One of my favorite things about hosting Fraud Forward is that sometimes a conversation goes somewhere I didn’t plan for. I sat down with Matt Janiga from Modern Treasury expecting to spend an hour on APIs, payment rails, and the general evolution of fintech infrastructure. And don’t worry, we covered all of that. But I walked away thinking about how fraud teams investigate financial crime, and how much of that work is about to change.
Of course fraud tactics are always changing and we're all used to adapting to whatever criminals throw at us year to year, month to month, and even week to week. But now at the same time, the infrastructure behind financial services is evolving faster than ever before.
Matt made a comment that stuck with me: Dodd-Frank was written for fax machines, corded telephones, and human-staffed call centers. The regulations that shaped modern banking predate real-time payments, cloud-native cores, embedded finance, and AI agents. We're asking those frameworks to govern a world their authors could barely have imagined. That gap is fixable, and I think regulators are starting to see it too.
The Conversation Has Changed
For years, the regulatory question was some version of "how do we stop this?" Matt's observation, and I've felt it in my own conversations, is that it's turning into "how do we enable this safely?"
That's a different posture. Regulators still care about fraud. What's shifted is the acceptance that this technology arrives whether or not the rulebook is ready, so the useful question is how institutions adopt it responsibly.
As a fraud professional, I'll take that trade. Waiting until a technology matures before anyone talks governance has a perfect record of failure. I'd much rather be in the room while it's being built. (Part of that proactive utopia we all yearn for.)
The Modern Fraud Stack: The Tools Fraud Teams Never Had Before
Matt described early fintech as everyone "banging sticks together," which of course made me laugh because it's so accurate. A lot of fraud professionals entering the industry today have never had to build fraud operations from scratch. Today we have:
- Consortium intelligence
- Device intelligence
- Behavioral analytics
- Cloud-native fraud platforms
- AI-assisted investigations
- Real-time orchestration
Fifteen years ago? Most of that didn't exist. Teams spent their days manually stitching together data from disconnected systems while engineering raced to close the next gap before criminals found it.
We've come a long way. And the payoff runs past today's caseload. That same tooling is what has us ready for what comes next.
Agentic AI in Payments Change the Investigation
I've been thinking about agentic commerce a lot lately. The part that interests me is how it changes the shape of an investigation.
Every fraud investigation I've ever worked started from one assumption: a human initiated the transaction. Someone logged in and clicked send, but now, agentic commerce blurs that, because software may now be acting on a customer's behalf.
The customer remains accountable. Attribution is where the work gets harder. "Who clicked the button?" becomes a set of questions:
- Who authorized the agent?
- What authority was granted, and did the agent stay inside it?
- What information did it act on?
- Was the behavior consistent with what the customer actually intended?
Different investigative questions than the ones most of us trained on.
Identity Doesn't Go Away. Context Becomes Everything.
A worry I hear a lot: does AI make identity impossible? I think the opposite. The customer relationship still exists. So does the account owner, and so does authorization. What changes is how much context an investigator has to reconstruct.
Tomorrow's fraud investigations may require us to review:
- Agent permissions
- Decision histories
- Audit logs
- Behavioral changes
- Human approvals
- Authority boundaries
The transaction itself may tell you a fraction of the story.
We've Been Preparing for This Without Realizing It
One of my favorite parts of the conversation was Matt's take on consortium intelligence. Institutions have wanted visibility beyond their own four walls for years. Banks wanted to know when a neighboring FI was seeing the same pattern. Fintechs wanted the shared intelligence that already existed in pockets of banking. Nobody has enough visibility alone.
Agentic commerce follows the same logic. A single bank, fintech, or fraud team will only ever see a sliver of how these agents behave across the ecosystem. Shared intelligence becomes the baseline.
The Fundamentals Haven't Changed
Underneath all of it, we're still answering the question we've always answered: can we explain what happened? That holds whether the activity involves a person, an agent, an orchestration platform, or all three at once.
Which is why I'd skip waiting for perfect regulatory guidance.
- Start documenting authority.
- Start thinking about attribution.
- Start asking how your investigations would change if software, not a customer, initiated the activity.
By the time agentic commerce is ordinary, the institutions in the best shape will be the ones who started asking these questions early.
If you'd like to hear the full conversation with Matt Janiga from Modern Treasury, including our discussion on payment infrastructure, consortium intelligence, real-time payments, and where fraud operations are headed next, you can listen to the latest episode of Fraud Forward. It was one of those conversations that left me with more questions than answers...and honestly, those are usually the best ones.
FAQs
What are agentic AI payments?
Agentic AI payments describe transactions initiated by AI agents on a customer’s behalf, rather than the customer directly. Instead of a human clicking 'Send,' software evaluates conditions, applies delegated authority, and authorizes the payment. Identity still matters. Attribution expands.
How does attribution change when an AI agent initiates a payment?
Attribution goes from single-question to multi-layered. Instead of 'who clicked the button,' investigators reconstruct who authorized the agent, what authority was granted, what information the agent used, whether it stayed within permissions, and whether the behavior was consistent with the customer's intent. Accountability is not lost. It is layered across permissions, decision histories, and audit boundaries.
How do agentic AI payments prevent fraud?
They don't prevent fraud by default. They change where the fraud surfaces. Agent-initiated transactions can look cleaner than human ones because they arrive without the friction cues traditional fraud models rely on. Prevention now means reconstructing the agent's permissions, decision history, and intent alignment, not just scoring the transaction.
How does the modern fraud stack investigate agent-initiated transactions?
The modern fraud stack (consortium intelligence, device intelligence, behavioral analytics, cloud-native fraud platforms, AI-assisted investigations, and real-time orchestration) provides the reconstruction surface. Investigators pull six artifacts: agent permissions, decision histories, audit logs, behavioral changes, human approvals, and authority boundaries.
What is Model Context Protocol (MCP) and why does it matter to fraud teams?
What is Model Context Protocol (MCP) and why does it matter to fraud teams?
What questions should fraud teams start documenting today?
Start with five: who authorized the agent, what authority was granted, what information the agent used, whether it stayed inside its permissions, and whether behavior was consistent with the customer's intent. Then document six artifacts per case: agent permissions, decision histories, audit logs, behavioral changes, human approvals, and authority boundaries.
