Source: Business Wire / Mercury — 2026-08-11
Summary
Mercury's business-banking platform launched "Spend," including "Agent Cards" — virtual payment cards created by a human but usable autonomously by an AI agent to complete purchases end-to-end without per-transaction approval. Each agent gets its own credential, spending limits, and audit trail, functionally treated like giving a new hire their own corporate card, but for software. Spending limits can't be altered by the agent itself, and cards remain human-created, human-trackable, and human-cancellable at any time; the launch also ships self-enforcing expense policies and spending pools that apply to human employee cards too.
Key Takeaways
- Agent Cards give AI agents their own regulated, auditable spending identity — a real financial credential separate from any human's card — rather than the common workaround of an agent using a borrowed or shared payment method.
- Spending limits are enforced outside the agent's control: the agent can spend within its limit but cannot raise, alter, or bypass that limit itself, keeping the human as the ultimate authority.
- Every Agent Card carries its own audit trail, so individual agent purchases are traceable back to a specific card and, presumably, a specific task or workflow.
- Cards are human-created and human-cancellable at any time, meaning the off-switch is never delegated to the agent.
- The same release ships self-enforcing expense policies and spending pools (travel, software, procurement) for human employee cards too — Mercury is unifying human and agent spend controls under one system rather than bolting agent support onto a separate product.
Reel Script
Hook (16s)
An AI agent can now book a flight, buy software, or pay a vendor with its own corporate card — no human clicking "approve" on each purchase. Mercury just made that a regulated banking product.
Core Concept (85s)
Up to now, if you wanted an AI agent to actually spend money — buy ad credits, subscribe to a tool, book travel — you had two bad options: give it a human's card number, which is a security and audit nightmare, or make a human approve every single transaction, which kills the whole point of automation. Mercury's fix is called an Agent Card: a virtual payment card that a human sets up once, but that the agent then uses on its own to complete purchases end-to-end. The mental model is simple — treat the AI agent like a new hire. You wouldn't give a new employee your personal credit card; you'd issue them their own card with a defined limit and a paper trail. Agent Cards do exactly that for software: each agent gets its own credential, its own spending cap, and its own transaction history, so nothing is borrowed and nothing is invisible.
Hands-On (110s)
Here's the control structure worth sketching: a human creates the card and sets the limit — say, a monthly cap for a procurement agent or a per-purchase cap for a travel-booking agent. From that point on, the agent can spend freely within that boundary without asking permission for each transaction. But the critical design choice is what the agent cannot do — it cannot raise its own limit, cannot alter the card's rules, and cannot prevent a human from canceling it instantly. That asymmetry is the whole security model: full autonomy inside the box, zero autonomy over the box itself. Layered on top, Mercury also shipped self-enforcing expense policies and spending pools — shared budgets for categories like travel, software, or procurement — and notably these same pools apply to human employee cards too, so a company isn't running two separate systems for people and agents, it's one spend-control system with two kinds of cardholders.
Takeaway (22s)
This is one of the first mainstream banking products treating an AI agent as a real, auditable financial actor instead of a script borrowing someone's login. If your team is deploying agents that need to actually transact, this is the control model to copy — full spend autonomy, zero limit autonomy.