a very large share of AI buyers and sellers.
• Agent traffic to Stripe’s documentation grew more than 10x in a year and is now about 40% of docs traffic.
• About 70% of Stripe CLI API resource requests are now from agents.
• Vibe deployment is emerging as the binding constraint after vibe coding.
• Stripe Projects is intended to help agents sign up for and integrate deployment services.
• Stripe’s core rationale is that deployment friction is now the bottleneck to building and shipping.
• Stripe positions orchestration as a central part of its mission to increase the GDP of the internet.
• AI breaks the SaaS billing model because prompts, API calls, and tasks now have real marginal cost.
• Usage-based billing is critical for AI companies because inference is not free.
• Stripe says few scaled AI companies remain purely subscription-based or seat-based.
• Lovable started on subscriptions and moved to a hybrid model with usage charges above thresholds.
• 11 Labs followed the same pattern, moving from subscriptions to pay-as-you-go.
• Agents can consume at machine speed, so real-time metering and real-time billing are needed.
• Stripe and Tempo are co-building streaming payments for token consumption.
• Traditional accounting systems struggle with AI microtransactions because of row proliferation.
• AI accounting work is shifting toward a hybrid accountant-engineer role focused on anomalies and abuse.
• Token theft is an under-discussed AI fraud category because stolen tokens can be resold or wrapped into products.
• More than 1 in 6 signups at AI companies are multi-account abuse.
• Free-trial abuse has more than doubled on Stripe, mostly because of AI businesses.
• Usage-based dine-and-dash is an abuse mode where users rack up token costs and never pay the invoice.
• Fraud teams are building real-time APIs that score accounts at login, free-trial start, and during usage accumulation.
• Stripe says most current AI fraud is first-party abuse rather than classic credential or payment fraud.
• As agents become buyers, companies need self-serve and PLG motions.
• Stripe says its AI fraud defenses depend on network density, not just model quality.
• 58% of Lovable’s volume flows through Link.
• The Machine Payments Protocol lets agents request a service, receive a payment request, and pay without checkout UI or a human in the loop.
• Streaming payments combine Metronome real-time tracking with Tempo instant settlement in stables.
• Stripe expects agents to become multifaceted economic actors that buy, sell, provision infrastructure, and run businesses.
• Stripe Directory and Stripe Projects are meant to help agents discover providers and integrate them directly.