PDPA and AI: Where Your Data Can Live
PDPA does not say all data must stay in Thailand, but it sets conditions for transfers. Here is how to map your AI data flows and choose where each workload runs.
PDPA AI data residency is the question of where the data behind an AI system is stored and processed, and who can reach it. Thailand's Personal Data Protection Act (PDPA) does not say that all data must stay in Thailand. It does set conditions for sending personal data abroad, and it expects you to protect that data wherever it sits. So the practical answer is: you can use different locations for different data, if you know what flows where and can justify it.
This guide is for IT, legal, risk and data leads at Thai organizations that are planning an AI project. It explains the places AI data can live, the data flows people forget, and a simple way to decide. It is not legal advice. Confirm your own position with qualified counsel and check current guidance from Thailand's Personal Data Protection Committee (PDPC).
What "data residency" means for an AI project
Residency is often confused with two related ideas. Keeping them apart makes the conversation with your lawyers much shorter.
- Data residency: the physical or legal location where data is stored and processed.
- Data sovereignty: which country's laws and authorities can reach the data, even if it is stored elsewhere.
- Data transfer: the act of sending personal data from one place to another, including to a provider in another country.
An AI project creates more copies of data than a normal application. Your documents are uploaded, split into chunks, turned into embeddings, stored in a search index, sent to a model in a prompt, and often written to logs. Each step is a place where personal data can land.
What the PDPA asks of you, in plain terms
The PDPA applies to personal data about people in Thailand. The points below matter most for AI. We describe them in general terms, because the rules, notifications and guidance are updated over time.
- You need a lawful basis for processing, such as consent, contract or legitimate interest.
- You should collect and use only the data you need for a stated purpose.
- You must protect the data with appropriate security measures.
- Sending personal data to another country is allowed under conditions. These usually relate to the destination's protection standards, safeguards you put in place, or an exception that applies.
- Data subjects have rights, such as access and correction, and you must be able to act on them.
The PDPC publishes the regulations and guidance. Read the current versions rather than relying on a summary from a blog, including this one.
The five places AI data can live
Most designs fall into one of five patterns. They sit on a spectrum from "most control" to "least effort".
| Pattern | Where data and models run | What you control | Typical trade-off |
|---|---|---|---|
| On-premise | Servers in your own data center | Everything, including hardware and access | You buy and run the GPUs and staff the operations |
| Thai-hosted private cloud | Dedicated capacity at a provider in Thailand | Data location, access, network | Less hardware to run, still depends on provider capacity |
| Thai region of a global cloud | A public cloud region located in Thailand | Data location by setting, shared platform | Check which services are actually available in that region |
| Foreign region of a global cloud | A public cloud region in another country | Contract and configuration | Widest choice of models; cross-border transfer rules apply |
| Model API from a foreign provider | Prompts sent to the provider's endpoint | Only what you choose to send | Fastest to start; retention and location depend on the provider's terms |
No row is "the compliant one". A hospital with patient records, a bank with customer financial data and a retailer with product questions have different risk, so they will land in different places.
The data flows people forget
Teams usually map where the model runs and miss everything around it. Before you decide on a location, trace these flows.
- Prompts and responses. If a user pastes a customer's details into a chat, that text goes wherever the model runs.
- Embeddings and vector stores. Embeddings of personal documents are derived personal data. Treat the vector database like the source documents.
- Logs and traces. Debug logs often capture full prompts. Decide how long they live and who can read them.
- Backups and replicas. A database in Thailand with a backup in another country is a cross-border copy.
- Support access. A vendor engineer who connects from abroad to fix a problem may be a transfer too. Check your contract.
- Training and fine-tuning data. Confirm in writing whether a provider uses your inputs to improve its own models.
- Third-party tools. OCR services, transcription tools and analytics plug-ins each have their own location.
Write this list down as a data-flow diagram. It becomes the core of your record of processing and your answer to any audit question.
A simple way to decide where AI data should live
Work through four questions in order. Stop at the first strong answer.
- How sensitive is the data? Health, financial, national ID and children's data call for the tightest options. Public product information does not.
- Does a regulator or sector rule add requirements? Banks, insurers, hospitals and government agencies often have sector rules or internal policies on top of the PDPA. Ask your compliance team, and check guidance from sector bodies such as the Bank of Thailand or the relevant ministry.
- Can you reduce the data first? Masking ID numbers, removing names before a prompt and sending only the relevant paragraph can move a use case into a lighter pattern.
- What does the use case need from the model? If a smaller open model that you can host in Thailand handles the task well on your own test set, the location question gets easier.
A common outcome is a split. Sensitive workloads run on in-country infrastructure. Low-risk workloads, such as drafting marketing copy from public information, use a foreign API with a clear contract. A model gateway in front of both lets you route each request by data class. For more on running models yourself, read Private LLM vs API Models: How to Decide.
What can go wrong
- Assuming a "Thailand region" covers everything. Some managed AI features in a local region may still call services elsewhere. Ask the provider to confirm where each component runs.
- Treating consent as a fix for everything. Consent has to be informed and specific. Employees and customers may not be free to refuse, and a vague notice does not cover an AI use you did not describe.
- Forgetting the people side. A secure server does not help if every employee can query every document. Match AI access to existing permissions.
- Sending real data to a free tool. Staff will paste data into public chatbots if you give them no approved alternative. A sanctioned internal tool is a control, not just a convenience.
- No exit plan. If you commit to one provider, check that you can move your data, indexes and prompts out later.
When data residency should not drive the decision
Sometimes location is the wrong first question. If the data is not personal and not confidential, such as public documentation, residency rules add little and a foreign API may be the sensible choice. If you cannot yet describe your use case, spend the time on that first. And if your real concern is quality in Thai, test models on your own examples before debating hosting.
How UPSPIRE approaches it
Our AI infrastructure and sovereign cloud practice designs AI on Thai-hosted GPU capacity, in a private cloud or in your own data center, so personal data does not need to leave Thailand when you decide it should not. This matters most for banks and financial services, hospitals and government agencies. We also build mixed designs, because one location rarely fits every workload. For a wider view of what a project involves, see our guide to AI automation in Thailand.
How to start
Pick one planned AI use case and draw its data flow on a single page: sources, processing steps, storage, logs, vendors and people with access. Mark each box with its country and data class. Then take that page to your legal or privacy officer and ask which boxes need action.
If you want a second pair of eyes on the diagram, book a consultation with our team. We will help you compare options, but your counsel makes the compliance call.
Frequently asked questions
Does the PDPA require AI data to stay in Thailand?
The PDPA does not generally require all personal data to stay in Thailand. It sets conditions for sending personal data abroad, such as destination safeguards or an applicable exception. Sector rules or internal policies may be stricter. Check current PDPC guidance and confirm with counsel.
Is a Thai region of a global cloud enough?
Not automatically. Confirm where every component runs, including managed AI features, backups, logs and support access. Some services in a local region may still call services elsewhere. Ask the provider to confirm in writing.
Are embeddings and vector databases personal data?
Embeddings made from personal documents can reveal or be linked back to personal information, so treat the vector store like the source documents: same access controls, location and retention rules.
Can we use a foreign AI API for some work?
Often yes, for low-risk content such as public information, with a clear contract on retention and training use. Keep sensitive data on in-country infrastructure and route requests by data class. Your legal team should approve the design.
Ready to put this to work?
Turn this playbook into a working system
From a first agent to a national platform, the same team plans it, builds it and runs it with you.
