The Bank of Thailand issued its policy guideline on risk management of the use of artificial intelligence systems on 12 September 2025, by circular ธปท.ว. 5994/2568 to commercial banks, specialized financial institutions and non-bank payment service providers. It holds a provider responsible for how its AI systems work and for decisions based on their outputs, asks the board and senior management to set an AI policy and risk appetite in line with responsible-AI principles such as FEAT (fairness, ethics, accountability, transparency), and sets controls for data, model development and AI-specific cyber threats. It is guidance rather than a rule with penalties, it applies to bought-in AI as well as AI built in-house, and the BOT published it in Thai.
OFFICIAL TEXT: bot.or.th ↗ · FINAL · BANK OF THAILAND
| Document | BOT AI Risk Management Policy Guideline — Policy Guideline on Risk Management of the Use of Artificial Intelligence Systems (แนวนโยบายธนาคารแห่งประเทศไทย เรื่อง การบริหารจัดการความเสี่ยงของการใช้งานระบบปัญญาประดิษฐ์) |
| Issued by | Bank of Thailand (BOT), Thailand's central bank and the supervisor of commercial banks, specialized financial institutions and payment service providers |
| Type | Guidance |
| Status | Final |
| Published | Sep 12, 2025 |
| Applies to | Commercial banks registered in Thailand, specialized financial institutions and non-bank electronic payment service providers, the addressees of circular ธปท.ว. 5994/2568; the guideline's own scope is financial institutions and specialized financial institutions under the Financial Institutions Businesses Act and regulated payment system and payment service providers under the Payment Systems Act. A policy guideline (แนวนโยบาย): its practices are what providers should do, applied by use case, with no stated effective date. Published in Thai; the English title is this tracker's translation |
| Supersedes | BOT Draft AI Risk Management Guideline |
| Official source | bot.or.th ↗ |
| Use cases | AI governance (general) · Model risk management · Generative & agentic AI · Customer-facing chatbots · Cybersecurity · Data & privacy · Third-party & vendor AI · Credit scoring & underwriting |
What are the key points of BOT AI Risk Management Policy Guideline?
- Definition: an AI system imitates human intelligence and can learn, remember, decide, perform tasks or create new content by learning from data and building conditions automatically; automated systems whose steps humans define are excluded, and the Q&A names robotic process automation, condition matching and automated scripts as outside scope.
- Risks: three families, namely data risk (poor-quality data, leakage of personal or sensitive data), model development risk (unexplainable, unreliable, biased or false outputs) and cyber threats, including prompt injection, model inversion, data poisoning and adversarial attacks; providers also weigh use-case risks such as credit and operational risk.
- Principles: the provider is responsible for the AI system and for decisions based on its outputs; governs AI under generally accepted responsible-AI principles such as FEAT; manages risk across the AI lifecycle so that systems are secure, accurate, reliable and explainable; and treats customers transparently and fairly.
- Governance (Part 1): the board and senior management are accountable for AI decisions and actions, oversee an AI policy, assign roles across the three lines of defence and build staff skills so people do not over-rely on AI outputs; the AI policy is reviewed regularly, and AI risks and the organisation's risk appetite are identified and monitored continuously.
- Human oversight: where AI is used in core functions tied to strategic decisions (the examples are approving loans, account openings and deposits, withdrawals or transfers) or in work that directly affects customers, providers may consider human-in-the-loop or human-over-the-loop oversight as appropriate, scaled to risk and impact.
- Customers: when AI converses or communicates with customers in place of staff, providers should have a way to tell customers that AI is serving them and may offer a route to a staff member for customers who do not want to deal with the AI.
- Data and models (Part 2): data-quality criteria covering currency, accuracy, volume and diversity, checked before training; access control, masking, hashing and input sanitization against leakage; a data boundary separating corporate data when using a third party's generative AI; evaluation metrics and testing before and after deployment with unseen, noisy and edge-case data; retrieval-augmented generation or prompt engineering against hallucination; and documentation, reasoning or source references that make outputs explainable.
- Cyber (Part 2): prompt and response filtering so AI does not reveal sensitive information, continuous testing against attacks such as data poisoning and adversarial inputs, and threat monitoring with reference to the OWASP Machine Learning Security Top 10, the OWASP Top 10 for Large Language Model Applications and MITRE ATLAS.
What did BOT AI Risk Management Policy Guideline change for banks?
The guideline is the BOT's AI-specific supervisory document. It adds to, rather than replaces, the BOT rules it names: the IT risk rules, the IT risk management and third-party risk management implementation guidelines, the data governance policy and the market conduct rules. Compared with the June 2025 draft, it narrowed the customer-facing duty from a disable-or-bypass option to informing customers and possibly offering a route to staff, made human oversight a matter for providers to consider by risk, and added a Q&A; the BOT declined requests to add high-risk or prohibited tiers of AI.
Is the Bank of Thailand AI risk management guideline binding?
It is a policy guideline sent by circular, not a notification with penalties. Its practices are framed as what providers should do, and the BOT's appended Q&A says providers apply its principles and approaches according to how they use AI. The same Q&A says providers must still comply strictly with relevant laws, such as personal data protection and intellectual property law.
Does the BOT guideline apply to off-the-shelf AI such as a generative AI chatbot?
Yes. The Q&A says providers can apply the relevant parts to manage the risk of their AI systems whether they develop them in-house or use a third party's system. Part 2 adds that, when using a third party's generative AI, a provider should set a data boundary that keeps its own data separate.
Is there an English version of the BOT AI guideline?
The BOT published the guideline and its Q&A in Thai. The English title used here is a translation; the June 2025 consultation draft carried the BOT's own English title, Guiding Principles for Artificial Intelligence Risk Management.
| Date | Document | Status |
|---|---|---|
| Jun 12, 2025 | BOT Draft AI Risk Management Guideline — Draft Policy Guideline on Risk Management of the Use of Artificial Intelligence Systems (Guiding Principles for Artificial Intelligence Risk Management) | Superseded |
| Sep 14, 2026 | BOT Minimum Practice for Digital Fraud Management — Minimum Practice for Digital Fraud Management (แนวปฏิบัติขั้นต่ำสำหรับการบริหารจัดการภัยทุจริตดิจิทัล) | In force |
| Nov 19, 2026 | BNM AI Feedback Statement — Feedback Statement: Artificial Intelligence (AI) in the Malaysian Financial Sector | Final |
| Nov 19, 2026 | BoE/PRA/FCA effective practices on AI cyber risk (Nov 2026) — Effective practices: Cyber risk and resilience management relating to artificial intelligence (AI) | Final |
| Nov 19, 2026 | EBA Risk Assessment Report (Nov 2026) — Risk Assessment Report of the European Banking Authority, November 2026, with Box 1 on artificial intelligence related risks and opportunities | Final |
| Nov 17, 2026 | FSB Plenary: priorities for 2027 (Nov 2026) — FSB sets out its priorities for 2027 (Plenary meeting, Basel, 16-17 November 2026) | Final |
Follow every document these regulators publish
when one of these regulators moves, the next morning's brief says so · six sourced stories · 7 am PT · free
plus every regulation, bank and agent page update, the morning after · leave any morning