Trust Is Becoming a Protocol

Software has been spending money for years. Advertising systems bid automatically, subscriptions renew without intervention and trading algorithms execute orders at machine speed. The novelty in agentic commerce is different: a general-purpose system may interpret a person’s intention, choose among merchants and exercise delegated purchasing authority.

That makes trust an infrastructure problem.

A merchant needs to know that an agent is legitimate. A payment network needs evidence that the user authorized the action. The consumer needs to know what was delegated, for how much and under which conditions. When something goes wrong, the parties need an audit trail capable of separating the model’s recommendation from the user’s mandate and the merchant’s acceptance.

The payment industry is increasingly treating those requirements as protocol rather than policy prose.

Payments were built for people, not shopping agents

Visa’s Intelligent Commerce Connect combines tokenization, authentication, spend controls and agent-compatible payment initiation. It is designed to work across several agent protocols rather than force merchants into a single ecosystem. Mastercard’s Agent Pay products similarly aim to make the agent an identifiable participant in the transaction.

OpenAI’s first implementation of Instant Checkout is intentionally conservative. Users explicitly confirm steps before action, while merchants continue to accept or reject orders and process payment through existing providers. The system is not attempting to make trust disappear; it is formalizing the handoff.

This is exactly what one should expect from financial infrastructure. A useful agent is not simply capable of action. Its authority must be bounded and legible to the other parties.

Autonomy should be granular

Current consumer research supports that model. Gartner’s May 2026 survey found strong reluctance to delegate final purchase decisions even where consumers were willing to delegate research and filtering. Deloitte’s July European study similarly found that AI agents are already assisting with search, comparison and evaluation while consumers still want transparency and final control.

The implication is that autonomy will probably expand by task rather than arrive as a binary switch. A consumer may permit an agent to compare prices freely, build a basket within a budget and reorder familiar household goods, while requiring confirmation for an expensive electronics purchase. A company may give a procurement agent broader authority inside an approved supplier list than outside it.

The economically important variable is not “how autonomous is the agent?” but “how cheaply can authority be delegated without making disputes, fraud and verification more expensive?”

Trust becomes portable

Protocols matter because bilateral trust does not scale. If every agent, merchant and payment provider defines consent differently, adoption becomes a custom integration problem. Agent identity, payment mandates, tokenized credentials and spend controls are attempts to make delegated authority portable across systems.

IBM’s guidance on production agents reaches a similar conclusion from the enterprise side: human oversight is not a failure mode. Systems should deliberately keep rigid controls around sensitive operations and use agentic reasoning where ambiguity produces value.

The best agentic commerce systems will therefore look less like unconstrained autonomy and more like well-designed institutions. They will make it easy to delegate the uninteresting decisions and difficult to exceed the mandate.

Trust is not the final UX layer. It is becoming part of the protocol stack.

The useful unit of autonomy is not the agent. It is the permission.

“Users stay in control — they explicitly confirm each step before any action is taken.” — OpenAI