Skip to main content
News

Thailand's PDPC Issues Its First PDPA Guide for AI: Prompts, Logs and AI Agents Are Now in Scope

In October 2026 Thailand's Personal Data Protection Committee published version 1.0 of its guide on applying the PDPA to the development and use of AI, covering generative and agentic AI. Here is what companies running chatbots, AI document search or AI agents should check before they scale up.

17 Oct 20266 minOffice of the Personal Data Protection Committee (PDPC), Thailand
PDPAPDPCAI GovernanceAgentic AIData PrivacyAI

Picture a company where sales runs a chatbot on the website, accounting uses AI to summarise supplier invoices, and IT is piloting an AI agent that pulls customer records from the back office to draft emails. All of it touches personal data. Until now, Thailand's data protection regulator had not written down how the PDPA applies to work like this.

Now one has.

What happened

On 7 October 2026 the Office of the Personal Data Protection Committee (PDPC) posted version 1.0 of its guidelines on personal data protection in the development and use of artificial intelligence. The document is an 80-page PDF in Thai.

The cover states that the Personal Data Protection Committee issued it under Section 16(3) of the Personal Data Protection Act B.E. 2562 (2019), and that it was prepared by the PDPC Office together with the Electronic Transactions Development Agency (ETDA).

The PDPC had opened a draft of the same guidance for public comment in February 2026, so this is the post-consultation version. Thai business daily Thansettakij reported the release on 13 October, and law firm Tilleke & Gibbins published an English summary on 14 October.

The guide says plainly that it is advice only. It has no binding legal force and does not bind any finding of an offence. The PDPA it interprets is already in force, though, so the document shows how the regulator reads existing law when AI is involved.

Five points from the executive summary

The guidance covers analytic, generative and agentic AI, from planning and data preparation through training, deployment and decommissioning. Its executive summary lists five points.

  1. Personal data in AI takes more forms than people expect. Prompts, usage logs, embeddings and AI-generated output can all be personal data if a person can still be identified directly or indirectly. A model trained on personal data is not automatically anonymous.
  2. Each organisation must state whether it is a controller or a processor for each activity, based on who actually decides purposes and means rather than on role labels. If an AI provider uses customer data to train its own models, that provider becomes a controller in its own right.
  3. Consent should be the last option, not the first. Organisations should work through the legal bases in Sections 24 and 26 first. Relying on legitimate interest requires a documented assessment (LIA). Relying on consent alone creates operational risk when people withdraw it.
  4. High-risk AI projects, such as automated decisions that affect people's rights or training on sensitive data, should go through a data protection impact assessment (DPIA) before they start, along with AI-specific controls such as input and output guardrails, restricted permissions for AI agents, and logging that supports later investigation.
  5. Data subject rights still apply when the technology is complex. Removing data from a trained model is technically hard, but that does not remove the legal duty. Systems should be designed so data can be deleted, for example by keeping it in a retrieval (RAG) store separate from the model and filtering output until the next model update.

What it says about AI agents

The security chapter recommends that every AI agent have its own identity that can be traced, fine-grained access rights, and a way to verify which user an agent is acting for, so that someone can be held responsible if data leaks.

It also recommends least-privilege access for agents, limits on write access to sensitive databases, checks on the connections between agents and other systems such as the Model Context Protocol (MCP), and sandboxed code execution to contain damage if an agent is attacked or misbehaves.

For everyday staff, the guide recommends an acceptable use policy for AI with explicit prohibitions. Its example is a ban on putting customers' personal data into prompts for public chatbots.

Why this matters for Thai businesses

Many companies adopted AI through overseas services without treating it as personal data processing. This guide says prompts and logs count. It also says that if a provider stores or uses personal data in an overseas data centre where others can access it, that may be a cross-border transfer subject to PDPA Sections 28 and 29.

The deadlines have not changed either. A personal data breach must be reported to the PDPC within 72 hours of becoming aware of it under Section 37(4). That is hard to do if your AI system keeps no record of who asked what and what the AI sent back.

What to check with your team

The guide includes six questions that boards and senior management should ask before approving an AI project. Here they are as a short checklist.

  1. Are we the controller or the processor in this project, and does our contract with the AI provider say whether our data may be used to train its models?
  2. What legal basis are we using, and if it is legitimate interest, is there a written LIA?
  3. Does this project meet the high-risk criteria that call for a DPIA?
  4. Is there a person with real authority to override the system wherever it affects someone's rights?
  5. If a customer asks to see or delete their data, do we know where it sits across prompts, logs and vector stores?
  6. If data leaks through AI, do we have the logs and the response plan to notify the PDPC within 72 hours?

The appendices also contain separate checklists for organisations that build AI and those that use it, which a DPO can use for a first readiness review.

What Enersys does

We have built and supported business software for 14 years, and our work covers AI and personal data protection. Our PrivacyHub system handles intake and tracking of data subject requests. If your company runs chatbots or AI agents on customer data and wants to map where that data goes, who can reach it, and how you would answer a rights request, talk to the Enersys team.

Sources

Back to Insights

Related Articles

AI That Writes Code Needs Boundaries: GitHub Adds a Sandbox for Coding Agents, and Google Stops Selling Gemini Code Assist

In the first week of October 2026, GitHub let organisations limit what AI coding agents can reach on developer machines and control spending on AI code review, while Google stopped selling Gemini Code Assist Standard and Enterprise. What businesses with in-house or outsourced developers should ask.

AIS x Thai SME Council Launches ProStart — Digital + AI Bundle with 200% Tax Deduction Thai SMEs Shouldn’t Miss

AIS has partnered with the Thai SME Council to launch ProStart, an AI + Digital package for SMEs with a 200% tax deduction worth up to ฿300,000 — a golden opportunity available only until the end of 2027.

Anthropic Hits $30B ARR — Enterprise AI Has Officially Crossed the Tipping Point

Anthropic tripled from $9B to $30B run rate in just four months, with 1,000+ enterprise customers each spending over $1M a year. This is no longer hype — it is proof that Enterprise AI has real ROI at scale, and Thai businesses need to decide when (not if) to move.

"Empowering Innovation,
Transforming Futures."

Contact us to make your project a reality.