
For three decades, fraud detection within the Financial Services industry has rested on a reasonably stable premise: a human being initiates a transaction, and that human being’s behaviour can be observed, modelled, and, when necessary, challenged. Humans are predictable creatures, and non-human behavioural signals are highly indicative of fraud risk. Predictive AI built on this premise has become the industry’s proven backbone: it has enabled real-time transaction monitoring and pattern recognition at scale, with machine learning models processing thousands of transactions per second to flag anomalies before they become losses.
Today we find ourselves on the brink of an unprecedented time: non-human entities are becoming economic actors with payment authority. The world-view and technological foundation that we have relied on for 30 years is now being asked to do something it was never designed to do: make sense of a world where the party behind a transaction may not be a person at all.
This is happening now, in pilot form, across retail, travel, and connected-device ecosystems. This is a fundamental shift, not just in the technology that supports the payments ecosystem, but a shift in the entire consumer shopping experience. Whilst this shift may start with small-scale transactions, the market signal is substantial: agentic commerce is projected to reach $1.7 trillion in transaction value by 2030i, a threshold at which more agents may be transacting than humans.
For fraud managers, this raises a question that goes well beyond incremental risk management. The industry’s central identity question has always been “who is this customer?” Increasingly, it will be “which agent is acting, on whose authority, and with what authorized intent?” Answering that question is going to require fraud prevention itself to evolve, leveraging agentic capabilities to match an agentic payment and technology landscape — a faster version of what banks already do will not suffice; a structurally different discipline is required.
A fundamentally different transaction flow
The shift from human-initiated to agent-initiated payments changes the anatomy of a transaction. A human outlines their intent and provides a preauthorization based on that intent, and sets permissible boundaries once; then, an agent identifies the need, initiates the payment, and only confirms with the human after the bank has already authenticated the agent and processed the transaction. The sequence is at once more complex and, from the bank’s vantage point, more compressed: fewer behavioural signals, no biometric variation, and the human effectively removed from the real-time loop.
This is the crux of the problem for fraud managers. Bot-attack detection today relies partly on a simple heuristic: a human cannot physically initiate hundreds of payments within seconds, so high-velocity transaction bursts are treated as suspicious by default. Agents erase that heuristic. If an authorized agent can legitimately transact at machine speed on a consumer’s behalf, velocity alone stops being a reliable fraud signal. The industry has not yet defined what does, in fact, distinguish a legitimate high-speed agent from a malicious one impersonating it.
Why established fraud prevention technologies are reaching their limits
The authentication toolkit that banks have spent years refining is built almost entirely around human signals: keystroke dynamics, device fingerprinting, behavioural biometrics, login patterns. Much of that toolkit becomes irrelevant the moment an agent engages a bank purely through an API rather than a user interface. There is no typing cadence to analyze, no mouse movement, no session behavior in the conventional sense — only a request, credentials, and a claimed intent.
Compounding this, agents still require human intent, authority, and permissions behind every action, but “intent” as a fraud signal doesn’t yet have an agreed data representation. Fraud teams are left asking how to measure deviation from intent when the very concept of intent has to be inferred from an agent’s configuration and its owner’s original instructions, rather than observed directly in behaviour. And unlike previous waves of payment innovation, there is no historical data to establish what “normal” agent behavior looks like versus anomalous behavior. Fraud models have always depended on a baseline, and for agentic commerce that baseline doesn’t yet exist.
At the same time, generative AI is being weaponized on the attacker’s side, enabling social engineering at a scale and level of sophistication that older training and detection methods were not built to counter. Industry data shows 95% of enterprise fraud managers already cite siloed data across systems as a top concern, even before agentic volumes and agentic payment schemas hit their books. Consumers reported $7.9 billion lost to investment scams in the United States in 2025 aloneii. Card fraud losses are projected to reach $404 billion over the next decade. And most pointedly for this discussion, potential fraud losses in the US driven specifically by fraudsters’ use of generative AI, are projected to reach $40 billion by 2027iii.
Whilst examples illustrate how Generative AI can be used to engineer sophisticated and highly tailored attacks, attack vectors at the other end of the spectrum demonstrate lower levels of sophistication, but significantly higher scale and widespread impact. A prime example of this is the emergence of “WormGPT” in 2023: an AI tool built specifically for malicious use and sold on dark web forums, explicitly stripped of the ethical guardrails found in commercial large language models. Security researchers reported that WormGPT enabled fraud operations, particularly adaptive phishing campaigns, to scale roughly ten times faster than traditional methods. This is the clearest evidence available that generative AI does not just change how individual fraud attempts are executed — it changes the throughput of fraud operations as a whole.
Speed, scale, complexity, attribution: the stakes are rising
Four characteristics distinguish agentic fraud risk from anything fraud teams have managed before.
The first is speed. Agent-to-agent transactions can occur in milliseconds, and detection windows that once spanned hours are compressing toward microseconds. A fraud team’s traditional review cycle — flag, investigate, decide, act — simply cannot execute within that window using conventional processes.
The second is scale. A single compromised agent could execute thousands of fraudulent transactions before any human becomes aware something is wrong, because agent-to-agent interactions can propagate a compromise far more rapidly than a single bad actor working through human-paced channels ever could.
The third is complexity. Multi-hop agent chains (where one agent enlists or instructs another, which enlists another) create transaction pathways that are difficult to audit and, in a fraud event, nearly impossible to unwind. If a compromise occurs three or four hops into a chain, tracing the reasoning and authority back to its origin becomes a forensic challenge unlike anything in traditional payments investigation.
The fourth, and in some ways the most consequential for the industry’s near-term planning, is attribution. When an agent commits fraud, liability is genuinely unresolved: is it the agent’s owner, the developer who built the agent, or the bank that authenticated and processed the transaction? Legal and regulatory frameworks have not caught up to this question, and until they do, banks are exposed to ambiguity that traditional fraud liability frameworks were never built to resolve.
Regulation is moving, but human-centric
Regulators are beginning to engage with this landscape, but today’s established frameworks were largely conceived with human actors in mind and are only now being extended, or reinterpreted, for non-human ones.
- The EU AI Act, which took effect in 2024, classifies AI systems by risk tier and imposes transparency requirements on high-risk applications. This structure will likely apply to many agentic payment systems once regulators formally categorize them.
- PSD3 specifically addresses requirements for Strong Customer Authentication (SCA) and caters for a vast array of transaction and payment types, but central and fundamental to the current framework is the assumption that a human has initiated the purchase. However, there is still opportunity for PSD3 to evolve and incorporate requirements for Agentic Commerce, as full implementation of the mandate is not expected to commence until 2027 at the earliest.
- Existing data privacy regimes, GDPR and CCPA, carry real implications for how AI agents collect, share, and act on consumer data, even though neither was drafted with autonomous agents in mind.
- The Basel Committee has begun articulating operational resilience requirements specifically for AI systems within financial institutions.
Practically, this leaves banks with a clear set of near-term obligations regardless of how quickly formal rules solidify: updating terms of service to establish agent authorization frameworks, defining explicit agent authentication policies, documenting AI decision-making processes in a form regulators can review, and proactively engaging with regulatory bodies that are actively shaping these emerging standards.
Fraud prevention becomes an agent-based system
The most important strategic implication of all this is that fraud prevention itself needs to adopt the same architecture and leverage the same capabilities as the threat it is defending against.
A workable model preserves the predictive AI core that has anchored fraud detection for decades: pattern recognition, anomaly detection, and real-time scoring remain foundational and are not being replaced. On top of that core, generative AI capabilities can be added specifically to increase the speed at which fraud teams react to new attack vectors, analyzing emerging trends and helping deploy countermeasures faster than manual analysis allows.
Underneath that model, a specific set of capabilities is becoming essential for any bank serious about operating in this space. Banks will need to:
- Profile new entities at scale, understanding agent identities and attributes, scoring agent-provider reputation, mapping agent authority and permissions, verifying intent, and enforcing transacting limitations.
- Use authentication controls that adapt dynamically rather than statically, including continuous authentication throughout a session, risk-based escalation, real-time control adjustments, and API-based authentication frameworks suited to non-interface actors.
- Connect agent and human identities explicitly, linking agents back to their human owners, correlating transaction types across an individual’s multiple agents, validating authorization chains, and tracking delegation and permissions.
- Adopt cross-agent pattern correlation: behavioral profiling of agents themselves, deviation detection operating at machine speed, multi-agent activity correlation, context-aware risk scoring, and network analysis of relationships between agents.
Why it’s important to act now
Industry analysis suggests banks acting on these capabilities today have a genuine three-to-five-year competitive advantage. Mass adoption of agentic payments is expected between roughly 2027 and 2030, at which point fraud patterns will emerge at real scale and regulatory frameworks will have solidified considerably. Banks that move early get to shape the standards, authentication norms, and case law that will govern this space; banks that wait are likely to face higher fraud losses, sharper regulatory scrutiny, and, eventually, customer trust erosion once agentic fraud incidents become public and frequent.
The transaction volume and the fraud tooling on the attacker’s side both indicate this agentic, AI-powered fraud is already underway. This is why banks need to build the identity, authentication, and cross-agent correlation capabilities before the regulatory floor rises to make catching up considerably more expensive. For fraud managers specifically, the shift from monitoring people to monitoring networks of autonomous agents may be the most consequential change to their discipline since the introduction of real-time transaction scoring itself.



