Mastercard spent decades building a fortress designed to keep bots out. Now it's learning to open the gates and invite them in to shop. That is the quiet upheaval at the heart of Greg Ulrich's presentation at VB Transform 2026, and it's worth pausing to appreciate how strange this moment really is. For more than twenty years, the company's risk rules treated the automated buyer as the enemy, the thing to block in under a hundred milliseconds across 175 billion transactions. Now, as Ulrich put it, the task is to enable the bot to transact, which means rewriting the very framework that made the network safe in the first place. If a company that built its entire trust infrastructure on saying no to machines is now engineering ways to say yes, then the rest of us are not just watching a trend; we are watching the operating system of commerce flip.
The practical consequence for anyone building or buying AI tools is that trust is no longer a feature of the transaction, it is the transaction. Ulrich's five layers, identity, verifiable intent, controls, execution, and intelligence, read like a checklist for every team trying to deploy agents in the real world. The identity layer alone is telling: only 32% of enterprise respondents in VentureBeat's June 2026 Pulse research give every agent its own scoped identity, and just 12% even consider agent-identity products. That gap is not a technical oversight; it is the difference between a bot that can buy a pair of Nikes and a bot that can explain why those Nikes were the wrong size on a final sale. Ulrich's point about verifiable intent, a tamper-proof record of what the agent was actually told to do, should land like a warning shot for every company rushing to ship an AI assistant without that paper trail. If you cannot prove what the agent intended, you cannot dispute what it did. The technology is not the hard part. The accountability is.
What makes this story more than a corporate update is the way Mastercard is treating the problem as an architectural one, not a policy one. Ulrich's confession that the company would build its internal agents differently today, after less than fourteen months, is the most honest thing said on that stage. He is not talking about tweaking a prompt or adding a guardrail after the fact. He is saying that the guardrails have to be embedded in the front end, that observability and accountability matter as much as the intelligence, and that trying to bolt on safety after the build is a path to failure. That is a direct challenge to the way most organizations are approaching AI right now, which is to treat it as a series of pilots that can be contained. The related lesson from our own coverage of AI clones and tax-season verification models is that the same principle holds at every scale: if you cannot audit what the system did, you cannot trust what it will do next. Mastercard is not just building an agentic factory for itself; it is building the blueprint for everyone else who will have to answer for an autonomous purchase gone wrong.
The specific thing to watch is the B2B procurement agent, not the consumer shopping bot. Ulrich pointed to a manufacturer with an always-on assembly line, an agent that tracks inventory, replenishes stock, and stays within budget, as the larger prize. That is where the five layers get stress-tested across companies, where a procurement agent, a supplier agent, and a banking agent all have to trust each other in an autonomous loop. Consumer purchases are a proof of concept. Business procurement is the pressure test. And the question that will define the next decade of commerce is not whether the models are smart enough to buy, but whether the infrastructure is honest enough to prove intent. Mastercard has spent 175 billion transactions learning that lesson. The rest of us are just starting to tap the card.