Agency Fraud: SHAP is No Longer Enough

Software agents are now responsible for a portion of financial fraud. The explainability methods associated with transaction fields are showing their limits in the face of this shift. In the United States, the NIST and a bill introduced by Senator Mark Warner are already targeting the identity and traceability of agents, marking the beginnings of a new audit framework.

Regulators and Legislators Target Identity and Traceability of Agents

In February 2026, the NIST launched a specific standards initiative for AI agents, considering them as identifiable non-human entities. The organization advocates for the establishment of certification cycles and specialized audit mechanisms, arguing that current authentication systems were developed for individuals and do not apply to autonomous software. Alongside this standards work, a bill presented in mid-2026 by Senator Mark Warner, the AI AGENT Act, would propose the creation of a federal registry of trusted agents and require them to operate transparently for the benefit of the user on major platforms. The parliamentary outcome of this proposal remains uncertain, but the trend is clear: the identity and accountability of agents are becoming central issues, while the sector has yet to define a shared standard for tracing an agent's reasoning in an auditable manner afterward.

The Human Signal Fades in Machine-to-Machine Flows

AI agents are now making purchases and transactions on behalf of users, to the point of being cited as a driver of fraud in a report mentioning "machine-to-machine chaos." In transaction logs, a legitimate purchasing agent and a fraudulent bot can appear almost identical, and no stabilized indicators reliably separate them. This disruption challenges detection systems built on the assumption of a human counterpart, leveraging behavioral footprints and deviations from a personal behavior baseline. The signals that fed these defenses seem to be evolving, with 60% of respondents anticipating a weakening of traditional protections in an AI-mediated banking context, while 84% is highlighted as a benchmark figure.

SHAP Has Established Explainable Decision-Making, but at the Transaction Level

The integration of SHAP above fraud classifiers has provided local justifications, transaction by transaction, where deep ensembles and boosted decision trees are difficult to interpret. In the absence of explanations, compliance teams can either approve by default or overlook alerts. Explainability approaches like SHAP allow for addressing the question "why does this transaction seem incorrect now," highlighting factors such as a high amount for this account, an account opened less than 48 hours ago, or a merchant already flagged. This type of justification is deemed more reliable than a simple risk percentage and remains suitable for fraud stemming from human behaviors.

Agents Do Not Follow Human Rhythms, Breaking the Basic Assumption

Transaction-centered models exploit descriptors such as amount, time, device, and velocity, without considering the identity of the sender. These variables correlate with human anomalies, for example, when a student suddenly initiates a transfer typical of an executive. Agents, on the other hand, have neither routine nor fatigue at a given time and can execute operations at a pace and consistency that a human cannot sustain even for five minutes. This asymmetry directly challenges the implicit assumption that deviations from human behavior are a reliable signal.

From Point Verdict to Trajectory Audit: The Need for a New Trace

In the face of fraud conducted by agents, the crucial question becomes "why did this actor act this way." Answering this requires inspecting the decision path: what tools were invoked, what permissions were in effect, and whether the action fell within the delegated framework or deviated several steps. An explainability engine designed to justify a point verdict on an amount does not indicate whether the agent deviated upstream of the transaction. The need resembles more of a trajectory audit than the analysis of a point. No widely recognized solution currently fully covers this scope. Some avenues consider assembling agent monitoring techniques and trajectory evaluation with existing explainability, or even building a new approach. SHAP has already helped defend systems before panels of experts, but the target has shifted to a space that characteristic-based explainability alone no longer covers.

What a Model Case Shows: An Explainable Classifier but Myopic to Agent Context

A fraud detection system based on a Random Forest trained on PaySim, compared to XGBoost and LightGBM, illustrates this gap. The model relied on transaction attributes and ignored the identity of the initiator. The addition of SHAP provided ranked explanations for each decision, offering a solid discussion basis before a panel tasked with examining weaknesses. This type of architecture remains effective for clarifying local decisions, yet it fails to capture the procedural context of a software agent.