Building Economic Infrastructure for AI with Foundation Models and Agentic Commerce

stripe · Finance · 2025 · source Agentic: yes · promotional bias: med · source-backed · extracted with prompt 0.5.5 (schema v2)

Generated from extraction.json — do not edit by hand. Every claim cites verbatim source evidence; stated means the quote directly asserts the claim, implied means it supports it indirectly.

Problem

Domain: Economic infrastructure for AI and agentic commerce

Stripe needed to adapt payments, fraud, billing, data access, and merchant integration for an AI economy where agents buy on behalf of users, AI companies have volatile inference costs, fraud can burn expensive GPU usage, and internal employees increasingly use LLM tools against company systems. The case combines external infrastructure such as ACP, shared payment tokens, Radar extensions, and Token Billing with internal AI infrastructure such as Toolshed MCP, Hubert text-to-SQL, and semantic data pipelines.

Constraints:

• AI companies face new fraud vectors where free trial or refund abuse can consume expensive inference rather than near-zero marginal SaaS cost.
• Agentic commerce needs a standard way for agents to discover merchant catalogs, prices, brand constraints, and payment flows.
• Payment credentials must not be exposed to agents, and merchants still need fraud/risk signals for good-bot versus bad-bot decisions.
• Payment fraud models must run on the charge path with very low latency while processing every transaction.
• Internal LLM tools need access to data and tools, but text-to-SQL depends on high-quality documentation, table discovery, and human validation.

Architecture

Stripe's AI infrastructure has several connected layers. Its payments foundation model treats charges as dense embeddings in behavior sequences, processes every transaction on the charge path, and powers Radar fraud and clustering use cases. Agentic Commerce Protocol standardizes how agents read merchant catalog, inventory, pricing, and brand information, while Shared Payment Tokens let agents complete purchases without seeing underlying credentials. Radar risk signals travel with the token so merchants can decide whether the agentic transaction is trustworthy. Token Billing tracks inference costs in real time for AI businesses. Internally, Toolshed acts as a centrally managed MCP-like tool layer over Slack, Drive, Git, Hubble, query engines, and other systems; Hubert and Sigma Assistant provide text-to-SQL experiences over internal and customer Stripe data; and semantic events work aims to make near-real-time canonical data available to dashboards, Sigma, data exports, and LLM agents.

Diagram confidence: reconstructed (reconstructed from prose).

flowchart LR
  stripe-payments-data[Stripe payments data] --> payments-foundation-model[Payments foundation model]
  stripe-payments-data --> semantic-events-pipeline[Semantic events pipeline]
  payments-foundation-model --> payment-embeddings[Payment embeddings]
  payment-embeddings --> radar-fraud-system[Radar fraud system]
  radar-fraud-system --> radar-risk-signals[Radar risk signals]
  radar-risk-signals --> shared-payment-token[Shared payment token]
  radar-risk-signals --> merchant-checkout[Merchant checkout]
  ai-buying-agent[AI buying agent] --> agentic-commerce-protocol[Agentic Commerce Protocol]
  agentic-commerce-protocol --> merchant-catalog[Merchant catalog]
  merchant-catalog --> ai-buying-agent
  agentic-commerce-protocol --> shared-payment-token
  ai-buying-agent --> shared-payment-token
  stripe-link-wallet[Stripe Link wallet] --> shared-payment-token
  shared-payment-token --> merchant-checkout
  token-billing-api[Token Billing API] --> ai-company-pricing-model[AI company pricing model]
  internal-llm-interface[Internal LLM interface] --> toolshed-mcp-server[Toolshed MCP server]
  toolshed-mcp-server --> company-tools[Company tools]
  toolshed-mcp-server --> hubble-data-catalog[Hubble data catalog]
  hubble-data-catalog --> hubert-text-to-sql[Hubert text-to-SQL]
  hubert-text-to-sql --> human-validator[Human validator]
  hubert-text-to-sql --> semantic-events-pipeline
  semantic-events-pipeline --> dashboard-and-sigma[Dashboard and Sigma]

Components

Data flow

Payments and merchant behavior data flow into Stripe's domain-specific foundation model, where charges become embeddings and behavioral sequences provide context for Radar decisions and suspicious-transaction clustering. In agentic commerce, an AI buying agent uses ACP to understand a merchant catalog and then passes a Shared Payment Token that is scoped, time-limited, amount-capped, revocable, and enriched with Radar risk signals; the merchant processes the charge and maintains the consumer relationship. Internally, employees and agents access company tools through Toolshed, query Hubble-backed data through Hubert, and increasingly rely on semantic events so the same near-real-time canonical feed powers Dashboard, Sigma, Stripe Data Pipeline, and future agent-facing insights.

Why this architecture

Stripe uses domain-specific models and protocols because payments and agentic commerce require trust, latency, and standardization that generic LLM wrappers do not provide. Payment embeddings exploit Stripe's transaction network to detect subtle fraud clusters in real time. ACP avoids custom merchant integrations per agent, while Shared Payment Tokens preserve payment credentials and direct merchant relationships. Internally, Toolshed goes beyond pure RAG by allowing LLMs and agents to retrieve from and act on company systems; semantic events aim to give those systems high-quality, near-real-time canonical data rather than fragmented dashboards and exports.

Models used: Stripe payments foundation model · LLMs · GPT-3.5 · GPT-4 · GPT-5 · GPT-4o · O3 Mini · Braintrust · LibraChat

Scale

• Stripe processes about $1.4 trillion annually, roughly 1.3% of global GDP.
  “we’re looking at on the order of 1.3% of global GDP. About $1.4 trillion a year is processed on Stripe.”article, stated
• Stripe sees around 50,000 transactions per minute.
  “we see, like, 50,000 transactions a minute.”article, stated
• Every transaction goes through the foundation model on the charge path in under 100 milliseconds.
  “Every single Stripe transaction goes through this model in <100ms.”article, stated
• Card-testing detection improved from 59% to 97% on large users.
  “The card testing detection numbers went from 59 to 97 percent.”article, stated
• Stripe Link has more than 200 million consumers.
  “The ACP is being coupled with Stripe’s Link product, which has 200m+ consumers”article, stated
• The initial Instant Checkout launch mentions over 1 million Shopify merchants coming soon plus Walmart and Sam's Club adoption.
  “There’s over 1 million Shopify merchants coming soon”article, stated

  “Walmart and Sam’s Club have just signed up to also make their inventory purchasable through ChatGPT and the Agentic Commerce Protocol”article, stated
• About 8,500 Stripe employees use LLM-based tools daily, and 65-70% of engineers use AI coding assistants day to day.
  “8,500 stripes a day use our LLM based tools.”article, stated

  “65, 70% of engineers use them on the day-to-day.”article, stated
• Hubert is used by about 900 people per week.
  “we have like 900 people who use it a week.”article, stated
• A pan-European local payment method integration went from about two months to two weeks using an LLM.
  “it takes like two months for a couple of engineers.”article, stated

  “they just launched a new pan-European payment method in two weeks using an LLM”article, stated
• The top 100 AI companies on Stripe reach revenue milestones two to three times faster and are in 55 countries by the end of year one.
  “how quickly do they get to a million or 10 million or 30 million in ARR? And the answer is two to three times as fast as a SaaS cohort.”article, stated

  “55 countries. They have more. Majority of their revenue coming from outside their home market”article, stated

Alternatives considered (roads not taken)

Use task-specific fraud models and hand-engineered features.

Rejected because: Stripe moved toward a domain foundation model that can embed full payment sequences and power many downstream applications.

Build custom merchant integrations for each agent platform.

Rejected because: ACP gives merchants one standard protocol that can work across many present and future agentic buying experiences.

Give agents direct wallets or expose buyer credentials.

Rejected because: Stripe uses Shared Payment Tokens or one-time virtual cards so agents can transact within scope without seeing card credentials or bearing transaction risk.

Use static pricing for AI products despite volatile model costs.

Rejected because: Token Billing lets AI companies track and price to inference cost in real time when upstream LLM costs fall or spike.

Rely on pure RAG for internal knowledge.

Rejected because: Stripe's Toolshed gives LLMs both retrieval and action across internal tools, which the interview frames as complementary to RAG.

Keep separate data feeds for Dashboard, Sigma, and exports.

Rejected because: Stripe is moving toward one near-real-time canonical semantic-events feed so LLMs and users see the same high-quality data.

Patterns

tool-use-orchestration (agent-integration)

Stripe's internal Toolshed exposes company systems as tools so LLMs and agents can both retrieve information and act across Slack, Drive, Git, Hubble, query engines, and team-owned integrations.

privacy-preserving-processing (payments)

Shared Payment Tokens and virtual-card flows let agents spend within constraints without exposing the buyer's underlying card credentials.

safety-policy-enforcement (risk)

Radar and ACP risk signals distinguish good bots from bad bots so merchants can accept beneficial agentic demand while blocking fraud, scalping, and suspicious traffic.

cost-aware-routing (billing)

Token Billing lets AI businesses price against real-time model-token costs so margins adapt when upstream LLM costs change.

rag-retrieval (internal-data)

Hubert and Sigma Assistant let users ask natural-language questions over Stripe data, with query generation, natural-language query explanation, and data-catalog grounding.

data-quality-loop (internal-data)

Stripe improves text-to-SQL reliability by deprecating low-quality tables, documenting canonical datasets with human owners, and experimenting with persona context for table selection.

human-in-the-loop-gate (quality)

Stripe keeps humans responsible for validating high-risk AI-generated SQL, documents, PRs, and fraud decisions rather than treating LLM output as self-authorizing.

Tradeoffs

Publish ACP as a standard protocol.

Chose open agentic commerce protocol over Stripe-only or agent-specific integrations.

Rationale: Merchants need one integration that can work across many agents and payment providers as the buying-agent ecosystem evolves.

Use Shared Payment Tokens for ChatGPT instant checkout.

Chose scoped payment token over agent holding raw credentials or becoming transaction merchant of record.

Rationale: The token lets the merchant process the charge and maintain the consumer relationship while the agent avoids payment-risk custody.

Keep multiple agent payment mechanisms alive.

Chose SPT, virtual cards, wallets, stablecoins, and stored balances depending on use case over one universal payment primitive.

Rationale: Physical-goods commerce, travel booking, microtransactions, and agent-to-agent payments have different economics and risk shapes.

Use domain foundation modeling on the charge path.

Chose payments foundation model embeddings over small task-specific fraud models.

Rationale: Stripe's transaction volume allows one richer representation to support many downstream fraud and optimization tasks in real time.

Give Toolshed decentralized extension points while keeping it centrally managed.

Chose central MCP layer with team-added tools over every tool request blocked on a central team.

Rationale: The central layer provides consistency while individual teams can expose domain tools such as Salesforce without waiting for the core owner.

Treat build-versus-buy as sequential and revisitable.

Chose build when no safe provider exists, then replace or buy when ecosystem matures over permanent homegrown ownership or premature vendor lock-in.

Rationale: Stripe built early LLM access when enterprise-safe options were missing, later moved to open-source LibraChat, and uses RFPs/POCs for tooling such as evals.

Gotchas (what bit them)

AI fraud can crush unit economics because abuse consumes expensive inference.

Root cause: Free trial, refund, and nonpayment abuse now burn GPU and model costs rather than only consuming near-zero marginal SaaS service.

Resolution: Stripe extends Radar and payment risk tooling for AI-specific abuse while trying to preserve legitimate growth loops like free trials.

Traditional fraud models missed card-testing hidden inside large merchant volume.

Root cause: Attackers can sprinkle small-value tests into high-volume legitimate transaction streams.

Resolution: Stripe uses payment embeddings from a foundation model to detect clusters in real time.

Agentic commerce cannot simply allow every bot to buy or block every bot.

Root cause: Some bots represent valuable demand, while others are scalpers, account enumerators, or suspicious actors that may not produce chargebacks.

Resolution: Pass risk scores with the payment token and reason about fraud before checkout where possible.

Text-to-SQL failure is often data discovery, not SQL syntax.

Root cause: Enterprise data contains many nuanced, poorly documented, or low-quality tables, so the model may choose the wrong table or field.

Resolution: Deprecate low-quality tables, document canonical datasets with humans, and use persona context to bias table selection.

LLM-generated work can look polished without preserving expert reasoning.

Root cause: Generation can skip the deep thinking that careful writing or code review used to force.

Resolution: Require LLM citation and human sign-off; judge the underlying reasoning rather than pages of generated output or lines of code.

Static AI pricing can go underwater when upstream model costs spike.

Root cause: Many AI products wrap third-party models whose prices and capabilities change quickly.

Resolution: Use Token Billing or other dynamic monetization mechanisms tied to inference cost.

Evaluation

Stripe evaluates these systems with production business and infrastructure metrics rather than only offline model scores. The payments foundation model is judged by charge-path latency and card-testing detection lift. ACP is judged by merchant/platform adoption and ability to work across payment providers and agents. Token Billing is tied to unit economics under changing LLM prices. Hubert is evaluated through table-selection failures and human/domain validation. Internal AI adoption is measured by daily tool usage, engineering adoption, local payment method integration speed, costs, and perceived productivity rather than lines of code or PR count.

Metrics: card-testing detection accuracy · charge-path latency · transactions per minute · merchant and AI platform adoption · Link consumer reach · risk signal quality for good bots versus bad bots · free trial and refund abuse reduction · LLM tool daily active internal users · AI coding assistant adoption · local payment method integration cycle time · text-to-SQL table and field selection accuracy · human validation of generated SQL · LLM tool cost and business return · AI company revenue milestone speed and country count

Outcomes

• Stripe deployed a payments foundation model that processes every transaction under 100ms
• large-user card-testing detection improved from 59% to 97%
• ACP and Shared Payment Tokens power ChatGPT instant checkout and broader agentic-commerce integrations
• Radar risk signals travel with payment tokens so merchants can reason about good and bad bots
• Token Billing gives AI businesses a way to price against real-time model-token costs
• Toolshed exposes internal systems as LLM/agent tools and supports both retrieval and action
• Hubert provides internal natural-language analytics over Hubble, with about 900 weekly users
• 8,500 Stripe employees use LLM tools daily and 65-70% of engineers use AI coding assistants
• a local payment method integration dropped from roughly two months to two weeks using an LLM
• Stripe is re-architecting payments and usage-billing data around near-real-time semantic canonicals

The case states Stripe processes $1.4T/year or 1.3% of global GDP, sees 50,000 transactions/minute, runs every transaction through the foundation model in under 100ms, improved card-testing detection 59% to 97%, couples ACP to Link's 200M+ consumers, has 8,500 daily internal LLM users and 65-70% engineering coding-assistant adoption, has Hubert at about 900 weekly users, and reports top AI companies reaching $1M/$10M/$30M ARR two to three times faster than SaaS peers.