Take one purchase order. We have made it up, so nobody has to defend its numbers: a small distributor, a pallet of packing tape, a reorder that fires when the warehouse count falls below a set level. It is not a client's order and none of its figures is data. We will follow it from the count that triggers it to the moment money leaves, and we will follow it twice. In the first walk the agent holds a payment credential. In the second it only drafts.
Everything this article argues sits in the difference between those two walks.
The count drops
Software has raised this trigger for decades. A reorder rule notices stock below the line and raises a requisition, and nobody calls that agentic. What an agent adds is judgment around the trigger. It can see that the usual supplier now quotes a longer lead time, compare two quotes, and notice that a promotion next week will use the tape faster than the rule assumes.
That is useful, and it is also where the first problem appears. The reorder rule never needed to know who may commit the company's money, because a human always signed the result. Once an agent does the noticing, somebody has to have decided how much, for which categories, and who approves. If that approval matrix lives in one finance manager's memory, the agent will find out the first time it orders at the wrong price.
The draft, in the ERP
In the second walk the agent writes a purchase requisition inside the ERP: supplier, quantity, price, delivery date, and the reason it chose them. It lands in the same approval chain a person's requisition would, tiered by value. Below a threshold it goes to the warehouse lead, above it to finance. The agent has no payment credential and no way to release money.
Nothing in this step is exotic, and that is deliberate. The draft is a record with a history in a system that has permissions, so a reviewer can see what the agent proposed and an auditor can see who approved it. It is the shape of work that ERP buyers already know how to control, which is why the buyer's question is usually not about "agentic commerce" but something plainer: what an agent does to the purchasing process they already run.
The branch where the agent holds the card
Now rewind to the moment of purchase and give the agent a card. The payment industry has built real infrastructure for exactly this. Visa launched its Trusted Agent Protocol so merchants can tell a legitimate purchasing agent from a hostile bot. Mastercard has Agent Pay, a framework for merchants to accept agent-initiated payments. Google published its Agent Payments Protocol, AP2, as an open standard, and OpenAI and Stripe co-developed the Agentic Commerce Protocol behind Instant Checkout in ChatGPT.
These answer two questions well. Is this agent who it says it is, and did a human authorize the spend? They say nothing about a third question, the one your finance team asks: was this a good purchase?
Suppose the agent misreads a unit of measure and orders a hundred times the quantity, from a supplier you dropped last year. It is authenticated. It is authorized, within its spending cap. The rail reports a clean transaction, because by its own definition it is one. Only your own rules, held in your own system, could have stopped it, and in this walk they were never consulted.
Four problems around this branch have no clean answer yet. The first is liability: when an agent buys the wrong thing, spend caps and revocation exist, but dispute and refund processes for agent-initiated transactions are still being built, protocol by protocol. The second is manipulation. An agent can be steered by what it reads, and a vendor page or an email with hidden instructions becomes a payment attack the moment the agent can pay. See AI agent security and prompt injection. The third is shadow procurement: a team adopting an agent-driven buying tool without finance or IT knowing has built an unmonitored channel to a company card. Our shadow AI governance piece covers the policy side. The fourth is lock-in. Four large players have four different protocols, and committing early to one network's identity and dispute model could be expensive to reverse if the market consolidates elsewhere.
The branch where accounts payable pays
Back to the second walk. The requisition has been approved. The supplier ships, the warehouse records the receipt, and the invoice arrives. Accounts payable matches the invoice against the order and the receipt, which is the three-way match finance teams have run for years, and releases payment through the existing controls.
If the agent had misread the unit of measure here, the receipt would not agree with the order and the match would fail. The mistake would be caught by a step that already existed, run by people who were already accountable. The agent did the part that benefits from judgment and the money still moved through controls built over years.
That is our position, and it comes from reasoning and not from a client history: most of the risk in the first walk comes from one design choice, letting the agent hold payment credentials, and you can avoid it. For a low-value, repeating category such as consumables on a standing order, you can later let the agent approve inside a tight cap. Do that only after it has run in shadow mode, proposing orders that a person actually places, long enough to see where it disagrees with them. Widen from there one category at a time.
What the forecasts do and do not tell you
Juniper Research forecasts global agentic commerce transaction value rising from $8 billion in 2026 to $3.5 trillion in 2031, a factor of more than 400 in five years, or growth of about 3.4 times every year without a pause. It may come true. We cannot inspect the model behind it, and a forecast from a firm that sells reports on the topic is not independent evidence that the topic matters. Plan for the capability and ignore the curve, because a procurement agent that can pay is a decision you face whether the market reaches $3.5 trillion or $35 billion.
One more bet, held less firmly. The first agents most SMEs meet in commerce will probably sit on the other side of the table, representing marketplaces and suppliers, and not buying on the SME's behalf. Written approval rules are the defense in both cases. Write the matrix down, with values, categories and approvers. Route agent purchases through the same chain as human ones. Name an owner for what the agent may buy and from whom, and pilot on a low-risk category. If your team has tightened vendor approvals after an ERP implementation, you already know the shape of this work. If you are weighing where agents belong in procurement, talk to ThinqHub.
Sources: Juniper Research, Agentic Commerce Market 2026-2031 | Protocol coverage: Digital Commerce 360, Visa and Mastercard launch agentic AI payments tools



