Does the PDPA apply to AI tools?
Yes. PDPC’s guidance says the PDPA applies to all collection and use of personal data by an organisation, including to develop, test, monitor and deploy AI systems. Using AI does not create a separate set of rules. The same obligations cover personal data in prompts, uploaded files and AI outputs.
PDPC has published two advisory guidelines on AI:
- Use of Personal Data in AI Recommendation and Decision Systems (issued 1 March 2024): when personal data can be used to develop AI systems, and how to be open about how those systems use it.
- Use of Personal Data in Generative AI (issued 20 July 2026): training data, who is responsible for what across the generative AI lifecycle, and requests from individuals about their data.
Both are advisory. They are not legally binding, and the Act prevails where they differ.
Most businesses are “system deployers”.
The 2026 guidelines name three roles. Model providers develop and make available generative AI models. System providers build systems and applications on them. System deployers use those systems, or let others use them, under their authority. A business that gives its staff an AI assistant, or puts a chatbot on its website, is usually a system deployer. One organisation can hold different roles in different projects.
PDPC says system deployers “bear primary responsibility for ensuring that the Generative AI Systems they have chosen to use can meet their obligations under the PDPA”. In practice, the guidelines ask deployers to:
- Get the facts before choosing a system. Have enough information about the provider’s safeguards to assess them. PDPC’s examples of what providers can share: access controls, where data is stored and processed, retention, input and output filters, data-leakage testing, and incident and breach procedures.
- Define the purpose. Specify what the system is for and how much personal data that needs, then limit its access to match. PDPC’s example is a records-triage system that cannot reach billing systems or contact lists.
- Protect the new data. Prompts, uploaded inputs, generated outputs and agent or tool activity can all contain personal data. Make someone responsible for each source, and tell users which personal data may go into the system.
- Write it down. Keep written policies on your safeguards. PDPC encourages making them available before anyone asks, for example on your website.
- Review regularly, especially for agents. Agents plan and act across several steps with tools, which can raise the data protection risk. PDPC’s example adds role-based limits on an agent’s file and network access, tags for more sensitive data, and escalation to a person for complex or high-risk tasks.
The PDPA obligations that come up most.
A plain-language summary, with the question to ask about each AI workflow. Section numbers refer to the PDPA.
| Obligation | What it requires, in short | Ask about your AI workflow |
|---|---|---|
| Consent, purpose and notification (s 13, 18, 20) | Use personal data only for purposes a reasonable person would consider appropriate and that people were told about, with consent unless an exception applies. | Is this AI use within what we told customers and staff? |
| Protection (s 24) | Make reasonable security arrangements against unauthorised access, use, disclosure, copying, modification or disposal, and against losing storage devices. | Who can read the prompts, logs, document stores and backups? |
| Retention limitation (s 25) | Stop keeping personal data, or remove what links it to individuals, once it no longer serves its purpose and is not needed for legal or business reasons. | When are chat logs, uploads and document indexes deleted? |
| Transfer limitation (s 26) | Transfer personal data outside Singapore only under the prescribed requirements, which give it protection comparable to the PDPA. | Where does each service process and store our data? |
| Accuracy (s 23) | Make a reasonable effort to keep personal data accurate and complete if it is likely to be used for a decision about the person or disclosed to another organisation. | Does a person check AI output before it is used to decide something about someone? |
| Access and correction (s 21, 22) | Respond to people’s requests to see or correct their personal data, including data a provider holds for you. | Could we find one person’s data in our AI system, such as in a document index? |
| Data intermediaries (s 4(3)) | You have the same obligation for personal data a provider processes on your behalf as if you processed it yourself. | Which providers process personal data for us, and on what terms? |
| Breach notification (s 26B–26D) | If a breach is likely to cause significant harm or is of significant scale, notify PDPC no later than 3 calendar days after assessing it as notifiable, and in significant-harm cases generally the affected individuals too. | Would we notice personal data leaking through an AI tool, and who would assess it? |
Training on customer data needs its own check.
Fine-tuning a model on your records is a different use from asking a model questions about them. Before you use customer data to train or fine-tune a generative AI model, check whether an exception such as business improvement applies, or whether you need consent.
If you rely on consent, PDPC’s 2026 guidelines say a general notice such as “new product development” is not enough for large-scale training or fine-tuning. The notice must say the data will be used for AI, so people can understand what data is used, how, and what the model does. PDPC also encourages using only the data attributes the model needs, and pseudonymised or anonymised data where possible.
Does local AI make you compliant?
No setup is compliant on its own. Running an open-weight model on hardware you control changes where the data goes. It does not change whose obligations they are.
When the model runs on your own workstation or server, the prompts and documents it answers from do not need to go to a model provider. PDPC notes that open-weight model providers without access to the personal data in systems built on their models are not subject to PDPA obligations for that data. In that setup, your organisation holds the personal data and the obligations that come with it.
That can make two questions easier to answer: which providers process personal data for you, and whether it leaves Singapore. It does so only if the interface, document index, connected tools, remote support and backups also stay under your control. Local AI does not secure the machine, set retention for logs, check outputs or answer access requests for you. Trace the whole path first, as described in “Local” is a data-flow question.
If we deploy a local setup for you, the handover documentation lists the interface, model and connections in it, so you can check them against your own policy.
A checklist before you start.
- List the AI tools in use, including staff’s own accounts, browser extensions and AI features inside software you already use.
- Write down each use: its purpose, the personal data it needs and who uses it. Remove data the task does not need.
- Map the data flow: interface, model, document index, tools, logs and backups. Note where each runs, what it stores and whether anything leaves Singapore.
- Ask each provider about access controls, where data is stored and processed, retention, whether it uses your prompts or files to develop its models, and how it handles incidents. Check that the contract matches.
- Set rules for staff: which personal data may go into prompts and uploads, and who can read prompt logs.
- Set retention for chat logs, uploads and document indexes, and make someone responsible for deleting them on schedule.
- Keep a person in the loop for decisions about people and for significant steps an agent takes. Limit what agents can access.
- Update your data protection policy to cover AI use, and agree who assesses a suspected breach, so you can notify PDPC within 3 calendar days of assessing it as notifiable.
For agents, IMDA’s Model AI Governance Framework for Agentic AI groups good practice into four areas: assess and bound the risks up front, make humans meaningfully accountable, implement technical controls and processes, and enable end users to act responsibly.
Next: map your own setup.
Start with the data-flow and evaluation steps in Local AI vs cloud AI. If you want help scoping a local deployment, see our deployment process.
Sources.
- Personal Data Protection Act 2012, Singapore Statutes Online (sections 4, 13–26 and 26B–26D)
- PDPC: Advisory Guidelines on Use of Personal Data in Generative AI (issued 20 July 2026)
- PDPC: Advisory Guidelines on Use of Personal Data in AI Recommendation and Decision Systems (issued 1 March 2024)
- PDPC: Advisory Guidelines on Key Concepts in the PDPA
- IMDA: Model AI Governance Framework for Agentic AI (PDF)
This guide is general information, not legal advice. PDPC’s guidelines are advisory, and the Act prevails. For a specific situation, read PDPC’s guidance or ask a qualified adviser.