AI Agents and Corporate Spend: The Coming Decade of Agentic Payments

An AI agent buying a 39 USD GitHub Copilot seat for a new engineer should not require a human approval. An AI agent buying a 12,000 USD enterprise software contract absolutely should. The hard part of agentic payments is not the buying. It is the boundary.

What "agentic payments" actually means

An agentic payment is a transaction initiated by an AI agent on behalf of a human principal, under a pre-authorized scope. The scope (the "mandate") defines what the agent may buy, from whom, up to what amount, and for how long. The mandate is enforced at authorization time by the issuer or the platform, not by trust in the agent's prompt.

This is fundamentally different from "AI helps a human fill out a form". The agent is the buyer. The card is the agent's. The audit log identifies the agent, not a human, as the actor.

The AP2 (Agent Payments Protocol) backdrop

In 2025, major payment networks and several wallet providers proposed AP2, an emerging standard for representing, transmitting, and revoking agent mandates between buyers, merchants, and processors. AP2 defines a mandate format, a verification flow, and a revocation path. It is not yet universally supported, but the direction is clear: every issuer will need to represent agent mandates in the next 24 months or risk losing the agentic-commerce traffic to those who do.

What changes for finance teams

Three things change practically:

1. A new actor class in the audit log. Today, every transaction has a human owner. Tomorrow, transactions will have a human principal and an agent actor. Your audit trail needs to capture both.
2. A new approval boundary. "Approval" historically meant a person clicking approve. With agents, the approval happens at mandate issuance, not at transaction time. Finance approves the agent's authority once, the agent executes within it many times.
3. A new failure mode. Agents fail differently than humans. A misconfigured agent can submit 10,000 transactions in an hour. Velocity limits, not just amount limits, become primary controls.

What to do today, before AP2 is universal

You do not need to wait for AP2 to start operating agentically. Three concrete steps:

Issue agent-specific virtual cards. When a new agent is deployed (a code assistant, a procurement bot, a marketing automation), issue it a dedicated virtual card. Tag the card with the agent name. Every transaction is automatically attributed.

Set tight merchant allow-lists. A documentation agent might only need access to OpenAI, Anthropic, and Pinecone. A research agent might need search APIs and a small list of data vendors. Never give an agent a general-purpose card. The blast radius of a hijacked agent is the blast radius of its card.

Set hard velocity limits. Not just per-day amount, but per-hour transaction count. A reasonable agent does not submit 100 transactions in an hour. If yours does, the agent is broken or compromised, and the card should auto-pause.

The contrarian take

Most "agentic payments" coverage in 2026 imagines a future where AI agents do everything from booking flights to negotiating supplier contracts. That future will come, but not soon. The agents shipping today are narrow: code copilots, content generators, support automation, basic research. Their financial footprint is small individual transactions to a small set of vendors.

The right finance posture today is to operationalize the small case well, not to architect for the imagined enterprise scenario. Get one card per agent, tight allow-lists, velocity caps, and clear audit attribution. Then expand as the agent population grows and the protocols mature.

What we are building

At FlyExpense, every virtual card can be tagged as an agent card with a structured mandate (subject, merchants, MCC categories, per-transaction limit, daily cap, hourly velocity cap, expiry). We export an AP2-compatible mandate representation for issuers that support it, and fall back to internal enforcement for those that do not. Every transaction is attributed in the audit log to the agent and the human principal who issued the mandate.

The goal is not to make agentic payments magical. It is to make them as boring and observable as a corporate card. That is what production finance looks like.