Agentic Commerce: The CFO's New Control Problem

The CFO Didn't Approve the Agent — But the Agent Just Approved the Spend

Visa processed $3.3 trillion in volume last quarter. This Cyber Monday, an unknowable percentage of those transactions had no human finger on the "buy" button — and no executive remembers greenlighting that shift.

Visa built something called Intelligent Commerce. Mastercard launched Agent Pay. The card networks aren't testing this in a lab. They've already built the plumbing for machines to pay machines, and they're lighting it up now.

I've spent the week after Thanksgiving talking to finance leaders who just realized the gap: the agents their teams spun up to "save time on procurement" are now making purchase decisions. Not recommendations. Decisions. And nobody thought to brief the CFO.

Who Decided That the Agent Could Decide?

Here's what's easy to miss. When retail moved to mobile, it changed where people bought things — couch instead of mall, thumb instead of mouse. But it never changed who decided. A person still tapped "buy." Every fraud control, every approval workflow, every audit trail we built in the last twenty years assumed that tap existed.

Agents take the tap away.

Pantera Capital published an essay on agentic commerce that drew a line I keep coming back to. Most agents today act for a person. They borrow your identity, your credit card, your legal standing. You're still the principal. The agent is just faster hands.

The next step — the one the card networks are building toward — is agents paying other agents, with no human identity behind the transaction at all. Each step moves the purchase further from the person who answers for it when the auditor shows up.

And make no mistake: the auditor will show up.

The Questions Nobody's Asking (Until the Spend Report Arrives)

Ravi Loganathan at Sardine laid out three questions that should be in every CFO's deck by January. I'm rephrasing them slightly, but the thrust is the same:

• How do you know the agent is acting within the scope of authority you actually granted it?

• How do you tie each payment to a verified identity when the "buyer" is a line of code?

• How do you stop fraud against the agent — or by it?

Read those again. Not one of them is a payment question.

They're governance questions, and they land on the finance leader's desk whether or not anyone asked finance to approve the pilot. What this agent may buy. Up to what limit. Who handles the exception when the agent tries to buy something outside its lane. What an auditor can rebuild six months later when the question is "who authorized this?"

Your procurement agent won't get distracted by a doorbuster sale. It also can't tell you why it ordered 400 toner cartridges instead of 40.

We've Seen This Movie Before

In 1999, companies rolled out corporate cards to speed up T&E. No more expense reports for every taxi ride. Easier for everyone. Then six months later, finance realized: we gave every traveler a card, but we never updated the policy, the monitoring, or the reconciliation process. We automated the transaction and left the controls behind.

The fix took years. Amex and the big issuers built entire businesses around retroactive control layers — transaction-level data feeds, policy engines, exception workflows. The gap was expensive.

Agentic commerce is that gap, but faster and harder to see. The agent "lives" inside your Slack or your ERP. It feels like a productivity tool, not a payment instrument. By the time finance notices it's making purchases, it's already woven into twenty workflows across four departments.

Nobody gets fired the day the agent goes live. The damage shows up in Q2, when someone asks "why did our software spend jump 40% and who approved it?"

The New Job: Coaching Agents, Not Reviewing Receipts

Phil Jackson never ran a play himself. He set the conditions his players executed within — spacing, tempo, reads. That's the CFO's job now.

You're not approving every transaction. You can't. The agent operates at machine speed. What you can do — what you must do — is set the boundaries before the agent starts spending:

• Authority limits. Not just dollar thresholds. Category limits. This agent can buy cloud services up to $5,000/month. It cannot buy hardware. It cannot buy subscriptions longer than one year. Specific.

• Approval hooks. What triggers a stop-and-check? When does the agent have to surface a decision to a human before proceeding? If you don't define this, the agent will optimize for speed — which means it'll default to "yes."

• Audit trail by design. What does the agent log, and where? If your ERP can't ingest machine-generated metadata about why the purchase happened, your audit is already compromised.

The companies that get this right in 2025 will be the ones that treat agentic purchasing like a new vendor relationship — complete with onboarding, delegation of authority documentation, and a quarterly review. The ones that get it wrong will treat it like a browser extension.

What This Looks Like Monday Morning

This isn't theoretical. If your firm is running agents that touch procurement, travel, software subscriptions, or vendor payments, here's what to do this week:

1. Inventory your agents. Not the ones IT approved. The ones your teams actually deployed. Slack agents, Zapier workflows, ChatGPT plugins with access to credit cards. You need the list.

2. Map their permissions. Can they initiate spend? Up to what amount? In what categories? If the answer is "we think so, but we're not sure," you have a control gap.

3. Define the policy now. What are agents allowed to buy without human approval? Write it down. Circulate it. Make it as specific as your existing procurement policy, because it is your procurement policy now.

4. Set the review cadence. Monthly reconciliation won't cut it. Agents move faster than that. You need real-time or near-real-time visibility into what they're doing — or you need to throttle their permissions until you have it.

The card networks have already decided this is the future. Visa and Mastercard didn't build Agent Pay as a concept demo. They built it because they see the transaction volume coming. Your job is to make sure the governance arrives before the spend does.

This November it was your doorbuster impulse purchase. Next November it might be your office's toner cartridge order — times four hundred, because the agent optimized for "never run out" and nobody told it about budget constraints.

But what do I know — I've only watched finance teams clean up after the technology team's "productivity improvements" four times now.

---

What are you doing to govern the agents your teams are already running? If the answer is "we're still figuring that out," you're not alone — but you're also not early. The agents are already live. The question is whether your controls are.