Fourteen markets. Fourteen different routes to the same question: who owns the model, and who cleared it?

Oonagh van den Berg | RAW Compliance

A bank cannot put "APAC" in the jurisdiction field of its AI inventory. It can use the same model across the region, but it cannot assume the same regulatory route to production.

Singapore is consulting on an AI risk-management framework for financial institutions. South Korea has updated sector guidelines. Hong Kong has long-standing supervisory principles and an expanded sandbox. Indonesia has issued bank-specific governance guidance. Elsewhere supervisors are applying licensing, conduct, data, cyber and outsourcing duties to AI without pretending that a discussion paper is a law.

Issue 1 of this map examined Africa, where Kenya's draft bank-AI guidance proposed a written pre-deployment approval gate. APAC shows a different pattern: some supervisors define the control spine, some open testing lanes, and some make firms prove that existing obligations survived a new technology. None is a permission slip to deploy first and find an owner later.

Singapore - from FEAT principles to proposed lifecycle controls

MAS's FEAT principles put fairness, ethics, accountability and transparency on the table. Its November 2025 consultation proposed AI risk-management guidelines for all financial institutions, covering oversight, policies, lifecycle controls and capability. The consultation closed in January 2026. A consultation is not a final binding rule. But the direction is clear enough to test an inventory now: what is the use case, where is the model, who approves it, how is it monitored, and what does the firm do when it fails? A regional vendor that only offers a model card will need to show how that card feeds the bank's own control record.

Source: https://www.mas.gov.sg/publications/consultations/2025/consultation-paper-on-guidelines-on-artificial-intelligence-risk-management

Australia - the governance gap is the story

ASIC's review of 23 financial and credit licensees warned of a gap between AI adoption and governance. It is a supervisory finding, not a new AI-specific licensing regime. APRA's operational-risk standard CPS 230 adds live duties around critical operations and material service providers for regulated entities. Put the two together and the board question is practical: can the firm see where AI touches a customer outcome or important service, and can it manage the provider behind it? "Our supplier handles the model" does not answer who owns the consumer and resilience risk.

Sources: https://www.asic.gov.au/regulatory-resources/find-a-document/reports/rep-798-beware-the-gap-governance-arrangements-in-the-face-of-ai-innovation ; https://www.apra.gov.au/news-and-publications/apras-new-prudential-standard-operational-risk-management-comes-force

Hong Kong - principles plus a supervised test lane

HKMA's high-level AI principles for authorised institutions predate the current generative-AI rush. In March 2026 it expanded its GenAI Sandbox with other regulators. The sandbox is a place to test, not a general exemption from bank governance or customer protection. Banks should keep the experimental boundary explicit: which use case was tested, with what data, what humans can override, and what evidence would justify wider deployment? A successful demonstration is not itself a production approval.

Sources: https://brdr.hkma.gov.hk/eng/doc-ldg/docId/20191101-1-EN ; https://brdr.hkma.gov.hk/eng/doc-ldg/docId/20260305-2-EN

India - FREE-AI is a framework proposal, not a licence

RBI's August 2025 FREE-AI committee report sets out a responsible and ethical enablement framework for the financial sector. It is a committee report, not an enacted blanket AI code. Its value for a bank is the discipline of joining innovation to accountability and human oversight. That is where a procurement checklist often falls short: approval of the tool is not approval of every decision it will influence. The decision path needs named owners and a way to challenge an output.

Source: https://www.rbi.org.in/Scripts/PublicationReportDetails.aspx?ID=1306

Indonesia - bank-specific governance guidance

OJK's 2025 Artificial Intelligence Governance for Indonesian Banks was issued as guidance and a minimum benchmark for responsible bank AI. OJK positions it alongside existing digital-transformation, IT, cyber, digital-maturity and resilience requirements. That means the implementation work is not another slide deck called "AI ethics". It is the join between model design, data quality, prudential risk, cybersecurity, customer treatment and ongoing assurance. Guidance is not the same thing as a new pre-deployment approval rule.

Source: https://ojk.go.id/en/Publikasi/Roadmap-dan-Pedoman/Perbankan/Pages/Indonesia-Artificial-Intelligence-Governance-for-Banking.aspx

South Korea - sector guidelines, with a legal perimeter

The FSC introduced updated financial-sector AI guidelines in June 2026, effective from 22 June. It calls them self-regulatory and recommends risk-proportionate use by firms. Seven principles cover governance, legality, human-supervised assistance, credible data and models, financial stability, good faith and security. The FSC also says high-impact AI covered by Korea's AI Framework Act must meet the applicable law. This is the distinction to preserve: voluntary sector guidelines do not erase binding obligations. For agentic payments, ask who can authorise the agent and who remains responsible when its recommendation becomes a transaction.

Source: https://www.fsc.go.kr/eng/pr010101/87140

Japan - dialogue before prescriptive AI rules

Japan's FSA updated its financial-sector AI discussion paper in March 2026. It describes opportunities, misuse, misinformation and the risk of inaction, and invites continued dialogue. It is a discussion paper, not a detailed mandatory AI deployment code. The practical response is neither to wait for a perfect rule nor to call every experiment compliant. Keep a risk-ranked inventory, demonstrate controls under existing financial and data requirements, and record what needs supervisory discussion before scaling.

Source: https://www.fsa.go.jp/en/news/2025/20250304/aidp.html

Malaysia - the consultation stage matters

Bank Negara Malaysia's August 2025 discussion paper explores its proposed approach to AI in the financial sector and asks for feedback. Do not label that paper a final AI policy document. For financial institutions, it is a signal to connect AI use to customer outcomes and existing technology-risk controls while the policy develops. A board should know which deployments are customer-facing, which depend on third-party models and where human challenge is real rather than a box on a form.

Source: https://www.bnm.gov.my/-/dp-aifs25

New Zealand - technology-neutral does not mean responsibility-neutral

The FMA says it wants responsible AI adoption and expects good governance as firms manage market manipulation, systemic, consumer and ethical risks. It points to board and senior-management accountability and clear roles across the AI lifecycle, including human intervention. That is a supervisory position rather than a new dedicated bank-AI rule. For firms using AI in fraud, advice or customer service, the test is whether oversight can follow the decision from input to customer outcome.

Source: https://www.fma.govt.nz/library/opinion/ai-and-good-governance/

Mainland China - a binding bank and insurance AI opinion

Reported, not regulator-confirmed: On 18 June 2026 the National Financial Regulatory Administration issued Guiding Opinions on the Secure Development and Application of AI in the Banking and Insurance Sectors, effective on issue. They cover governance, data, risk and security. They require full life-cycle management and AI risk grading. Generative AI in scenarios such as fund transactions, credit approval, underwriting and claims is high-risk and needs risk management committee approval. Externally sourced generative AI models must be filed with the cyberspace authorities. Personal information such as names, ID and card numbers may not be used to train or tune generative models. Public-facing or high-risk use must be reported to the NFRA. This is the most prescriptive text in the region.

Sources: https://www.grandwaylaw.com/en/Legalupdate/4679.html ; https://policy.mofcom.gov.cn/claw/clawContent.shtml?id=106182

Taiwan - FSC guidelines for AI in the financial industry

Reported, not regulator-confirmed: The Financial Supervisory Commission released its Guidelines for AI Applications in the Financial Industry on 20 June 2024. They are guidelines for financial institutions, not a statute. I read the listing and the first page only, so the specific control requirements are not summarised here.

Source: https://law.fsc.gov.tw/EngLawContent.aspx?id=2690&lan=E

Thailand - Bank of Thailand AI risk management policy

Reported, not regulator-confirmed: The Bank of Thailand issued a Policy Framework on AI System Risk Management on 12 September 2025. It applies to financial institutions under the Financial Institution Act and payment service providers under the Payment System Act, and aims at reliable, unbiased, transparent and secure AI use. I relied on a law-firm summary of it and on the BOT document link.

Sources: https://www.lexology.com/library/detail.aspx?g=2944cbce-8539-4091-92a8-df74be1e83f0 ; https://www.bot.or.th/content/dam/bot/fipcs/documents/FOG/2568/ThaiPDF/25680178.pdf

Philippines - BSP governance principles for AI

In June 2026 the Bangko Sentral ng Pilipinas issued Memorandum M-2026-031 with a guidance paper, Governance Principles for AI in Financial Services, to all BSP-supervised institutions. It is a principles-based paper, proportionate to the nature and complexity of each institution’s AI use, built around what it calls the STARS of AI governance, and it maps controls to the AI system lifecycle. It sets minimum supervisory expectations and asks institutions to build their own AI governance framework.

Source: https://www.bsp.gov.ph/Regulations/Issuances/2026/M-2026-031.pdf

Vietnam - an AI law and a draft banking circular

Reported, not regulator-confirmed: Vietnam passed a Law on Artificial Intelligence, No. 134/2025/QH15, on 10 December 2025. The State Bank of Vietnam has also released a draft circular on AI in banking, reported in April 2026. It would require risk classification of AI systems, human oversight, impact assessments for high-risk AI, disclosure that a customer is dealing with AI, and incident reporting to the SBV within 24 hours. A draft is not a rule. I did not confirm whether the circular has been finalised or when the law takes effect.

Sources: https://english.luatvietnam.vn/law-no-134-2025-qh15-dated-december-10-2025-of-the-national-assembly-on-artificial-intell ; https://www.bakermckenzie.com/en/insight/publications/2026/04/vietnam-draft-circular-on-ai-deployment-in-banking

The operating model that travels

One region, no single checklist. Build a common control spine - use-case inventory, accountable owner, model and data provenance, independent challenge, human authority, testing for performance and bias, monitoring, third-party assurance, incident response and exit. Then add a jurisdiction layer: instrument status, scope, supervisory engagement, local data and customer duties, and whether the system can move from trial to production.

Do not flatten a consultation, a sandbox, a guideline and a binding prudential standard into the same green cell in a spreadsheet. The status is part of the risk. If a bank cannot say which version of the model made a decision, which local rule applied and who could stop it, it does not yet have a regional AI operating model. It has a shared demo.

Next issue: Middle East.

Source and status note

Positions checked against official regulator materials available by 25 September 2026. This is a market map, not legal advice; verify subsequent finalisation, amendments and institution-specific obligations before deployment. The country entries deliberately distinguish consultation, report, guidance, sandbox and live prudential duties. Mainland China, Taiwan, Thailand, the Philippines and Vietnam were added on 11 October 2026. Taiwan and Thailand rest on a listing and a law-firm summary, and Vietnam’s circular is a draft.

THE AI REGULATION MARKET MAP • APAC • RAW COMPLIANCE

Fourteen markets. Fourteen different routes to the same question: who owns the model, and who cleared it?

Oonagh van den Berg | RAW Compliance

A bank cannot put "APAC" in the jurisdiction field of its AI inventory. It can use the same model across the region, but it cannot assume the same regulatory route to production.

Singapore is consulting on an AI risk-management framework for financial institutions. South Korea has updated sector guidelines. Hong Kong has long-standing supervisory principles and an expanded sandbox. Indonesia has issued bank-specific governance guidance. Elsewhere supervisors are applying licensing, conduct, data, cyber and outsourcing duties to AI without pretending that a discussion paper is a law.

Issue 1 of this map examined Africa, where Kenya's draft bank-AI guidance proposed a written pre-deployment approval gate. APAC shows a different pattern: some supervisors define the control spine, some open testing lanes, and some make firms prove that existing obligations survived a new technology. None is a permission slip to deploy first and find an owner later.

Singapore - from FEAT principles to proposed lifecycle controls

MAS's FEAT principles put fairness, ethics, accountability and transparency on the table. Its November 2025 consultation proposed AI risk-management guidelines for all financial institutions, covering oversight, policies, lifecycle controls and capability. The consultation closed in January 2026. A consultation is not a final binding rule. But the direction is clear enough to test an inventory now: what is the use case, where is the model, who approves it, how is it monitored, and what does the firm do when it fails? A regional vendor that only offers a model card will need to show how that card feeds the bank's own control record.

Source: https://www.mas.gov.sg/publications/consultations/2025/consultation-paper-on-guidelines-on-artificial-intelligence-risk-management

Australia - the governance gap is the story

ASIC's review of 23 financial and credit licensees warned of a gap between AI adoption and governance. It is a supervisory finding, not a new AI-specific licensing regime. APRA's operational-risk standard CPS 230 adds live duties around critical operations and material service providers for regulated entities. Put the two together and the board question is practical: can the firm see where AI touches a customer outcome or important service, and can it manage the provider behind it? "Our supplier handles the model" does not answer who owns the consumer and resilience risk.

Sources: https://www.asic.gov.au/regulatory-resources/find-a-document/reports/rep-798-beware-the-gap-governance-arrangements-in-the-face-of-ai-innovation ; https://www.apra.gov.au/news-and-publications/apras-new-prudential-standard-operational-risk-management-comes-force

Hong Kong - principles plus a supervised test lane

HKMA's high-level AI principles for authorised institutions predate the current generative-AI rush. In March 2026 it expanded its GenAI Sandbox with other regulators. The sandbox is a place to test, not a general exemption from bank governance or customer protection. Banks should keep the experimental boundary explicit: which use case was tested, with what data, what humans can override, and what evidence would justify wider deployment? A successful demonstration is not itself a production approval.

Sources: https://brdr.hkma.gov.hk/eng/doc-ldg/docId/20191101-1-EN ; https://brdr.hkma.gov.hk/eng/doc-ldg/docId/20260305-2-EN

India - FREE-AI is a framework proposal, not a licence

RBI's August 2025 FREE-AI committee report sets out a responsible and ethical enablement framework for the financial sector. It is a committee report, not an enacted blanket AI code. Its value for a bank is the discipline of joining innovation to accountability and human oversight. That is where a procurement checklist often falls short: approval of the tool is not approval of every decision it will influence. The decision path needs named owners and a way to challenge an output.

Source: https://www.rbi.org.in/Scripts/PublicationReportDetails.aspx?ID=1306

Indonesia - bank-specific governance guidance

OJK's 2025 Artificial Intelligence Governance for Indonesian Banks was issued as guidance and a minimum benchmark for responsible bank AI. OJK positions it alongside existing digital-transformation, IT, cyber, digital-maturity and resilience requirements. That means the implementation work is not another slide deck called "AI ethics". It is the join between model design, data quality, prudential risk, cybersecurity, customer treatment and ongoing assurance. Guidance is not the same thing as a new pre-deployment approval rule.

Source: https://ojk.go.id/en/Publikasi/Roadmap-dan-Pedoman/Perbankan/Pages/Indonesia-Artificial-Intelligence-Governance-for-Banking.aspx

South Korea - sector guidelines, with a legal perimeter

The FSC introduced updated financial-sector AI guidelines in June 2026, effective from 22 June. It calls them self-regulatory and recommends risk-proportionate use by firms. Seven principles cover governance, legality, human-supervised assistance, credible data and models, financial stability, good faith and security. The FSC also says high-impact AI covered by Korea's AI Framework Act must meet the applicable law. This is the distinction to preserve: voluntary sector guidelines do not erase binding obligations. For agentic payments, ask who can authorise the agent and who remains responsible when its recommendation becomes a transaction.

Source: https://www.fsc.go.kr/eng/pr010101/87140

Japan - dialogue before prescriptive AI rules

Japan's FSA updated its financial-sector AI discussion paper in March 2026. It describes opportunities, misuse, misinformation and the risk of inaction, and invites continued dialogue. It is a discussion paper, not a detailed mandatory AI deployment code. The practical response is neither to wait for a perfect rule nor to call every experiment compliant. Keep a risk-ranked inventory, demonstrate controls under existing financial and data requirements, and record what needs supervisory discussion before scaling.

Source: https://www.fsa.go.jp/en/news/2025/20250304/aidp.html

Malaysia - the consultation stage matters

Bank Negara Malaysia's August 2025 discussion paper explores its proposed approach to AI in the financial sector and asks for feedback. Do not label that paper a final AI policy document. For financial institutions, it is a signal to connect AI use to customer outcomes and existing technology-risk controls while the policy develops. A board should know which deployments are customer-facing, which depend on third-party models and where human challenge is real rather than a box on a form.

Source: https://www.bnm.gov.my/-/dp-aifs25

New Zealand - technology-neutral does not mean responsibility-neutral

The FMA says it wants responsible AI adoption and expects good governance as firms manage market manipulation, systemic, consumer and ethical risks. It points to board and senior-management accountability and clear roles across the AI lifecycle, including human intervention. That is a supervisory position rather than a new dedicated bank-AI rule. For firms using AI in fraud, advice or customer service, the test is whether oversight can follow the decision from input to customer outcome.

Source: https://www.fma.govt.nz/library/opinion/ai-and-good-governance/

Mainland China - a binding bank and insurance AI opinion

Reported, not regulator-confirmed: On 18 June 2026 the National Financial Regulatory Administration issued Guiding Opinions on the Secure Development and Application of AI in the Banking and Insurance Sectors, effective on issue. They cover governance, data, risk and security. They require full life-cycle management and AI risk grading. Generative AI in scenarios such as fund transactions, credit approval, underwriting and claims is high-risk and needs risk management committee approval. Externally sourced generative AI models must be filed with the cyberspace authorities. Personal information such as names, ID and card numbers may not be used to train or tune generative models. Public-facing or high-risk use must be reported to the NFRA. This is the most prescriptive text in the region.

Sources: https://www.grandwaylaw.com/en/Legalupdate/4679.html ; https://policy.mofcom.gov.cn/claw/clawContent.shtml?id=106182

Taiwan - FSC guidelines for AI in the financial industry

Reported, not regulator-confirmed: The Financial Supervisory Commission released its Guidelines for AI Applications in the Financial Industry on 20 June 2024. They are guidelines for financial institutions, not a statute. I read the listing and the first page only, so the specific control requirements are not summarised here.

Source: https://law.fsc.gov.tw/EngLawContent.aspx?id=2690&lan=E

Thailand - Bank of Thailand AI risk management policy

Reported, not regulator-confirmed: The Bank of Thailand issued a Policy Framework on AI System Risk Management on 12 September 2025. It applies to financial institutions under the Financial Institution Act and payment service providers under the Payment System Act, and aims at reliable, unbiased, transparent and secure AI use. I relied on a law-firm summary of it and on the BOT document link.

Sources: https://www.lexology.com/library/detail.aspx?g=2944cbce-8539-4091-92a8-df74be1e83f0 ; https://www.bot.or.th/content/dam/bot/fipcs/documents/FOG/2568/ThaiPDF/25680178.pdf

Philippines - BSP governance principles for AI

In June 2026 the Bangko Sentral ng Pilipinas issued Memorandum M-2026-031 with a guidance paper, Governance Principles for AI in Financial Services, to all BSP-supervised institutions. It is a principles-based paper, proportionate to the nature and complexity of each institution’s AI use, built around what it calls the STARS of AI governance, and it maps controls to the AI system lifecycle. It sets minimum supervisory expectations and asks institutions to build their own AI governance framework.

Source: https://www.bsp.gov.ph/Regulations/Issuances/2026/M-2026-031.pdf

Vietnam - an AI law and a draft banking circular

Reported, not regulator-confirmed: Vietnam passed a Law on Artificial Intelligence, No. 134/2025/QH15, on 10 December 2025. The State Bank of Vietnam has also released a draft circular on AI in banking, reported in April 2026. It would require risk classification of AI systems, human oversight, impact assessments for high-risk AI, disclosure that a customer is dealing with AI, and incident reporting to the SBV within 24 hours. A draft is not a rule. I did not confirm whether the circular has been finalised or when the law takes effect.

Sources: https://english.luatvietnam.vn/law-no-134-2025-qh15-dated-december-10-2025-of-the-national-assembly-on-artificial-intell ; https://www.bakermckenzie.com/en/insight/publications/2026/04/vietnam-draft-circular-on-ai-deployment-in-banking

The operating model that travels

One region, no single checklist. Build a common control spine - use-case inventory, accountable owner, model and data provenance, independent challenge, human authority, testing for performance and bias, monitoring, third-party assurance, incident response and exit. Then add a jurisdiction layer: instrument status, scope, supervisory engagement, local data and customer duties, and whether the system can move from trial to production.

Do not flatten a consultation, a sandbox, a guideline and a binding prudential standard into the same green cell in a spreadsheet. The status is part of the risk. If a bank cannot say which version of the model made a decision, which local rule applied and who could stop it, it does not yet have a regional AI operating model. It has a shared demo.

Next issue: Middle East.

Source and status note

Positions checked against official regulator materials available by 25 September 2026. This is a market map, not legal advice; verify subsequent finalisation, amendments and institution-specific obligations before deployment. The country entries deliberately distinguish consultation, report, guidance, sandbox and live prudential duties. Mainland China, Taiwan, Thailand, the Philippines and Vietnam were added on 11 October 2026. Taiwan and Thailand rest on a listing and a law-firm summary, and Vietnam’s circular is a draft.

THE AI REGULATION MARKET MAP • APAC • RAW COMPLIANCE

Fourteen markets. Fourteen different routes to the same question: who owns the model, and who cleared it?

Oonagh van den Berg | RAW Compliance

A bank cannot put "APAC" in the jurisdiction field of its AI inventory. It can use the same model across the region, but it cannot assume the same regulatory route to production.

Singapore is consulting on an AI risk-management framework for financial institutions. South Korea has updated sector guidelines. Hong Kong has long-standing supervisory principles and an expanded sandbox. Indonesia has issued bank-specific governance guidance. Elsewhere supervisors are applying licensing, conduct, data, cyber and outsourcing duties to AI without pretending that a discussion paper is a law.

Issue 1 of this map examined Africa, where Kenya's draft bank-AI guidance proposed a written pre-deployment approval gate. APAC shows a different pattern: some supervisors define the control spine, some open testing lanes, and some make firms prove that existing obligations survived a new technology. None is a permission slip to deploy first and find an owner later.

Singapore - from FEAT principles to proposed lifecycle controls

MAS's FEAT principles put fairness, ethics, accountability and transparency on the table. Its November 2025 consultation proposed AI risk-management guidelines for all financial institutions, covering oversight, policies, lifecycle controls and capability. The consultation closed in January 2026. A consultation is not a final binding rule. But the direction is clear enough to test an inventory now: what is the use case, where is the model, who approves it, how is it monitored, and what does the firm do when it fails? A regional vendor that only offers a model card will need to show how that card feeds the bank's own control record.

Source: https://www.mas.gov.sg/publications/consultations/2025/consultation-paper-on-guidelines-on-artificial-intelligence-risk-management

Australia - the governance gap is the story

ASIC's review of 23 financial and credit licensees warned of a gap between AI adoption and governance. It is a supervisory finding, not a new AI-specific licensing regime. APRA's operational-risk standard CPS 230 adds live duties around critical operations and material service providers for regulated entities. Put the two together and the board question is practical: can the firm see where AI touches a customer outcome or important service, and can it manage the provider behind it? "Our supplier handles the model" does not answer who owns the consumer and resilience risk.

Sources: https://www.asic.gov.au/regulatory-resources/find-a-document/reports/rep-798-beware-the-gap-governance-arrangements-in-the-face-of-ai-innovation ; https://www.apra.gov.au/news-and-publications/apras-new-prudential-standard-operational-risk-management-comes-force

Hong Kong - principles plus a supervised test lane

HKMA's high-level AI principles for authorised institutions predate the current generative-AI rush. In March 2026 it expanded its GenAI Sandbox with other regulators. The sandbox is a place to test, not a general exemption from bank governance or customer protection. Banks should keep the experimental boundary explicit: which use case was tested, with what data, what humans can override, and what evidence would justify wider deployment? A successful demonstration is not itself a production approval.

Sources: https://brdr.hkma.gov.hk/eng/doc-ldg/docId/20191101-1-EN ; https://brdr.hkma.gov.hk/eng/doc-ldg/docId/20260305-2-EN

India - FREE-AI is a framework proposal, not a licence

RBI's August 2025 FREE-AI committee report sets out a responsible and ethical enablement framework for the financial sector. It is a committee report, not an enacted blanket AI code. Its value for a bank is the discipline of joining innovation to accountability and human oversight. That is where a procurement checklist often falls short: approval of the tool is not approval of every decision it will influence. The decision path needs named owners and a way to challenge an output.

Source: https://www.rbi.org.in/Scripts/PublicationReportDetails.aspx?ID=1306

Indonesia - bank-specific governance guidance

OJK's 2025 Artificial Intelligence Governance for Indonesian Banks was issued as guidance and a minimum benchmark for responsible bank AI. OJK positions it alongside existing digital-transformation, IT, cyber, digital-maturity and resilience requirements. That means the implementation work is not another slide deck called "AI ethics". It is the join between model design, data quality, prudential risk, cybersecurity, customer treatment and ongoing assurance. Guidance is not the same thing as a new pre-deployment approval rule.

Source: https://ojk.go.id/en/Publikasi/Roadmap-dan-Pedoman/Perbankan/Pages/Indonesia-Artificial-Intelligence-Governance-for-Banking.aspx

South Korea - sector guidelines, with a legal perimeter

The FSC introduced updated financial-sector AI guidelines in June 2026, effective from 22 June. It calls them self-regulatory and recommends risk-proportionate use by firms. Seven principles cover governance, legality, human-supervised assistance, credible data and models, financial stability, good faith and security. The FSC also says high-impact AI covered by Korea's AI Framework Act must meet the applicable law. This is the distinction to preserve: voluntary sector guidelines do not erase binding obligations. For agentic payments, ask who can authorise the agent and who remains responsible when its recommendation becomes a transaction.

Source: https://www.fsc.go.kr/eng/pr010101/87140

Japan - dialogue before prescriptive AI rules

Japan's FSA updated its financial-sector AI discussion paper in March 2026. It describes opportunities, misuse, misinformation and the risk of inaction, and invites continued dialogue. It is a discussion paper, not a detailed mandatory AI deployment code. The practical response is neither to wait for a perfect rule nor to call every experiment compliant. Keep a risk-ranked inventory, demonstrate controls under existing financial and data requirements, and record what needs supervisory discussion before scaling.

Source: https://www.fsa.go.jp/en/news/2025/20250304/aidp.html

Malaysia - the consultation stage matters

Bank Negara Malaysia's August 2025 discussion paper explores its proposed approach to AI in the financial sector and asks for feedback. Do not label that paper a final AI policy document. For financial institutions, it is a signal to connect AI use to customer outcomes and existing technology-risk controls while the policy develops. A board should know which deployments are customer-facing, which depend on third-party models and where human challenge is real rather than a box on a form.

Source: https://www.bnm.gov.my/-/dp-aifs25

New Zealand - technology-neutral does not mean responsibility-neutral

The FMA says it wants responsible AI adoption and expects good governance as firms manage market manipulation, systemic, consumer and ethical risks. It points to board and senior-management accountability and clear roles across the AI lifecycle, including human intervention. That is a supervisory position rather than a new dedicated bank-AI rule. For firms using AI in fraud, advice or customer service, the test is whether oversight can follow the decision from input to customer outcome.

Source: https://www.fma.govt.nz/library/opinion/ai-and-good-governance/

Mainland China - a binding bank and insurance AI opinion

Reported, not regulator-confirmed: On 18 June 2026 the National Financial Regulatory Administration issued Guiding Opinions on the Secure Development and Application of AI in the Banking and Insurance Sectors, effective on issue. They cover governance, data, risk and security. They require full life-cycle management and AI risk grading. Generative AI in scenarios such as fund transactions, credit approval, underwriting and claims is high-risk and needs risk management committee approval. Externally sourced generative AI models must be filed with the cyberspace authorities. Personal information such as names, ID and card numbers may not be used to train or tune generative models. Public-facing or high-risk use must be reported to the NFRA. This is the most prescriptive text in the region.

Sources: https://www.grandwaylaw.com/en/Legalupdate/4679.html ; https://policy.mofcom.gov.cn/claw/clawContent.shtml?id=106182

Taiwan - FSC guidelines for AI in the financial industry

Reported, not regulator-confirmed: The Financial Supervisory Commission released its Guidelines for AI Applications in the Financial Industry on 20 June 2024. They are guidelines for financial institutions, not a statute. I read the listing and the first page only, so the specific control requirements are not summarised here.

Source: https://law.fsc.gov.tw/EngLawContent.aspx?id=2690&lan=E

Thailand - Bank of Thailand AI risk management policy

Reported, not regulator-confirmed: The Bank of Thailand issued a Policy Framework on AI System Risk Management on 12 September 2025. It applies to financial institutions under the Financial Institution Act and payment service providers under the Payment System Act, and aims at reliable, unbiased, transparent and secure AI use. I relied on a law-firm summary of it and on the BOT document link.

Sources: https://www.lexology.com/library/detail.aspx?g=2944cbce-8539-4091-92a8-df74be1e83f0 ; https://www.bot.or.th/content/dam/bot/fipcs/documents/FOG/2568/ThaiPDF/25680178.pdf

Philippines - BSP governance principles for AI

In June 2026 the Bangko Sentral ng Pilipinas issued Memorandum M-2026-031 with a guidance paper, Governance Principles for AI in Financial Services, to all BSP-supervised institutions. It is a principles-based paper, proportionate to the nature and complexity of each institution’s AI use, built around what it calls the STARS of AI governance, and it maps controls to the AI system lifecycle. It sets minimum supervisory expectations and asks institutions to build their own AI governance framework.

Source: https://www.bsp.gov.ph/Regulations/Issuances/2026/M-2026-031.pdf

Vietnam - an AI law and a draft banking circular

Reported, not regulator-confirmed: Vietnam passed a Law on Artificial Intelligence, No. 134/2025/QH15, on 10 December 2025. The State Bank of Vietnam has also released a draft circular on AI in banking, reported in April 2026. It would require risk classification of AI systems, human oversight, impact assessments for high-risk AI, disclosure that a customer is dealing with AI, and incident reporting to the SBV within 24 hours. A draft is not a rule. I did not confirm whether the circular has been finalised or when the law takes effect.

Sources: https://english.luatvietnam.vn/law-no-134-2025-qh15-dated-december-10-2025-of-the-national-assembly-on-artificial-intell ; https://www.bakermckenzie.com/en/insight/publications/2026/04/vietnam-draft-circular-on-ai-deployment-in-banking

The operating model that travels

One region, no single checklist. Build a common control spine - use-case inventory, accountable owner, model and data provenance, independent challenge, human authority, testing for performance and bias, monitoring, third-party assurance, incident response and exit. Then add a jurisdiction layer: instrument status, scope, supervisory engagement, local data and customer duties, and whether the system can move from trial to production.

Do not flatten a consultation, a sandbox, a guideline and a binding prudential standard into the same green cell in a spreadsheet. The status is part of the risk. If a bank cannot say which version of the model made a decision, which local rule applied and who could stop it, it does not yet have a regional AI operating model. It has a shared demo.

Next issue: Middle East.

Source and status note

Positions checked against official regulator materials available by 25 September 2026. This is a market map, not legal advice; verify subsequent finalisation, amendments and institution-specific obligations before deployment. The country entries deliberately distinguish consultation, report, guidance, sandbox and live prudential duties. Mainland China, Taiwan, Thailand, the Philippines and Vietnam were added on 11 October 2026. Taiwan and Thailand rest on a listing and a law-firm summary, and Vietnam’s circular is a draft.

THE AI REGULATION MARKET MAP • APAC • RAW COMPLIANCE