On March 11, we shipped Agent Cards on Visa Intelligent Commerce. On March 31, we announced the Visa Trusted Agent Protocol. On April 29, the Procurement Fleet went live.

Then I decided to actually run Ramp on the thing we built.

For 90 days, I let the agent make decisions I would normally make myself. Not all of them. There are categories where human judgment still matters. But the ones that can be rule-governed, policy-enforced, and verified? The agent got them.

What 26.1 million decisions looks like from the inside

The number sounds abstract until you break it down. 26.1 million autonomous finance decisions across our network in one month isn't a benchmark. It's a baseline. The interesting question isn't the volume. It's the four places the agent got it wrong.

Failure mode 1: Ambiguous vendor categories. The agent has no trouble with known vendors and known categories. It gets confused when a vendor spans categories: a contractor who is also a software tool. We've since trained the policy layer to ask one clarifying question before categorizing, rather than guessing.

The three things we'd never go back from

Idle cash management. The agent swept $5.5M into a 4% treasury yield without anyone asking. The team was asleep. The money earned overnight returns it never would have under the old system. Nobody asked it to do that. It just noticed.