Taiwan's Financial Supervisory Commission has opened a 30-day consultation on how financial institutions should govern, tier and secure agentic AI

Key takeaways

• This is a consultation. The FSC said it does not replace existing laws, supervisory rules, outsourcing management or internal controls.
• Chapter 3 would make the use-case scenario the unit of management, with review, monitoring, downgrade, disable and recovery mechanisms scaled to risk.
• Chapter 4 would tier scenarios on three axes and apply the higher tier where they differ.
• Chapter 5 would treat agents as non-human identities with basic and advanced controls under zero-trust principles.

What did the FSC publish, and is it binding?

Key dates and periods

On 13 October 2026 the FSC issued a press release announcing the document, with the document attached. It asks for views on governance, risk tiering and security controls when financial institutions adopt agentic AI. The FSC said the document draws on international security frameworks and rests on a risk-based approach, proportionality and technology neutrality.

It is not a rule. The FSC said the document is not meant to replace current laws, supervisory rules, outsourcing management or existing internal control requirements, and that it chose the consultation format because agentic AI techniques, use patterns and governance models are still developing. In the release's words: "本文件並非要取代現行法令、監理規範、委外管理及金融業既有內部控制要求。" The FSC said it hopes responses will serve as a reference for later policy work, guidance and supervisory communication. That is a statement of intent, not a commitment to issue guidance.

The release says agentic AI differs from chat-style AI that mainly generates replies because it can plan steps, call tools and reach data across systems. It names four risks if it is poorly managed: drift from the original goal, tool misuse or overreach, leakage of sensitive data, and errors that compound across multi-step processes. For related Taiwanese FSC material, see our articles on the FSC financial AI guidelines and the FSC frontier AI attack resilience measures, which are separate instruments.

How would governance by use-case scenario work?

Chapter 3 says it is advisable for institutions to manage agentic AI by application scenario. According to the release, the scenario would first pin down its business purpose, division of responsibility, accessible data, callable tools, executable actions, output or write destinations, and how abnormal events are handled, how the agent is disabled and how actions are traced. Controls for review, monitoring, inspection, downgrade, disabling and recovery would then be set according to risk level.

The FSC asks whether the governance principles are complete, whether there are ways to avoid excessive burden or strengthen accountability, and for other suggestions.

How would the three-axis risk tiering work?

Chapter 4 would rank each scenario on data sensitivity, action impact, and capability boundary or propagation (how widely it can spread). The release says the higher result governs, using the principle "採就高不就低原則". The tier would feed the full lifecycle: the approval threshold, the pilot or go-live baseline, the depth of controls, and later oversight and audit, with the tier reassessed over time. The FSC calls the tier a common language across those steps.

The FSC asks whether the three axes are appropriate and whether a more operable tiering method exists. The press release does not set out the tier levels or thresholds. We did not read the attached document.

What would the zero-trust controls in Chapter 5 cover?

Chapter 5 would treat an agent as a non-human identity and workload that can act. Following zero-trust principles of no default trust, continuous verification, least privilege and traceability, it sets out basic and advanced security requirements and control objectives across seven areas: identity, device or execution environment, network, application and workload, data, behaviour control, and visibility and analytics.

The FSC asks whether the design is enough to control risks specific to agentic AI, and whether existing security mechanisms can be linked in or an alternative reach the same goal. For a comparable approach in another jurisdiction, see our article on Singapore MAS agentic AI in finance, a separate instrument with its own scope.

What should Taiwan financial institutions and vendors do now?

This consultation creates no new mandatory requirements, and existing duties continue to apply. Our suggestion is to treat it as a prompt to test your agent inventory against the structure the FSC proposes, and to decide whether to comment.

• List each agent use case and fill in the Chapter 3 fields: purpose, owner, data, tools, permitted actions, write destinations, and shutdown and tracing method.
• Trial a tier assessment on the three axes for two or three live or planned agents, recording where the higher-tier rule changes the outcome.
• Editorial suggestion, not an FSC instruction: check whether each agent has its own identity and least-privilege access, rather than sharing a person's or a service account's credentials. The release says only that Chapter 5 treats the agent as a non-human identity under least-privilege principles.
• If you plan to file comments, follow the FSC's request: name the scenario and risk level, describe practical limits and alternatives, and avoid naming a specific model, product or vendor.

The FSC said the consultation lasts 30 days. The release states no start or end date. If the period were counted from the 13 October 2026 release date, which is our assumption, it would close around 12 November 2026; confirm the actual deadline on the public policy participation platform where the FSC says the full document is posted.

What we did not verify

What we opened: the FSC press release of 13 October 2026 on the FSC website (Chinese), on 24 October 2026. Every detail on this page comes from that release, including the chapter summaries.

What we did not open: the attached consultation document itself, the public policy participation platform page, the tier definitions, the list of basic and advanced controls, and any comment already filed. The 30-day start and end dates are our assumption and calculation, not dates stated by the FSC.

What we refuse to claim: that the document is binding, that any control or tier is mandatory, that the FSC will issue guidance or rules, or what the final text will contain.

Related briefings

• AI Regulation News hub
• Asia-Pacific AI regulation tracker
• AI regulation for CPAs and finance professionals
• Taiwan FSC financial AI guidelines
• Taiwan FSC frontier AI attack resilience measures
• Singapore MAS agentic AI in finance
• OSFI bulletin on generative and agentic AI cyber risk