Agentic AI has moved fast in customer support, software engineering, back-office document processing. Core banking is a different story. The same autonomous, multi-step, tool-calling agent architecture transforming other industries is running headlong into the regulatory apparatus that governs how financial institutions get to make decisions about customer money — and the collision is producing a genuine architectural problem, not just a governance inconvenience.
This post covers why agentic AI’s core strength — autonomous multi-step decision-making — is exactly what banking regulation is built to constrain, where agentic workflows are actually succeeding in banking today, and the control architecture emerging to reconcile the two.
Why Banking Regulation and Agentic Autonomy Are Structurally in Tension
Explainability requirements were written for deterministic decision systems. Fair lending regulations (ECOA, Regulation B), model risk management guidance (SR 11-7), and similar frameworks require a financial institution to explain, on demand, exactly why a specific credit or account decision was made — in terms both a regulator and a consumer can understand. An agent that dynamically decides which tools to call, in what order, off a probabilistic reasoning trace, produces a fundamentally different kind of decision record than a rules engine, or even a traditional ML model with a fixed feature set and a SHAP-value explanation.
Autonomous multi-step action is exactly what heightened control frameworks exist to prevent. Segregation of duties, maker-checker controls, transaction authorization limits — all of it exists because a single actor, human or system, taking an unsupervised sequence of actions that moves money is the core risk pattern banking controls are built to interrupt. An agent that plans, executes a funds-transfer tool call, verifies the result, and adjusts its next move based on that result is, architecturally, exactly the pattern regulators have spent decades building controls against — just wearing a new kind of actor.
Figure 1: The unconstrained agentic loop — plan, act, observe, re-plan — collides directly with maker-checker and segregation-of-duties controls when any step in the chain can move money or alter an account.
Where Agentic AI Is Actually Winning in Banking Today
Advisory and research workflows, not transactional ones. Agents that gather information, synthesize research across internal and external data sources, and hand a recommendation to a human underwriter or relationship manager are seeing genuine production deployment. The autonomy stays bounded to information gathering and synthesis; the decision and the action stay human.
Internal operations and reconciliation, where the “customer” is another system, not an account holder. Agentic workflows reconciling transaction breaks, chasing settlement discrepancies, gathering internal audit evidence are lower-risk because they touch internal control data, not customer-facing money movement — and errors get caught by existing reconciliation controls before they ever reach a customer.
Fraud and AML alert triage, with a human decision gate preserved. Agents that investigate an alert — pulling transaction history, cross-referencing sanctions lists, summarizing a case file — and hand a fully assembled recommendation to a human investigator are seeing strong adoption, because the final disposition still sits with a licensed, accountable human.
The Control Architecture Emerging to Bridge the Gap
- Bounded action spaces: Restrict the specific tools an agent can call in a given workflow to a pre-approved, auditable set. An agent doing loan research can query data sources but can’t call anything that alters an account or initiates a transaction.
- Mandatory human-in-the-loop checkpoints: Insert a hard stop before any action that crosses a regulatory threshold — funds movement above a limit, anything touching a credit decision — no matter how confident the agent is in its own reasoning.
- Full decision trace logging: Capture not just the agent’s final output but the complete reasoning chain, tool calls, and intermediate observations, in a format that satisfies model risk management documentation requirements — treat the agent’s trace the way SR 11-7 treats a traditional model’s validation evidence.
- Deterministic guardrail layers wrapping probabilistic agents: A rules-based control layer validates every proposed agent action against hard regulatory and risk limits before execution, regardless of what the underlying model reasoned. The control logic itself stays deterministic and auditable even when the agent’s reasoning isn’t.
Figure 2: The emerging banking agent control architecture — a deterministic guardrail and human checkpoint layer wraps every probabilistic agent action before it reaches execution.
The Investment Signal
- AI governance and control-layer platforms purpose-built for regulated industries are an emerging category distinct from general-purpose agent orchestration frameworks — and banking is likely to be the proving ground, given the stakes.
- Core banking vendors (Fiserv, FIS, Temenos, Thought Machine) that build agentic control architecture directly into their platforms have a durable edge over point-solution agent vendors trying to bolt compliance onto a generic framework after the fact.
- Model risk management tooling is being forced to evolve from its statistical-model validation roots into something that can validate and continuously monitor probabilistic, multi-step agentic reasoning — a genuinely unsolved tooling gap today.
Agentic AI’s autonomy is its selling point everywhere else in the enterprise. In core banking, that same autonomy is exactly the liability the regulatory apparatus exists to catch. The institutions winning this transition aren’t the ones deploying agents fastest — they’re the ones building control architecture that lets agents operate inside boundaries regulators already understand.
This is Part 5 of an advanced series on AI infrastructure economics. Follow @VamsiTalksTech for updates.
Discover more at Industry Talks Tech: your one-stop shop for upskilling in different industry segments!
Ready to master the future of telecom? My book, “Cloud Native 5G – A Modern Architecture Guide: From Concept to Cloud: Transforming Telecom Infrastructure (Industry Talks Tech)” is now available on Amazon.
Featured image designed by Frepik