Agentic Commerce

The buying decision, and everything after it

The consumer framing is the obvious one: an agent decides brand, store, price, alternatives, delivery, and buys your clothes. Pedro Franceschi, co-founder of Brex, argues the more interesting frontier is B2B, precisely because the decision-making is far more complex, and complexity is exactly what agents are good at absorbing.¹

A corporate purchase is not one decision, it is a graph of them. Pre-purchase questions include whether it is in policy, price comparison, terms, the security implications of onboarding a vendor, and compliance requirements. The personas involved include the requesting employee, their manager, a finance budget-holder, a procurement specialist, legal, IT and security, the line of business, and accounting. Franceschi's model gives each persona an agent and has the platform orchestrate the chain. The underrated half of the problem is what happens after the purchase: how it is booked, to which cost center, who approves it, who reviews it, how it reconciles. Even a single travel-card swipe hides a tail of receipt capture, memo, attendee list, approval, and reimbursement accounting. The end-user interface collapses to intent, "I need to buy this piece of software," and the orchestration happens on the back end.

The dogfood data point

Franceschi's own AI agent makes a purchase on a Brex card about once a week, versus his own online purchases every two to three days, putting roughly a third of his purchases under agent control. A representative flow: a friend forwards a message about movie seats, the agent finds the matching showing, notices those specific seats are taken, proposes adjacent ones, enters the card, pays, and forwards the confirmation email, untouched by a human at any step.² His tax prediction extends the same logic: "everybody's going to be audited because the cost is going to be zero."

Which rail wins: a live disagreement

Franceschi's picture assumes the card is the accountable rail agents transact on. Sam Broner of a16z crypto argues the opposite: cards are a structurally bad fit for agents, because the interchange model assumes a $20 to $1,000 human spending band and a roughly 30-cent fee floor, so it cannot serve machine-scale flows, and agents behave like businesses rather than shoppers, meaning most volume should route through pre-negotiated B2B settlement, largely ordinary wires, ACH, and batch transfers, with the net-new, programmable slice going to stablecoins rather than cards.³ Both readings can be partly right at once: a platform can own the orchestration and accountability layer while the underlying settlement migrates beneath it.

CZ supplies the clearest marker of how close that future is. His claim: "today AI can search for the best hotels and the best flights for us, but they don't book it for us." He expects booking to arrive "within months, definitely not years," and treats that crossing as the trigger for agent-native money, since a credit card at the point of purchase is workable but "very clunky."⁴

The consumer side, and the honest gap

Micky Malka of Ribbit Capital expects function-specific agents making commerce, payment, and investment decisions on a person's behalf to arrive "very, very soon," through blockchains plus AI acting as intelligent smart contracts on the user's behalf.⁵ He also concedes the gap in his own thesis: money is roughly four to five years behind knowledge in the adoption of AI, which sits in tension with how much building on the payments side is already underway. The likelier reconciliation is that consumer-facing decision-making lags while the underlying rails do not.