AI agents in financial services: control in production

chatbot that gives an incorrect answer creates a poor experience. An AI agent that blocks an account, changes a limit, or initiates the wrong transaction can cause financial losses, regulatory consequences, and an investigation that is difficult to reconstruct.

That difference changes the infrastructure problem.

AI agents do more than process an input and return an answer. They access systems, select tools, chain actions, and use the result of one step to decide the next. In financial services, this loop can involve personal data, credit engines, fraud detection systems, payment APIs, and Open Finance services.

Protecting the endpoint alone is therefore not enough. Institutions need to control what each agent can access, recommend, and execute, and stop a flow when its behavior exceeds expected boundaries.

The main risks emerge when agent autonomy grows faster than the controls around it.

---

Risk 1: the agent has system access but no identity of its own

Many AI agents enter production using credentials already assigned to existing services. The integration works, but the institution loses the ability to distinguish which agent performed each action.

If several agents share a service account, logs can show that a credential accessed an API or changed a record. They cannot necessarily show which agent initiated the action, which task it was performing, or which permissions it actually required at that moment.

The problem grows when credentials have broad scopes. An agent designed to check transaction status may be technically authorized to initiate payments. Another agent intended to support a credit assessment may be able to access customer data unrelated to the case under review.

Excessive permissions are already a risk in traditional systems. With agents, that risk gains scale and speed: an incorrect decision can be repeated across hundreds of sessions before anyone detects the pattern.

The institution’s architecture should treat each agent as an independent operational identity. This can involve dedicated credentials, least-privilege scopes, short-lived tokens, and different policies for reading, recommending, and executing.

Machine-to-machine authentication is only the beginning. A valid credential answers who is calling an API. Authorization policy must determine what that identity may do, against which resources, in what context, and within which limits.

Risk 2: a probabilistic decision produces a deterministic financial action

Models operate on probabilities. Payments, account blocks, customer data changes, and credit decisions produce concrete effects.

When an agent connects these two layers, an inaccurate response is no longer just incorrect content. It can become an action executed in a real system. The risk depends not only on model quality, but also on which tools are available and which controls exist between a decision and its execution.

A customer service agent may retrieve an invoice without additional approval. Changing a billing address may require additional identity validation. Increasing a limit or initiating a transfer may require reauthentication, human approval, or validation by another system.

This separation reduces the blast radius without removing all agent autonomy. Institutions can define controls according to the risk of each action:

• low-risk queries may proceed automatically;
• data changes may require additional validation;
• decisions with financial impact may be subject to limits on value, frequency, and recipient;
• critical actions may require human approval before execution.

The control point must exist outside the model’s instructions. A prompt telling an agent not to approve transactions above a certain amount is guidance, not a security boundary. The limit must be enforced by a mechanism the agent cannot ignore or modify.

WithAzion Functions, institutions can execute programmable logic in the request path before a request reaches the origin API. This capability can be used as part of an enforcement layer to evaluate requests and apply institution-defined rules before forwarding them to origin systems.

Risk 3: individually valid actions can form dangerous behavior

Not every incident begins with a malformed request or invalid credential.

An authenticated agent may make only permitted calls and still behave in a way that is inconsistent with its intended function. It may access too many customers in a short period, follow an unusual sequence of endpoints, repeat an action after inconclusive results, or continue a flow that should have stopped.

Each request may appear normal in isolation. The risk becomes visible in the sequence.

There is a difference between checking one transaction and observing an entire session. A balance inquiry is legitimate. Accessing thousands of different accounts with the same identity within minutes warrants investigation. One payment attempt may be expected. Repeating the operation after an ambiguous response can create duplicates or increase costs.

Generic IP-based rate limits do not solve the entire problem. An agent-control architecture needs to consider factors such as identity, action type, frequency, call sequence, and accessed resources.

Azion Bot Manager combines behavioral signals, fingerprinting, and network reputation to classify automated traffic. This classification provides an additional signal for security policies and mitigation of unwanted automation. Authorization for an AI agent to perform a particular action, however, should remain defined and enforced by the identity and authorization mechanisms adopted by the institution.

Risk 4: incoming traffic volume does not reveal the blast radius

In a conventional application, a request typically generates a predictable set of calls. With an AI agent, one task can initiate a loop involving queries, inference, tool calls, and retries.

A bank may receive only one hundred customer requests while the agent turns each one into dozens of interactions with internal APIs. If the logic is incorrect, downstream activity can grow significantly without the initial traffic appearing anomalous.

Capacity and risk should therefore not be measured only by incoming requests. Institutions also need to monitor how many actions each task generates, how long the flow remains active, which systems are called, and how often the agent retries.

The agent architecture can define circuit breakers and contextual limits to reduce this blast radius. For example, the system may stop a task when it exceeds the expected number of steps, repeats an operation excessively, or exceeds limits established by the institution.

Implementing these controls requires the architecture to maintain the context and state needed to evaluate accumulated task behavior. Controls in the request path can complement this design by applying rules to individual calls and protecting origin systems.

These measures also protect operations against failures that are not attacks. An endpoint returning an ambiguous response, an unavailable integration, or a misinterpreted instruction can keep an agent in a loop and consume origin capacity even when valid credentials are being used.

Risk 5: API logs do not form a decision trail

Traditional dashboards can show whether an API was available, how long it took to respond, and how many errors occurred. Those signals are necessary when auditing an AI agent, but they are not sufficient.

When a financial action is questioned, the institution needs to reconstruct the complete flow:

• which agent received the task;
• which identity and permissions it used;
• which data and systems it accessed;
• which tools it called and in what order;
• which action it recommended or executed;
• which policy authorized, blocked, or routed the action for approval;
• what happened in each system involved.

Without an identifier connecting these steps, investigations become fragmented across authentication logs, APIs, models, tools, and transaction systems. The institution knows that an operation occurred but cannot clearly reconstruct how the agent reached it.

Azion Real-Time Events provides visibility into events associated with traffic processed by the platform.Data Stream enables continuous delivery of events to observability and analytics destinations, including platforms such as Splunk, Azure Monitor, Datadog, and S3.

When the application propagates consistent context identifiers, this telemetry can contribute, together with records from the agent itself, identity mechanisms, and financial systems, to a correlated trail in the institution’s observability or SIEM environment.

Traffic infrastructure provides part of this visibility. Fully reconstructing a decision requires instrumentation across all components involved in the agent’s flow.

Organizing controls by level of autonomy

Not every financial AI agent requires the same architecture. Controls should reflect the authority granted to the system.

Autonomy level	Example	Controls the institution should consider
Query	Retrieve payment status or account information	Dedicated identity, read-only access, and per-resource limits
Recommendation	Recommend a credit decision or flag fraud	Records of data accessed, policy version, and human review when required
Modification	Update customer data, block a card, or change parameters	Additional identity and context validation, recovery mechanisms, and rollback
Transaction	Initiate a payment, transfer, or credit issuance	Value and frequency limits, segregation of duties, approval, and interruption mechanisms

This classification avoids two extremes. The first is applying controls so restrictive that the agent cannot operate. The second is granting production automation the same level of trust as a traditional internal service even when it dynamically decides which actions to execute.

The greater the autonomy, the more the architecture must account for the possibility that an agent can make a mistake, be induced to act outside expected boundaries, or use a valid permission in an inappropriate context.

A control layer before financial systems

Not every control required for AI agents belongs to the same architectural layer.

Identity, authorization, human approval, segregation of duties, and task state must be defined by the appropriate systems and processes within the institution. At the same time, controls applied in the request path can prevent inappropriate calls from reaching critical APIs and systems while providing security signals and telemetry for investigation.

At this traffic layer, the architecture can apply checks and protections to requests before forwarding them to the origin.

On the Azion platform,Functions enables programmable logic in the request path;Bot Manager uses behavioral signals, fingerprinting, and reputation to classify automated traffic;Web Application Firewall inspects and protects application-layer requests; andReal-Time Events andData Stream provide visibility and telemetry delivery capabilities.

These capabilities do not replace an institution’s identity, authorization, or transactional workflow mechanisms. They can form part of an enforcement, protection, and observability layer between agents and APIs, integrated with the other controls in the financial architecture.

The question to ask before putting an agent into production

Much of the discussion around AI agents in financial services focuses on whether the model is accurate, whether it may introduce bias, or whether it is susceptible to hallucinations. Those questions remain important. But they do not determine on their own whether a system is ready for production.

Before deployment, the institution needs to answer another question: what is the highest-impact action this agent can perform with the permissions it has today?

The answer reveals which controls are necessary.

If the agent only retrieves information, the primary risks involve unauthorized access and data exposure. If it recommends decisions, traceability and review become important. If it modifies systems or moves funds, transaction limits, segregation of duties, approval, and interruption mechanisms become part of the security design.

AI agents in production need to be treated as operational identities with bounded authority, not as intelligent interfaces connected to service credentials.

The difference between a compelling demonstration and a secure financial operation does not lie only in the model. It lies in an architecture capable of defining, enforcing, and demonstrating what each agent was permitted to do — and in infrastructure controls capable of protecting the systems that receive those actions.

---

Read also:

• Understanding agentic AI infrastructure
• Financial API security