Thailand - AI Risk Management Guidelines

Overview

The Bank of Thailand (BOT) published for public consultation a draft entitled Guiding Principles for Artificial Intelligence Risk Management on 12 June 2025 and invited comments through 30 June 2025. The draft is targeted at financial institutions and payment-service providers under BOT supervision and supplements existing guidance on information-technology risk, third-party risk, data governance and market conduct. It defines AI to include machine learning, deep learning, generative AI such as large language models (LLMs) and agentic AI, and explicitly excludes simple rule-based automation such as robotic process automation (RPA). The draft organises expectations across governance (board and senior management oversight, FEAT principles), lifecycle risk management (inventory, classification and testing), development and deployment controls (data quality, model validation and cybersecurity) and consumer protections (disclosure, human review). The original draft and public-hearing announcement are available on the BOT website: BOT public-hearing page (12 June 2025) and the full draft PDF: BOT draft PDF (Guiding Principles for AI Risk Management).

Definitions

The draft provides precise definitions that frame obligations and scope. "AI system" covers systems that emulate human intelligence to learn, memorise, decide, operate and generate new content by learning from data, including ML, DL, generative models (LLMs) and agentic systems. Excluded are RPA and pre-defined condition-matching automation. "Financial service provider" means banks, specialised financial institutions and payment-system providers under the Financial Institutions Business Act and the Payment Systems Act. The draft also defines three principal risk categories—data risk, model-development risk (including hallucination and opacity), and AI-specific cybersecurity threats (prompt injection, model inversion, data poisoning, adversarial attacks)—to guide downstream controls and testing requirements.

Governance and Institutional Framework

The BOT draft places primary accountability with boards and senior management: boards must ensure an AI usage policy aligned with organisational strategy and FEAT (fairness, ethics, accountability, transparency) principles, approve risk appetite for AI use-cases, and oversee implementation. Management must assign clear responsibilities across the three lines of defence, create an AI governance function or centre of excellence where appropriate, and ensure workforce training and awareness. The guidance expects comprehensive AI inventories (model registries), categorisation by materiality and risk, and periodic reporting to the board. Governance expectations include formal change-management controls, pre-deployment approvals, and documented human-in-the-loop arrangements for material customer-facing or high-impact decisions. For reference and submission of comments the BOT provided the public-hearing page and the supporting PDF: BOT public-hearing page and draft PDF.

Key Focus Areas

The draft outlines core focus areas for risk management across the AI lifecycle. First, data governance: institutions must assess and ensure data quality, provenance, freshness, representativeness and appropriate labeling; implement privacy-preserving measures and cross-border safeguards where data or models are hosted offshore. Second, model development and validation: define metrics (accuracy, reliability, fairness, robustness), conduct out-of-sample and adversarial tests, and validate models on holdout and real-world data including edge cases. Third, generative-AI mitigations: for LLMs and other generative systems, the draft requires controls to reduce hallucinations (e.g., retrieval-augmented generation, factuality checks, constrained prompts), output verification and traceability of supporting sources. Fourth, cybersecurity and model security: integrate defenses against prompt injection, model inversion and poisoning; apply access controls, secrets management, and robust monitoring. Fifth, third-party and vendor management: perform due diligence, require contractual audit rights, and maintain contingency plans. Sixth, transparency and consumer protection: disclose material AI use to affected customers, provide mechanisms for human review or opt-out where decisions materially affect customers, and monitor outcomes to detect discriminatory or unfair treatment.

Implementation Framework

The draft advises a pragmatic, risk-based implementation approach. Steps include: (1) inventorying AI systems and classifying them by impact and materiality; (2) defining AI policies and governance structures; (3) integrating AI risk controls into existing IT and model-risk frameworks (MLOps, documentation and audit trails); (4) establishing pre-deployment assurance (validation, bias and robustness testing, security assessment) and operational controls (monitoring, drift detection, retraining criteria); and (5) codifying incident response and rollback procedures. The BOT expects institutions to adopt measurable KPIs for model performance and safety, maintain versioned records (model lineage), and implement continuous monitoring and periodic independent reviews. The draft recommends that organisations align development pipelines with privacy and data-protection obligations under Thailand's PDPA and applicable contractual commitments to third-party providers.

Monitoring and Evaluation

Ongoing monitoring is emphasised as essential. The draft requires continuous performance monitoring against pre-defined metrics, drift detection, fairness and discrimination checks, and logging sufficient to support investigations and audits. Institutions should maintain metrics dashboards, alerting thresholds, and processes for rapid remediation and rollback. Periodic independent model reviews and stress-testing (including adversarial scenarios) are recommended. Supervisory reporting lines should be established for material incidents, and institutions should retain records of tests, validation results and governance approvals for supervisory inspection.

Penalties, Liability, and Appeals

The draft itself is framed as supervisory guidance and does not set out a new standalone schedule of monetary fines. Enforcement is expected to be through the BOT's existing supervisory powers and existing legal frameworks (e.g., Financial Institutions Business Act, Payment Systems Act, and data-protection law). The BOT may apply corrective measures, require remediation plans, restrict or suspend use of AI systems, and apply public supervisory actions where necessary. Liability for legal claims or consumer redress remains governed by existing civil and consumer-protection laws; institutions are required to maintain channels for redress and have procedures to investigate and remediate customer harms arising from AI-driven decisions. The draft indicates institutions should document appeals or human-review processes for customers affected by AI decisions.

Relationship to Other Instruments

The draft is explicitly intended to complement, not replace, existing BOT guidance on information-technology risk management, third-party risk management, data governance and market conduct. It should be read together with Thailand's Personal Data Protection Act (PDPA), the Financial Institutions Business Act, the Payment Systems Act, and other sectoral regulations. The BOT positions the guidance as a sector-specific articulation of internationally accepted principles and intends it to fit within existing supervisory frameworks so that institutions can leverage existing compliance processes and controls. The BOT public-hearing page contains links and contextual references to those frameworks: BOT public-hearing page.

International Alignment

The draft aligns with internationally recognised good practice for AI in finance: risk-based oversight, lifecycle controls, model validation and explainability expectations comparable to supervisory work by authorities such as the Monetary Authority of Singapore, the UK PRA/FCA, and selected guidance emerging from EU and US regulatory initiatives. The BOT highlights FEAT principles (fairness, ethics, accountability, transparency) and references global approaches to generative AI safety, model security and third-party governance to help Thai financial institutions apply internationally accepted controls while remaining proportionate to risk and institutional size.

Implementation Timeline

Sources and References