AI Governance in Thailand 2026: BOT and OIC Guide
Updated: 4 days ago
AI is already part of financial services in Thailand. Banks use it for fraud detection and credit scoring. Fintechs use it for onboarding, transaction monitoring and customer support. Insurers use it for underwriting, pricing and claims handling.
The question in 2026 is no longer whether AI should be used. The question is how it should be governed.
Thailand does not yet have a single, dedicated AI law. Instead, financial institutions must follow a risk-based approach shaped by sector regulators, especially the Bank of Thailand and the Office of Insurance Commission.
This guide explains the key expectations from BOT and OIC, where they overlap, and what banks, fintechs and insurers should do now.
This article is for general information only. It is not legal advice or regulatory advice.

Thailand is using risk-based AI governance
Thailand currently relies on a mix of existing laws, regulator guidelines and sector-specific supervision. This matters because AI risk is not the same in every use case.
An AI chatbot that answers general product questions creates one type of risk. An AI model that rejects a loan, prices an insurance policy or flags a transaction as suspicious creates a much higher risk.
A risk-based approach means institutions should apply stronger controls where AI can affect:
Customer rights
Access to financial products
Pricing, underwriting or credit decisions
Claims outcomes
Fraud alerts and account restrictions
Personal data and sensitive data
Operational resilience
Cybersecurity
Market trust
For financial services, regulators expect AI to be governed like any other material risk. That means clear ownership, board oversight, proper testing, ongoing monitoring and records that show how decisions were made.
The practical message is simple. If AI can affect a customer, a balance sheet, a claim or a regulated process, it needs governance.
BOT expectations for banks and fintechs
The Bank of Thailand has moved toward clearer AI governance through its September 2025 AI Risk Management Guidelines. These guidelines are especially relevant for banks, financial institutions and fintechs operating in regulated financial services.
The BOT approach focuses on accountability, fairness, explainability, data quality, model risk, cybersecurity and customer protection.
Boards remain accountable for AI
The BOT expects senior leadership and boards to understand where AI is used and what risks it creates.
This does not mean every board member must become a machine learning engineer. It does mean the board should receive enough information to challenge management and make informed decisions.
At a minimum, board and senior management oversight should cover:
Where AI is used across the organisation
Which AI systems are high risk
Who owns each AI system
How AI risks are measured and reported
Whether controls are working
How customer complaints and incidents are handled
Whether outsourcing or third-party AI tools create extra risk
Board accountability also means AI cannot sit only with the technology team. Legal, compliance, risk, cybersecurity, data protection, product and business teams all need a role.
A useful governance model assigns clear responsibility across three layers:
Business owner Owns the use case, customer impact and performance.
Model or technology owner Owns technical design, testing, monitoring and changes.
Independent control functions Review risk, compliance, data protection, cyber controls and audit readiness.
BOT expects the FEAT principles
The BOT guidance refers to principles commonly described as FEAT.
FEAT stands for:
Fairness AI should not create unfair outcomes or unjustified discrimination.
Ethics AI should align with responsible conduct and customer protection.
Accountability People remain responsible for AI decisions and outcomes.
Transparency Customers and internal users should understand when and how AI is used, especially when it affects them.
These principles sound simple, but they require practical controls.
For example, a credit scoring model should not only be tested for accuracy. It should also be checked for unfair outcomes across relevant customer groups, within the limits of applicable data protection rules.
A fraud model should not only reduce losses. It should also avoid excessive false positives that block customers from legitimate transactions without a fair review path.
Data risk is a core BOT concern
AI systems depend on data. Poor data can create poor outcomes.
Financial institutions should be able to show that data used for AI is:
Lawfully collected and used
Relevant to the purpose
Accurate enough for the decision being made
Protected against unauthorised access
Checked for bias or imbalance
Governed through clear retention and access rules
This also connects with Thailand’s Personal Data Protection Act. If AI uses personal data, the institution should confirm its lawful basis, notices, consent where needed, data minimisation and cross-border transfer controls.
For higher-risk AI, teams should keep records of:
Training data sources
Data cleaning steps
Excluded fields
Known data limits
Data quality checks
Data access approvals
Sensitive data handling
Good data governance makes AI easier to explain, test and defend.
Model risk needs full lifecycle control
AI models change over time. Their performance can drift as customer behaviour, fraud patterns, economic conditions or product rules change.
BOT-style model governance should cover the full lifecycle:
Use case approval
Risk rating
Model design
Data selection
Validation before launch
Approval before deployment
Monitoring after launch
Change control
Incident response
Retirement or replacement
Higher-risk models should have stronger validation. That may include independent testing, challenger models, stress testing and review of false positives and false negatives.
Institutions should avoid treating AI as a one-time technology project. AI governance is an ongoing process.
Cyber and third-party risks must be managed
AI can widen the attack surface.
A bank or fintech using AI should consider risks such as:
Model manipulation
Data poisoning
Prompt injection
Unauthorised access to training data
Leakage of customer information
Over-reliance on external AI vendors
Weak controls in cloud-based model environments
Third-party AI tools need careful review. The institution remains responsible even when a vendor supplies the model, platform or data service.
Vendor governance should cover:
Contractual audit rights
Data use limits
Security controls
Incident notification
Subcontracting
Model performance reporting
Exit plans
Business continuity
If a vendor cannot explain how a material AI system works, that should raise a red flag.
Customers should be told when AI is used
A key BOT expectation is that customers should be informed when AI is used in ways that affect them.
This is especially important when AI supports:
Credit decisions
Product recommendations
Fraud alerts
Account restrictions
Customer service chats
Complaint routing
Risk profiling
Customer notices do not need to reveal trade secrets. They should be clear enough for customers to understand that AI is involved, what it is used for, and how they can ask for help or review.
Plain language matters. A notice that says “automated decisioning may be deployed to enhance service delivery” is not helpful.
A better version would be:
We may use automated tools to help assess your application. If your application is declined, you can contact us to request more information or ask for a review.
That kind of wording supports transparency without overloading the customer.

OIC expectations for insurers
The Office of Insurance Commission supervises the insurance sector in Thailand. Its AI governance expectations are especially important for underwriting, pricing, claims and customer servicing.
Insurers use large amounts of data. They also make decisions that can have serious effects on customers. An AI tool that influences premium pricing or claim approval needs stronger controls than a tool that sorts internal documents.
The OIC focus is likely to centre on governance, explainability, fairness and human oversight.
AI governance should match insurance risk
Insurers should create a clear AI governance framework that fits the way insurance products work.
This framework should define:
Which AI systems are allowed
Which AI systems need approval before launch
How use cases are risk-rated
Who owns the model
Who reviews customer impact
What records must be kept
When human review is required
How complaints are escalated
Insurance AI governance should cover both internal and customer-facing systems.
Examples include:
Underwriting tools
Dynamic pricing models
Claims fraud detection
Claims triage tools
Chatbots
Renewal recommendation engines
Customer risk segmentation
Agent support tools
Even if the AI does not make the final decision, it can still influence the outcome. A claim marked as “high risk” by an AI tool may receive different treatment. That influence should be governed.
Explainability matters in underwriting and claims
Insurance decisions must often be explained to customers, agents, brokers, regulators and internal review teams.
This creates a challenge for complex AI models. Some models are accurate but hard to explain. For high-impact decisions, this can be a problem.
Insurers should be able to explain:
What main factors influenced a decision
Whether AI was used
Whether a human reviewed the result
How a customer can challenge the outcome
What data was considered
Which factors were not allowed or were controlled
Explainability does not always mean opening the full model code. It means giving meaningful reasons that people can understand.
For example, if an AI tool supports a travel insurance claim review, the insurer should be able to explain the main reason for delay, rejection or escalation. A vague reference to “system rules” will not be enough.
Fairness is central to pricing and eligibility
AI can create unfair outcomes even when no one intended it.
A model may use data points that act as proxies for sensitive traits. A pricing model may treat similar customers differently in ways that cannot be justified by risk. A claims model may flag some groups more often because historical data reflects past bias.
Insurers should test AI systems for fairness before launch and during use.
Fairness reviews may include:
Checking whether model inputs are appropriate
Testing outcomes across relevant customer groups
Reviewing whether price differences can be justified
Removing or limiting problematic variables
Monitoring complaints and appeals
Reviewing edge cases manually
Keeping records of fairness testing
The goal is not to force every customer to receive the same price or outcome. Insurance depends on risk-based pricing and underwriting. The goal is to make sure differences are lawful, explainable and fair.
Human-in-the-loop control is needed for key decisions
For underwriting, pricing and claims, OIC expectations point toward meaningful human oversight.
Human-in-the-loop means a trained person can review, challenge and change an AI-supported outcome.
This is most important where AI affects:
Policy acceptance or rejection
Premium pricing
Coverage exclusions
Claim approval or rejection
Claim settlement amounts
Fraud referrals
Customer complaints
Human review must be real. If staff simply click “approve” without understanding the AI output, the control is weak.
A good human review process should include:
Clear authority to override the AI
Access to the main reasons for the AI output
Training on model limits
Escalation routes for unusual cases
Records of overrides and reasons
Quality checks on reviewer decisions
Human oversight also helps insurers spot model drift. If reviewers keep overriding the same type of AI decision, the model may need retraining or redesign.
BOT and OIC guidance compared
BOT and OIC share many themes. Both focus on risk, accountability, fairness, explainability and customer protection.
The main difference is the sector context. BOT guidance applies to banking and financial services risks. OIC guidance focuses on insurance-specific decisions such as underwriting, pricing and claims.
Area | BOT focus for banks and fintechs | OIC focus for insurers |
Main concern | Safe and responsible AI in financial services | Fair and explainable AI in insurance decisions |
Governance | Board accountability, senior management oversight and risk controls | Clear AI ownership, approval and insurance-specific controls |
Core principles | FEAT principles, fairness, ethics, accountability and transparency | Governance, explainability, fairness and human oversight |
High-risk use cases | Credit scoring, fraud detection, onboarding, transaction monitoring and customer service | Underwriting, premium pricing, claims triage, claims approval and fraud detection |
Data risk | Data quality, lawful use, privacy, security and third-party data controls | Appropriate data use, bias checks and explainable rating or claims factors |
Model risk | Validation, monitoring, drift checks and change control | Testing, outcome review, fairness checks and review of contested decisions |
Cyber risk | Strong focus on cyber controls, vendor risk and operational resilience | Relevant where AI tools process policyholder or claims data |
Customer notice | Tell customers when AI is used in relevant decisions or customer interactions | Explain AI-supported outcomes and provide review paths |
Human review | Needed for material customer impact and exceptions | Critical for underwriting, pricing and claims decisions |
Best evidence of compliance | AI inventory, risk ratings, board reporting, validation records and customer notices | AI governance policy, explainability records, fairness testing and human review logs |
A practical AI governance checklist for 2026
Banks, fintechs and insurers should not wait for a single AI law before acting. The safer path is to build a practical governance programme now.
Use this checklist as a starting point.
Build an AI inventory
Create a central list of all AI systems, including vendor tools.
Include:
System name
Business owner
Purpose
Data used
Customer impact
Vendor involvement
Risk rating
Launch date
Review date
Shadow AI is a real risk. Teams may use AI tools without formal approval. The inventory should capture both approved systems and tools under review.
Classify AI by risk
Not every AI tool needs the same level of control.
Classify systems as low, medium or high risk based on:
Impact on customers
Use of personal or sensitive data
Financial impact
Regulatory impact
Explainability needs
Degree of automation
Ability to reverse the decision
High-risk systems should need formal approval before launch.
Set clear ownership
Every AI system should have an owner.
The owner should be responsible for:
Intended use
Performance
Customer impact
Issue management
Review cycles
Documentation
Ownership should not sit only with IT. Business and control functions must share responsibility.
Test before launch
Pre-launch testing should cover more than technical accuracy.
Test for:
Data quality
Bias and unfair outcomes
Model performance
Explainability
Security weaknesses
Operational impact
Customer notice requirements
Human review process
Document the results. Regulators will care about evidence, not only policy wording.
Monitor after launch
AI performance can change. Ongoing monitoring should check:
Accuracy
Error rates
Drift
False positives
False negatives
Complaints
Overrides
Incidents
Unusual customer impact
Set trigger points for review. For example, a spike in complaints or overrides should lead to investigation.
Prepare customer disclosures
Write clear customer notices for AI use.
Good notices should explain:
That AI or automated tools may be used
The purpose of the tool
The type of decision it supports
How the customer can ask for help
Whether human review is available
Keep the wording short. Customers should not need legal training to understand it.
Strengthen vendor controls
For third-party AI, review:
Data access
Security controls
Model explainability
Subcontractors
Audit rights
Incident handling
Data retention
Exit plans
If the vendor is central to a regulated process, treat the relationship as a material outsourcing risk.
Train staff who use AI
Staff need to understand what AI can and cannot do.
Training should cover:
Approved AI use cases
Prohibited uses
Data protection rules
Customer fairness
Human review duties
Escalation steps
Incident reporting
This is especially important for staff who review underwriting, credit, fraud or claims decisions.
Keep evidence ready
AI governance depends on records.
Keep evidence such as:
AI policy
AI inventory
Risk assessments
Approval records
Model validation reports
Fairness testing
Customer notices
Training logs
Vendor due diligence
Incident reports
Board papers
Good records make supervision, audits and internal reviews much easier.

What this means for financial institutions in Thailand
AI Governance in Thailand 2026 BOT and OIC Guide for Banks Fintechs and Insurers is not about waiting for a future AI Act. It is about showing that AI is controlled today.
For banks and fintechs, the BOT direction is clear. AI must sit within risk management, board oversight, cyber resilience, data governance and customer protection.
For insurers, OIC expectations point to explainable, fair and reviewable AI in underwriting, pricing and claims.
The strongest institutions will do three things well:
Know where AI is used
Control the risks based on customer impact
Keep evidence that shows responsible governance
A written AI policy is not enough. Regulators will expect working controls, trained people, review records and clear customer communication.
AI can improve financial services in Thailand, but only if customers can trust how it is used. If your organisation is building or reviewing its AI governance framework for 2026, contact us to discuss how we can help assess your current controls and prepare a practical BOT and OIC readiness plan.


