Written in response to a news report. Facts and law may have changed since the publication date.
Taiwan's Financial Supervisory Commission (FSC) opened a 30-day consultation on agentic AI in finance on October 13, 2026, CNA reported. The regulator is considering systems that can call tools and interact with financial institutions' operational systems. Its announcement expressly leaves existing law, supervision, outsourcing rules and internal controls in place. It is neither a new AI statute nor an approval for a particular pilot.
For a US software company proposing a Taiwan bank pilot, the commercial scope could be as limited as summarizing a loan file for an employee. Adding permission to change a customer record, send a message or initiate a payment would create a different proposed function. A successful demonstration of the first task does not establish authority to perform the others.
The bank already has control duties
Banking Act Article 45-1(1) requires banks to establish internal-control and audit systems. Where the arrangement involves outsourcing bank operations, paragraph 3 also requires internal systems and procedures covering the outsourced scope, protection of customer rights, risk management and internal-control principles. The regulator prescribes the detailed rules.
The statutory duties fall on the bank, whose review needs to address how the supplier's implementation fits its requirements and agreed scope. Article 45-1 does not itself prescribe a human confirmation for every automated action. The October 13 consultation offers no exemption from existing duties for a project described as experimental.
The consultation proposes evaluating each use case by the data, tools and actions it permits, with controls proportionate to risk. That approach makes a vague statement of work difficult to evaluate. “Assist with customer requests” could describe preparing a staff draft or sending a response directly. Those versions need different descriptions of when the system acts and who can approve a change in its scope.
Read access can still expose protected information
Article 48(2) requires a bank to keep customer deposit, loan, remittance and similar information confidential, subject to the exceptions in that article. These include rules in other laws and circumstances prescribed by the regulator. The provision is not an absolute ban on every authorized service arrangement, but the consultation adds no general exception for AI.
Suppose a pilot limits the agent to summarizing a loan file. If producing the summary sends the file to an external model service, the software's inability to change the account does not answer the confidentiality question. The bank still needs to identify what information the recipient obtains and the legal basis for that arrangement. An agreement calling the service “read-only” describes a system permission, not the legal basis for disclosing banking information.
A supplier's confidentiality promise can be relevant to the contractual protections, but the bank still needs a legal basis for disclosing the records. Nor does a closed testing environment answer who can view retained prompts, outputs or diagnostic records. The actual route and recipients of protected information remain relevant even if the customer never sees an automated message.
A later permission change can alter the pilot
In another hypothetical pilot, the initial service prepares a response for bank staff to review. A later update could let it send the response without that step. The parties would then be evaluating a new action, even if the model and customer interface look unchanged. This is an illustrative scope change, not a report about a named bank or a conclusion that automated replies are prohibited.
In practical terms, the contract description, system permissions and bank approvals need to describe the same operation. The consultation's suggested monitoring, shutdown and recovery mechanisms also give the parties concrete matters to discuss: who can stop the agent, which actions remain pending, and what records would explain an unexpected result. The proposal leaves the choice of controls tied to the use case and its risks. A deployment decision still requires assessment under the applicable rules.
Fictional scene
Sources
- CNA, October 13, 2026: consultation and the regulator's focus on operational-system permissions.
- FSC, October 13, 2026: 30-day consultation, proposed use-case approach and preservation of existing requirements.
- Banking Act Articles 45-1 and 48: bank internal controls, outsourced operations and customer-information confidentiality.
Sources checked: October 14, 2026. The cited Banking Act provisions are in force; the new document is a consultation. This is general Taiwan-law information, not approval of a system or an assessment of a particular bank's arrangements.
Frequently Asked Questions
- Did the FSC introduce a new binding agentic-AI law on October 13?
- No. It opened a 30-day consultation on governance, risk classification and controls. The FSC expressly says the document does not replace existing law, supervision, outsourcing rules or internal-control requirements.
- Does calling the deployment a pilot remove banking confidentiality duties?
- The consultation gives no such exemption. Banking Act Article 48(2) requires banks to keep customer deposit, loan, remittance and similar information confidential, subject to the statutory and regulatory exceptions. The actual data access needs assessment.
- Does permission to read an account also authorize the AI to initiate a transaction?
- Those are different operational permissions. A pilot's data access does not establish its permitted actions. The bank needs to assess the actual function under its applicable authority, procedures and controls; the consultation itself approves neither use.
This article provides general information and is not legal advice on any individual matter.
