AI agents need permissioned funding, not separate bank accounts

AI agents may not need to hold money. They need authorization to spend it, plus controls that define what they can buy, how much, and when.

Treasury keeps liquidity centralized, agents get programmable authority

A central funding pool can remain with Treasury while software receives permission to create approved obligations. The control layer decides identity, limits, counterparties, purpose, and which payment rail to use at the moment of payment.

Agentic payments turn Treasury into a real-time routing engine

Transaction-by-transaction selection can route funding across credit, cash, and settlement options based on cost, liquidity, FX, rebates, and risk. In this model, Treasury acts as a live decision layer rather than a place where cash sits for each agent.

Virtual cards offer an early path for agentic commerce

Virtual cards generated programmatically can be restricted by merchant and limited by transaction amount or time period. They also support reconciliation against specific purchases, while underlying capital stays with the issuer until a transaction occurs.

Because cards already connect to merchant acceptance networks, agents can execute purchases without relying on a new settlement rail. This structure fits early use cases like airfare, advertising, and office supplies.

Stablecoins may help, but not every agent needs its own balance

On-chain payments can work for frequent, small purchases such as compute, API calls, data acquisition, content licensing, or compensating other autonomous services. Even so, the source questions whether each agent requires a dedicated pile of stablecoins.

A more capital-efficient approach mirrors corporate cash concentration: Treasury maintains a central stablecoin balance, while agents receive virtual wallets or sub-ledgers representing claims. Policy engines then control which agents can access which amounts, counterparties, and transaction types. The same architecture could apply to tokenized deposits or conventional bank money.

Authorization to buy can converge with authorization to borrow

The most disruptive funding mechanism described is purchasing power at the moment of need rather than prefunding. The source cites Visa announcing on Tuesday (Sept. 8) a new credit and funding product for its stablecoin-backed card suite, shifting the decision from whether a specific agent has $100,000 available credit to whether an agent can be permitted to create an $18,700 liability for a particular purchase at a specific time.

CFO adoption signals a readiness gap

A PYMNTS Intelligence report in September 2025 found nearly 7% of enterprise CFOs in the United States had deployed agentic AI in live finance workflows, with an additional 5% running pilots. Another PYMNTS Intelligence report cited a widening agentic readiness gap, with 75% of tech firms extremely familiar with agentic AI versus 33% of goods firms and 38% of services firms.

Why this matters

Agentic commerce depends less on where money sits and more on who can authorize obligations. Central liquidity paired with permissioned spending controls could reduce cash fragmentation and make payments execution more flexible as agent activity scales.