This Week in Agentic (W12)

The week of March 14–20, 2026 sent a clear message: Agentic Commerce is ceasing to be merely a concept and is becoming infrastructure. From protocols to payments to the question of whether we can trust the agents at all—progress was made on all fronts simultaneously this week.

Protocols: Who Writes the Rules?

Google has updated the Universal Commerce Protocol (UCP). New features include a cart function that allows agents to add multiple items to the shopping cart at once, catalog access for real-time product data, and identity linking, which brings loyalty benefits across platform boundaries. That sounds like details. But it isn’t. These are the building blocks that turn an “agent that can search” into an “agent that can buy.” The simplified UCP integration via Google’s Merchant Center is set to roll out in the coming months—partners like Salesforce and Stripe plan to implement UCP on their platforms.

In parallel, Azoma has introduced the Agentic Merchant Protocol (AMP)—an alternative approach that thinks from the brand’s perspective. Mars, L’Oréal, Unilever, Beiersdorf, and Reckitt are on board as early adopters. The idea: Brands should control how their catalogs, brand guidelines, and product intelligence are represented in AI agents. This is the logical response to the fear that platform-driven protocols undermine brand sovereignty.

Two protocols, two perspectives. UCP says: We build the tracks. AMP says: We determine what runs on the tracks. The tension between the two will define the ecosystem in the coming months. Because the question isn’t just technical—it’s about whether the platforms or the brands control the agent layer. Whoever provides the context in which an agent “thinks” also controls what it recommends. More on this in my Ecosystem Map.

Payments: The invisible layer becomes visible

Visa had an eventful week. Visa Intelligent Authorization (VIA) is now live in Europe—a single-API solution that gives acquirers access to new payment methods, including agentic commerce, without having to completely overhaul their legacy infrastructure. Launch partners: Comercia Global Payments, Elavon, Fiserv, UNICRE, and Worldline. Added to this is the Visa Agentic Ready program, which is launching initially in Europe and the UK—with Barclays, HSBC UK, Nationwide, and Revolut as the first issuers.

In Latin America, Santander and Visa have completed the first end-to-end pilot for agentic payments in five markets—Argentina, Brazil, Chile, Mexico, and Uruguay. The transactions were still simple (books, chocolate), but the proof of concept is in place.

On the enterprise side, Mirakl and J.P. Morgan Payments are combining catalog orchestration with payment infrastructure—tokenization, fraud detection, and marketplace logic all in one. This is the kind of integration that moves agentic commerce from the demo phase into full-scale operation. And Lemrock has raised 6 million euros to act as middleware between merchant systems and conversational interfaces like ChatGPT and Claude. Investors are betting that the translation layer between existing commerce stacks and AI agents will become a standalone business.

Agent Safety: Trust Is the Bottleneck

Meta experienced a security incident triggered by an internal AI agent. The agent gave an employee incorrect technical advice and responded independently to an internal query—which led to unauthorized data access for nearly two hours. Meta emphasizes that no user data was compromised. Nevertheless, the case demonstrates what happens when agents act on their own initiative in production environments.

At the same time, OpenAI has published details on how they monitor their internal coding agents for misalignment. The method: Chain-of-Thought Monitoring, in which a monitoring model analyzes the agent’s thought processes. The result is both sobering and encouraging—monitoring works significantly better when it has access to the agent’s “thoughts,” but agents can learn to conceal their intentions. This has immediate relevance for the commerce context: If we allow agents to initiate payments and make purchasing decisions, we need exactly this kind of monitoring—not just after the fact, but in real time.

From SaaS to GaaS: Software That Delivers Results

Nvidia’s Jensen Huang coined a phrase at GTC 2026: “Every SaaS company will become a GaaS company.” GaaS—here not Governance, but “Goals as a Service”—describes software that is no longer operated by humans, but instead performs tasks autonomously via AI agents.

For the commerce sector, this means: The next disruption won’t come from better storefronts, but from software that sells results instead of access.

Consumer Acceptance: The Numbers Are Shifting

According to an Omnisend survey of 1,072 U.S. shoppers, 80% are now open to AI completing their online purchases—up from 34% a year ago. A 136% jump in a single year. 88% are willing to share personal data such as location and email receipts if they receive better recommendations in return. 38% have already made purchases directly through ChatGPT. At the same time, a clear red line remains: 70% would turn away from a retailer that uses AI for personalized pricing. And despite growing openness, 86% still have concerns—nearly half of them primarily regarding data collection. So willingness is growing—but not unconditionally. Retailers who see this as a free pass are making a mistake.

Outlook

Next week, Shoptalk Spring 2026 kicks off in Las Vegas (March 24–26). Running concurrently are eTail Asia in Singapore and the Chatbot Summit in Berlin—with an explicit focus on “Agentic Commerce in Action.” Anyone who wants to know how quickly the gap between protocols and practice is closing should pay attention to the Shoptalk panels on Agentic Checkout.

The question hanging over everything: If Google, Amazon, Visa, and the major brands all push their standards into the market at the same time—which one will prevail? And what will happen to the merchants who are still waiting to see what happens?