The promise of agentic commerce is an assistant that goes out and buys things for you. Money20/20 this year suggested the reality is closer to an assistant that pays your parking ticket.

The demos on stage were genuinely impressive. What was more useful, across four days of panels, was the distance between what agents can now do and what anyone is willing to let them do.

Capability is not the question. On BNB Chain an agent booked a hotel from start to finish, a transaction that cost a third of a cent and was final in under a second. That has been possible for a while.

What has actually launched is far more careful, and the implementations share a pattern. A person approves every purchase. The agent never sees the underlying payment details. Where a merchant is not already set up for it, the agent is handed a single-use credential scoped to that one transaction and nothing else. Meta's Muse, which pays through Stripe's Link, works this way and is currently limited to US consumers.

The restraint is deliberate rather than timid. Teams building these systems describe spending months on the approval step alone, and one put the difference between enforcing those gates 95% of the time and 100% of the time as all the difference. Working out how often to ask without exhausting the user took months more.

All of which explains the category list. The common uses are retail, travel, and the chores nobody wants, like parking tickets and government payments. People are handing agents the admin rather than the spending, and not because an agent could not manage the spending. Approving every purchase is only tolerable when the purchase is boring.

The leash stays on because the question of who pays when something goes wrong has not been settled. One panel put the figure at 42% of embedded finance fraud losses having no clear accountability owner, which would mean nearly half of all losses landing somewhere nobody agreed in advance to stand.

What is missing is the infrastructure that would let an institution take that liability on: network rules, refund mechanics and compliance processes that make the risk something a balance sheet can hold. None of that exists yet. Business payments are moving more slowly still.

Which points at where the interesting companies are, and it is not the agents themselves. It is the layer beneath them: fraud detection that works at the speed software moves, verification an institution can rely on, and the accountability plumbing that lets someone sign off. Every demo at the conference was waiting on something from that list, which makes it the part of the stack with the most leverage and the least attention.

Agentic commerce will widen the day there are rules saying plainly who pays when an agent is wrong. Until then the ceiling is not model quality. It is how many approvals a person will tolerate in a week.