Who tests the agent? Quality engineering in the age of autonomous payments
AI agents now make autonomous payment decisions in real time. The testing frameworks used to validate them were designed for software that behaves consistently. The gap between those two realities is where the next class of payment failures will come from.
In 2023, the question in every technology leadership meeting I attended was whether AI belonged in the payment workflow at all. In 2026, that question has been settled. AI agents are routing payments, scoring fraud risk in milliseconds, managing intraday liquidity positions, and initiating transactions without per-action human instruction. Raymond James launched a proprietary AI agent in January 2026. J.P. Morgan runs autonomous treasury operations. Deloitte's Agent Pay is in production pilots. The agents are here.
The question nobody in the regulated payments industry seems to be asking is the one I always ask first: who is testing them?
The difference between generative AI and agentic AI
Most discussions of AI in payments conflate two fundamentally different things. Generative AI — the models that summarise, recommend, and explain — can be evaluated against known outputs. You prompt it, you assess the response, you adjust. Testing it, while complex, follows recognisable patterns.
Agentic AI is categorically different. An agentic payment system does not wait to be prompted. It observes a payment event, applies a policy, and takes action — routing, blocking, escalating, initiating — based on learned behaviour that updates continuously from new data. The system that made a decision this morning may make a different decision on the same input tomorrow, because it has seen ten thousand transactions in between. That continuous adaptation is the source of its value. It is also the reason traditional testing frameworks cannot adequately validate it.
Why traditional QA frameworks fail for agentic payment systems
Traditional functional testing operates on a foundational assumption: that the same input will always produce the same output. For deterministic software, this assumption is sound. For an adaptive AI agent, it is not. A test suite that validates a fraud detection agent today may not reflect how that agent behaves a week from now.
This creates a specific problem that the regulated payments industry has not yet confronted at scale. The regulatory frameworks — Consumer Duty, DORA, PRA model risk management — require that firms can evidence the quality of their technology outputs. When those outputs are produced by a system that cannot be reliably reproduced from a fixed input, the evidence question becomes genuinely difficult. Most firms have not built the testing framework to answer it.
The four dimensions of agentic AI testing
Behavioural boundary testing. Rather than testing for specific outputs, agentic AI requires testing the boundaries within which the agent is permitted to operate — the rules it cannot violate regardless of what it learns. For a payment routing agent, these boundaries include payment amount limits, counterparty validation requirements, and regulatory sanctions list checks. Boundary testing verifies that the agent's adaptive behaviour never crosses those lines, even as the model evolves.
Adversarial testing. Agentic AI payment systems are a target. Adversaries probe AI fraud detection models to identify the patterns that trigger false negatives — transactions that the model approves despite being fraudulent. Adversarial testing deliberately constructs inputs designed to exploit the model's learned patterns. For firms deploying AI fraud detection in regulated environments, this is not an optional testing category. It is the most commercially critical one your test estate almost certainly does not yet have.
Model drift monitoring and testing. Agentic AI models drift over time as they learn from production data. What constitutes correct behaviour shifts. A testing framework for agentic AI needs to include continuous monitoring of model outputs against defined performance benchmarks — and a structured process for re-testing when drift exceeds defined thresholds. Without this, model degradation is invisible until it produces a production incident.
Explainability and auditability testing. When an AI agent makes a decision that a regulator or customer challenges — why was this payment blocked? why was this transaction flagged? — the firm must be able to produce an evidence trail. Explainability testing validates that the agent's decision logic can be traced, interpreted, and reported in language that satisfies a regulatory request. For Consumer Duty purposes, this is not optional.
What a mature approach looks like
The firms building mature agentic AI testing frameworks in payments are not treating this as a one-time validation exercise. They are building continuous testing into the production lifecycle — monitoring model behaviour, testing boundaries on a defined cadence, running adversarial scenarios alongside standard regression, and maintaining explainability as a first-class quality attribute alongside functional correctness.
This is what we call AI-Native Quality Engineering at RegalTech: not adding an AI tool to an existing test estate, but redesigning the quality function for the reality of adaptive, autonomous systems. The AI-Native QE Readiness Assessment gives a structured read on where your firm stands across all five maturity levels — and what the path to a defensible agentic AI testing posture actually looks like.
— Nipun Kumar, Founder & CEO, RegalTech Global Delivery Systems
Comments