Save
What Mastercard’s AI lead understands about enterprise transformation that most organizations are still missing.
Federico Cohen Freue fields roughly a thousand AI requests a year. That’s the incoming volume to Mastercard’s central AI and data team — proposals, ideas, and asks from across a global enterprise trying to figure out where and how to deploy the technology.
A few years ago, more than half of those requests were for chatbots. Today, more than half are for agents.
Fed takes this as a positive signal, not because agents are better, but because it means people’s mental models of what AI can do have matured. They’re no longer just imagining a question-and-answer interface. They’re imagining AI that takes action, embedded in a workflow, doing something on their behalf.
The problem, as he’s careful to note, is that wanting agents and being ready for agents are two very different things.
The Ball Bearing Problem
Robb introduced an analogy in this conversation that’s worth sitting with. Ball bearings: two that look identical. One that’s been perfectly machined, the right material, the right tolerance. One that hasn’t. You cannot tell the difference by looking at them. Put the bad one in an airplane engine, and the engine fails.
Agent demos work the same way. Even a sophisticated observer can’t distinguish a demo that represents a viable solution from one that’s a beautifully polished failure waiting to happen. The visual experience is identical. The underlying engineering is not.
This is part of why Fed’s team invests so heavily in training and fluency before deployment. It’s not enough to let demand drive the agenda. If people understand what conditions make AI work — what “machined correctly” actually means — they’ll make better requests, build better things, and catch the failures before they matter.
A Framework Simple Enough to Be True
One of the more practically interesting things Fed described is how Mastercard approaches prioritization at scale. With a thousand incoming ideas and a technology that keeps expanding its own capabilities, how do you decide what to move on?
The answer is a framework that fits in a sentence: use AI to make commerce more secure, smarter, more personal, and to make Mastercard stronger.
That’s it. The simplicity is the point. When teams across the organization have a shared language for where AI belongs (a common framework that holds even as models improve and use cases proliferate), prioritization becomes a conversation instead of a negotiation. It doesn’t answer every question, but it answers the most common one: does this fit?
What’s notable is what the framework is not doing. It’s not generating guardrails. It’s not a compliance checklist. It’s a strategic lens that gives people a way to think before they ask.
Trust at the Moment of Transaction
The conversation eventually turns to agentic payments, and here the stakes become concrete. AI agents executing financial transactions on behalf of users isn’t a speculative scenario anymore. The consumer demand exists. The product discovery patterns are already shifting toward LLM-mediated search. The loop is about to close.
Mastercard’s response to this is telling. The first priority isn’t building the exciting downstream applications: multi-vendor trip booking, smart replenishment, algorithmic negotiation between buyer and seller agents. The first priority is making sure the base case works. Agent identity verification. Delegated authority frameworks. Acceptance standards for merchants. The rules infrastructure that ensures when an agent executes a transaction, every party in the ecosystem can trust what happened.
Fed put it plainly: trust is the currency of innovation. And as transactions become more complex (more parties, more dynamic pricing, more autonomous decisions in the chain), the role of a trusted network doesn’t diminish. It compounds.
The middleman who was supposed to become irrelevant becomes the most critical node in the system.
Knowledge Before Action
In the second half of the episode, we walk through a demo of something we’ve been building: an AI-first approach to knowledge management and learning.
The premise is a critique of how enterprise AI usually works. You build a knowledge base, then you wait for someone to query it. You build an AI that knows things, then you ask it questions. The chatbot model. It puts the burden of knowing what to ask on the person who most needs to learn.
The alternative is a system that’s proactive rather than reactive, one that doesn’t wait to be asked, but figures out what you need to know and delivers it.
The architecture starts with a knowledge model: a structured, canonically correct source of truth for a given domain. Not a document repository, where the same idea exists in seventeen versions across seventeen files. A map, where each idea lives once, connects to what it’s related to, and carries a history of how it’s changed over time.
From that map, the system builds a learning twin: a representation of what a particular person knows and doesn’t know. Then it solves what Robb calls a traveling salesman problem: given where you are and where you need to go, what’s the most efficient route? Not a fixed curriculum built for an average learner. A dynamic path, recalculated at each step, based on what you’ve just learned and what’s changed in the domain since you last looked.
GPS for expertise. Here’s your position. Here’s your destination. Turn by turn, we’ll get you there, and if the road changes, we’ll reroute.
Fed’s response was immediate: this isn’t just a technology problem. It’s a cultural one. Asking people to engage with knowledge differently (to treat learning as a dynamic, ongoing process rather than a thing you did once during onboarding) requires a shift in how organizations think about readiness and what they reward. The technology can be ready before the culture is.
What Comes Before Doing
The thread running through this conversation is sequence. At Mastercard, there’s a deliberate ordering: understand first, then act. Build fluency before you deploy. Verify identity before you authorize a transaction. Know what you know before you build a curriculum.
This sounds obvious. It’s not how most AI initiatives actually work. Most AI initiatives start with the doing. An agent to automate this. A model to replace that. A chatbot to answer questions no one knows they have. Then they fail, and people call it a technology problem.
It’s usually a knowledge problem. The system didn’t know enough or the people managing it didn’t do the thing reliably. You don’t fix that by improving the model. You fix it by treating knowledge as infrastructure.
That’s the reframe this conversation is pushing toward: before you ask what AI can do, ask what your organization actually knows. Because the agents are only as good as the knowledge they’re built on.
Listen to the full conversation with Federico Cohen Freue on Invisible Machines.
