AI agent observability in financial services: what evidence do your risk, compliance and audit teams actually ask you to show?
AI agent observability in financial services is the practice of capturing what AI agents do, with which data and under whose authority, in a form that risk, compliance and internal audit teams can use as evidence.
Financial supervisors, including in Southeast Asia, are moving from general principles on AI toward concrete expectations on governance, inventories, controls and accountability. An AI observability platform matters here not because it satisfies a regulation by itself, but because those expectations can only be shown to be met with records of what actually happened.
Why financial supervisors are focusing on AI agents
Two recent signals show the direction. On 7 October 2026 the Monetary Authority of Singapore issued Guidelines on Artificial Intelligence Risk Management. They apply to financial institutions and to all forms of AI technology, and they address AI agents explicitly.
They take effect on 7 October 2027, with governance expectations applying from that date and the remaining sections, which include lifecycle controls and supporting capabilities, phasing in by 7 October 2028. The authority also said it plans to consult the industry in 2027 on further guidance for agentic AI.
In Malaysia, the Governor of Bank Negara Malaysia said in an opening address on 8 July 2026 that more than 70% of Malaysian financial service providers have implemented at least one AI application. He argued that AI belongs on the board agenda, that accountability cannot be handed to an algorithm, and he asked risk, compliance and internal audit teams whether institutions can explain outcomes and challenge decisions, not just confirm that controls exist.
These are statements of supervisory expectation, not a single rulebook. Requirements differ by jurisdiction and by institution, and interpreting them is a matter for the institution's own compliance and legal teams.
From requirement to evidence
The expectations in this area cluster into five themes. Each one can be restated as a question about telemetry.
AI inventory: which agents actually exist?
Supervisory expectations call for institutions to identify their AI use and keep an inventory at a sensible level of detail. A register maintained by hand tends to drift from reality. Telemetry gives a second view: the agents, models and tools that are actually running in production, including ones that never made it into the register. Comparing the register with what is observed is a practical check on the inventory.
Risk materiality: how risky is each use case?
Institutions are expected to assess how material the risk of each AI use case is. That assessment is stronger when it rests on facts about behavior: which tools an agent can call, which data it touches, how often it acts and with what outcomes. Observability supplies those facts.
AI lifecycle controls: evidence for testing, oversight and change
The guidelines name controls across the AI lifecycle, including data governance, testing, human oversight, cybersecurity, monitoring and change management. Observability supplies evidence for several of them: evaluation results over time, records of human approvals and overrides, changes to versions and configurations, and quality signals that show when behavior shifts.
Third-party AI: what changed when the provider changed?
Institutions remain accountable for third-party AI used in their services, and are expected to obtain sufficient assurance and assess whether it suits the intended use. A record of which model version served which task, and of how behavior changed after a provider update, makes that assessment concrete instead of contractual only.
Accountability and explainability: can a reviewer challenge the outcome?
The strongest test is whether a reviewer can reconstruct a decision: the input, the retrieved context, the tool calls, the output and any human intervention. End-to-end traces turn an opaque outcome into a path that risk, compliance and audit teams can examine and challenge.
What audit-grade evidence looks like
Volume of data is not the same as evidence. Records need to be complete, covering the whole path of a task and not only the final answer. They need integrity, with clear provenance and protection against alteration, so that an audit trail can be trusted.
Retention has to match the institution's policy, access has to be controlled, and sensitive or personal data in prompts and outputs has to be redacted. Where data residency applies, the deployment model becomes part of the evidence story, because telemetry that leaves the jurisdiction may create a new problem.
What observability cannot do
Observability does not set risk appetite, assign accountability or replace governance. It does not make an institution compliant. What it provides is the evidence layer: the record that lets governance, risk and audit teams verify what the AI agents did. Treating it as more than that leads to false confidence.
Evaluating an AI observability platform in a regulated setting
For regulated institutions, the evaluation questions shift.
- Can the platform show the full path of an agent task?
- Does it support retention, redaction and access control that match policy?
- Does the deployment model, such as private or hybrid, fit data residency needs?
- Does it support open standards and connect with existing audit and governance workflows?
Bonree ONE, the observability platform from the author's employer, collects traces and usage data from AI applications built on frameworks such as LangChain, LangGraph, Dify and OpenClaw, which is one possible source of this kind of evidence. Whether any set of records meets a specific requirement is for the institution to assess.
FAQ
What is AI agent observability in financial services?
It is the capture and analysis of AI agent behavior, including tool calls, data access, outcomes and human interventions, in a form that risk, compliance and audit teams can use as evidence for AI governance.
Does AI observability make a bank compliant with AI regulation?
No. It provides evidence that supports governance and audit, but compliance depends on the institution's own controls, policies and interpretation of the rules.
What is an AI agent inventory, and why does observability help?
It is a register of the AI agents, models and tools an institution uses. Observability helps by showing what is actually running in production, so the register can be checked against reality.
What evidence do auditors need from AI agents?
Complete task traces, records of human approvals and overrides, version and configuration changes, evaluation results over time, and records of which model served which task, all with appropriate retention and access control.
Do these expectations apply outside Singapore?
The Singapore guidelines apply to institutions under that authority. Other supervisors take their own approaches, but the underlying questions about inventory, control, accountability and evidence are similar.
Disclosure: the author works at Bonree, the vendor of Bonree ONE, which is mentioned once above. Beyond that, this article names no products and ranks none. It is not legal or regulatory advice.
Comments
No comments yet. Start the discussion.