The index held flat this week, Organization at 72, Product at 68, Governance steady at 79. Flat is the wrong read. Underneath a quiet composite, the agentic commerce stack took its most concrete step yet from policy language to money on the line.
Start with the two signals that matter most together. American Express published its ACE Developer Kit specifying agent registration, verification, and tokenized payments, then paired it with Agent Purchase Protection that covers charges from registered agents (American Express). The structure is deliberate. An unregistered agent sits outside the protection boundary, and the merchant eats the chargeback. In the same window, Skyfire and Forter reported a Jomashop pilot where agent identity verification lifted checkout success to near 100% with no rise in fraud (The Paypers). That result retires the oldest objection to Know Your Agent controls, that identity verification adds friction. It removed friction.
Here is the argument. For eighteen months, agent registration has been a slide in a governance deck, a thing legal would get to. This week it acquired a dollar figure. The question in front of a principal is now quantifiable: how many agents transact against our rails, how many are registered, and what is the uncovered chargeback liability on the rest. That is a number a CFO can demand by Friday. The Identity Control Surface has a price now, and the price of not having one is a measurable loss column.
The complication is that the standard is splitting before it set. Amex and Skyfire are US-centric. PPRO has convened a Local Payments Council with BLIK, EPI, Bancontact, and DIAS to align European schemes with agentic protocols (Financial IT). Any firm operating across markets should assume multiple, possibly incompatible agent registration regimes. A centralized control surface built to one spec will need to speak several.
Meanwhile the architecture layer is confirming where this governance has to live. MIT Technology Review argues the durable advantage is the serving harness, not the model (MIT Technology Review), and OpenAI's Decisions API pushes bounded machine-speed decisions below the model layer (OpenAI). Both put the control surface in the plumbing teams were treating as neutral. Product at 68 reflects that gap: integrations are outrunning the governance meant to wrap them.
Do this: pull your agent inventory, map it against registered status on every payment rail you touch, and price the uncovered exposure. That is a board-ready artifact available today.
Watch item: whether a second EU member-state AI agency announces a bilateral safety arrangement on the OpenAI-NASK template within 60 days. If one does, a distributed model-layer inspection regime is forming ahead of formal EU AI Act enforcement.
WatchOpenAI and Poland's NASK partnership on AI safety and security is positioned as a bilateral research arrangement, but the underlying structure, shared access to safety tooling, coordinated red-teaming, government-adjacent technical collaboration, is the template that national AI agencies are adopting to gain visibility into frontier model behavior. Watch whether other EU member-state agencies announce equivalent bilateral arrangements in the next 60 days. If they do, the practical effect is a distributed regulatory inspection regime operating at the model layer, before formal EU AI Act enforcement reaches that level.