How does Ramp use AI across its payment ecosystem?
Ramp runs AI at three points in a corporate payment: when a card is authorized, when a bill is approved, and when an AI agent is the buyer. Its latest card fraud model, described on October 7, 2026, adds a small transformer that reads each cardholder's recent transactions to an XGBoost model and runs in real time on every Ramp card authorization; run side by side with its predecessor, it caught 32% more fraud dollars while flagging the same number of transactions. On bills, agents for accounts payable, launched October 7, 2025, use more than 60 fraud signals to flag suspicious invoices before approval; Ramp says they identified more than $1 million in fraud. For agents as buyers, Ramp Agent Cards, in early access in April 2026, are tokenized credentials issued through Visa Intelligent Commerce and scoped to a single agent and transaction, and Ramp for Agents lets agents work across cards, bills, expenses and approvals through MCP and a CLI, under an agent identity tied to a human sponsor. On March 31, 2026 Ramp and Visa renewed a multi-year issuing agreement and said they would introduce AI agents for corporate bill pay using Visa Intelligent Commerce and the Visa Trusted Agent Protocol. Ramp's card program runs on Stripe Issuing. Ramp does not disclose fraud-loss rates or agent transaction volumes.[1][2][3][4][5][6]
Data as of Oct 19, 2026 · Ramp profile updated Oct 19, 2026 · 14 sources cited on this page · 52 fintechs profiled on this site · Markdown
Where in the payment flow does Ramp use AI?
One row per layer, with Ramp's product name and the latest dated figure Ramp has published for it. All results are Ramp's own measurements.
| Layer | Product | What the record shows | Date |
|---|---|---|---|
| Card authorization fraud scoring | Card fraud model (transformer plus XGBoost) | Runs in real time on every Ramp card authorization; caught 32% more fraud dollars than the previous model at the same flag budget, with average precision 12% higher than the same XGBoost without the transformer | October 7, 2026[1] |
| How the fraud model learns | Card fraud model | The transformer is pre-trained without fraud labels on two years of card transactions, in under an hour and a half on four L40S GPUs; about 47,000 tokens represent millions of merchants | October 7, 2026[1] |
| Fraud scoring latency | Card fraud model | After a rollback and fixes, the step that runs Ramp's card fraud models went from 145 to 65 milliseconds at the median; a week later it ran at 54 milliseconds at the median and 84 at p99 | October 7, 2026[1] |
| Expense policy review | Policy agents | Review expenses against company policy; Ramp says they catch 15x more out-of-policy spend than non-AI alternatives, with 99% accuracy, and escalate 10-15% to humans | July 9, 2025[7] |
| Invoice fraud before approval | Agents for AP | Trained on hundreds of thousands of invoices and use 60-plus fraud signals; identified $1M+ in fraud before invoices were sent for approval, based on Ramp's September 2025 internal data | October 7, 2025[2] |
| Paying bills by card | Agents for AP | Surface bills that accept card and input card details directly in the vendor's payment portal | October 7, 2025[2] |
| Purchases by AI agents | Ramp Agent Cards | In early access; tokenized credentials issued by Visa Intelligent Commerce, each scoped to a single agent and transaction, running on the Visa network | April 28, 2026[3] |
| Agent card controls | Ramp Agent Cards | A single-use card number and CVV minted at purchase time, scoped to one merchant, capped at the requested amount, and expiring after the first authorization or 12 hours | Undated docs, retrieved October 19, 2026[8] |
| Agent access and attribution | Ramp for Agents, agent identity | CLI and MCP interfaces let agents work across Ramp cards, bills, expenses and approvals; each agent is tied to a human sponsor, can never have more access than that sponsor, and appears in logs as, for example, "Sarah (via Codex)" | May 4, 2026[4] |
Who does Ramp work with on AI in payments?
| Partner | What | Date |
|---|---|---|
| Visa | Renewed multi-year issuing agreement, plus work with Visa Intelligent Commerce and the Visa Trusted Agent Protocol to introduce AI agents that automate corporate bill pay; Visa will also use Ramp for some corporate services | March 31, 2026[5] |
| Visa | Ramp says Visa is becoming a Ramp customer while the two deploy Visa Intelligent Commerce and the Trusted Agent Protocol across Ramp's 50,000-plus customers | March 31, 2026[9] |
| Visa (as reported) | American Banker reported that Visa and Ramp will use Visa's trusted agent protocol and Ramp's network of more than 50,000 corporate clients | April 2, 2026[10] |
| Stripe | Ramp chose Stripe as its issuing partner shortly after its early-2020 launch; Ramp built custom authorization logic on Stripe Issuing and uses it to evaluate spending context and make real-time decisions | Undated page, retrieved October 19, 2026[6] |
| Anthropic's Model Context Protocol | Ramp built an MCP server on its developer API; with Claude Desktop as the client, a prototype issued cards on demand for a company offsite | March 25, 2025[11] |
| Agent clients | Agent Cards work through MCP runtimes such as Cursor, Claude, ChatGPT or Perplexity, or through the Ramp CLI | Undated docs, retrieved October 19, 2026[8] |
| Coding agents | Ramp for Agents works with coding agents that support skills, including Claude Code, Codex and ChatGPT, and lets founders apply for Ramp's corporate card from them | July 6, 2026[12] |
What has Ramp disclosed about its fraud model's results, and what does it not disclose?
Most of the card fraud model's 32% gain in fraud dollars comes from a dollar-weighted training loss: XGBoost retrained with that loss alone caught 29% more fraud dollars than the previous model. Ramp says the transformer is what improves the ranking, and that on fraud transactions its outputs carry about 15% of the score.[1]
By size, the new model caught 16% more of the fraud between $100 and $1,000 and 34% more of the fraud of $1,000 or more, a small group that holds three quarters of the fraud dollars; on the first fraudulent charge in a run on one card, it caught 40% more of the fraud dollars than the previous model.[1]
The first launch was rolled back: the model call took about twice as long as the one it replaced and too many authorizations hit the deadline. Ramp says the fixed build matched the original's scores on all 80,149 live authorizations it compared and went live less than a day after the rollback.[1]
What is not disclosed: Ramp's fraud-loss rate or the dollar amount of card fraud its models stop, how many Agent Cards have been issued or what they spend, and results from the agentic bill pay work announced with Visa. The $1M+ in invoice fraud is Ramp's own count of flagged bills that users then deleted.[1][3][5][2]
What does the Ramp profile record on this topic?
The rows below are the entries of the Ramp AI strategy profile that bear on this page, as of Oct 19, 2026; the profile holds the full record of 17 dated moves, 10 use cases and 22 sources.
| Use case | What the record shows | Status | Source |
|---|---|---|---|
| Policy agents for expense review | Agents that review expenses against company policy and escalate 10-15% to humans; available to Ramp Plus customers through the Early Access tab at launch. | Rolling out | [7] |
| Agents for accounts payable | Invoice coding, approval recommendations and card payments; customers enable them or join a waitlist in the Early Access tab. | Rolling out | [2] |
| Invoice fraud checks by AP agents | AP agents check invoices for fraud before approval; Ramp says they identified more than $1 million in fraud over their first few months. | Rolling out | [2] |
| Agentic bill pay with Visa | AI agents to automate corporate bill pay using Visa Intelligent Commerce and the Trusted Agent Protocol, announced March 2026. | Announced | [5] |
| AI token spend management and Router.com | Tools that show companies what they spend on AI models and route tasks to cheaper models; token spend management launched in early access in April 2026 and Router.com opened to all developers in August 2026. | Rolling out | [13][14] |
| Date | Move | Detail | Source |
|---|---|---|---|
| Aug 19, 2026 | Router.com opens to all developers | One API to access, evaluate and route traffic across AI models, based on routing infrastructure Ramp has run on its own production workloads for three years; Ramp says it saved customers 40% on cost. | [14] |
| Mar 31, 2026 | Visa partnership extended to agentic bill pay | A renewed multi-year issuing agreement and work with Visa Intelligent Commerce and the Visa Trusted Agent Protocol to introduce AI agents that automate corporate bill pay. | [5] |
| Oct 7, 2025 | Agents for accounts payable | Agents that code invoice line items, check for fraud, recommend approvals and complete card payments; Ramp says they got 85% of accounting fields right the first time, identified $1M+ in fraud before approval and issued 90,000 approval recommendations with a 90% acceptance rate. | [2] |
| Jul 9, 2025 | Ramp launches its first AI agents | Policy agents for Ramp Plus customers that review expenses; Ramp says they catch 15x more out-of-policy spend than non-AI alternatives, enforce policy with 99% accuracy and escalate only 10-15% of expenses to humans. | [7] |
How does Ramp detect card fraud?
Ramp scores every card authorization in real time with a model that combines a small transformer, which reads the cardholder's recent transactions and scores how unusual the new one is, with an XGBoost model. Ramp said on October 7, 2026 that it caught 32% more fraud dollars than its predecessor at the same number of flagged transactions.
What are Ramp Agent Cards?
Virtual cards that let AI agents pay. Ramp says they are tokenized credentials issued through Visa Intelligent Commerce, scoped to a single agent and transaction; its developer docs describe a single-use card number scoped to one merchant, capped at the requested amount and expiring after the first authorization or 12 hours. They were in early access in April 2026.
What is Ramp doing with Visa on AI?
On March 31, 2026 Ramp and Visa renewed a multi-year issuing agreement and said they would introduce AI agents that automate corporate bill pay using Visa Intelligent Commerce and the Visa Trusted Agent Protocol. Ramp said Visa is also becoming a Ramp customer. No results from the agentic bill pay work have been published.
Do Ramp's AP agents check invoices for fraud?
Yes. Ramp's agents for accounts payable, launched October 7, 2025, are trained on hundreds of thousands of invoices and use more than 60 fraud signals to flag suspicious bills before they are routed for approval. Ramp says they identified more than $1 million in fraud, based on bills users deleted after a flag in September 2025.
How do AI agents connect to Ramp?
Through Ramp for Agents, which offers an MCP server and a command-line interface covering cards, bills, expenses and approvals. Each agent gets an identity tied to a human sponsor, with access scoped to no more than that sponsor's role, expiry by default, and audit logs that show the agent acting on the sponsor's behalf.
- Putting a Transformer Inside Every Card Authorization
- AP just became autonomous
- Virtual cards for AI agents: How they work and what to look for
- Agentic identity: modeling agents to keep users in control
- Ramp and Visa Deepen Partnership to Power the Next Era of Autonomous Finance
- Behind the scenes: How Ramp Built a Global Corporate Card Program (undated page, retrieved 2026-10-19)
- Ramp agents: Let finance teams do finance
- Agent cards overview (undated page, retrieved 2026-10-19)
- The world's most demanding companies run on Ramp
- US card networks accelerate bets on agentic AI
- Building Ramp's MCP server
- Builders can now access financing and incorporation via Ramp on day 0
- Ramp raises $750 million, plans AI spending software
- Introducing Router.com: The best model changes. Your application shouldn't.
When Ramp moves on AI, you'll read it here first.
when a bank, fintech or regulator moves on AI, it is in the next morning's brief · six sourced stories · 7 am PT · free
plus every regulation, bank and agent page update, the morning after · leave any morning