Agents Are About to Spend $15 Trillion. Who's Watching the Money?
Keel Editorial Team
Research on AI governance, budgets, and auditability
Gartner just put a number on the quiet revolution in purchasing: by 2028, AI agents will intermediate more than $15 trillion in B2B purchases, and 90% of B2B buying will be mediated by an agent rather than a person.
Sit with that.
Within two years, most of what a company buys will be initiated, negotiated, or executed by software acting on its behalf. The reorder, the vendor pick, the payment — increasingly an agent's decision, at machine speed and machine scale.
The question that scale forces
Money leaving a company is the one place where "who did this?" was never optional.
Every dollar is supposed to trace to someone who was allowed to spend it. That is the whole architecture of financial control: authorization, then audit.
An agent breaks the first assumption.
When the buyer is not a person, "who authorized this purchase?" stops being rhetorical and becomes a problem. The agent did, but the agent is not a person. It has no signature, no intent you can examine, no one to ask.
So the question rebounds to you:
What was this agent allowed to buy, up to what limit, from whom — and can you prove what it actually did?
Logs will not answer it when it counts
The comfortable answer is "we have logs."
But the moment that matters is not the happy path. It is the dispute. A duplicate order. A payment to the wrong vendor. A refund that should not have gone out. A charge a counterparty challenges.
In that moment, your own system's log of your own agent's behavior is precisely what an interested party discounts. It is your word about your own software, and you could have changed it.
What closes a dispute is evidence the other side can check.
Not "our records show," but:
This agent was authorized to spend this much, here, by this person — and here is the record of what it did, that you can verify yourself.
Decide what it can spend
Agentic commerce does not remove the need for financial control. It relocates it.
The control can no longer sit with a human approving each transaction. There are too many transactions, moving too fast. The control has to move into the agent's authority itself: a boundary set before it acts, and a provable record after.
That boundary needs to answer practical questions before money moves:
- is this agent allowed to buy this kind of thing?
- is this vendor in scope?
- is this amount inside the limit?
- who delegated that authority?
- what record will survive if the transaction is disputed later?
If those questions are answered after the invoice, the control already failed.
Prove what it spent
Fifteen trillion dollars is about to flow through actors that are not people.
The companies that come out ahead will not be the ones that trusted their agents the most. They will be the ones that could always prove what their agents did — to a vendor, a bank, an auditor, or anyone who did not take their word for it.
Decide what your AI can spend. Prove what it spent.