PDPA for AI Chatbots in Singapore: A Pre-Launch Checklist

Before a customer-facing AI chatbot goes live in Singapore, settle five PDPA questions: what personal data it collects, what you tell users, whether chat logs will train or improve any model, what role each vendor plays under the Act, and how long you keep transcripts. PDPC's July 2026 guidelines on generative AI now spell out what good practice looks like.

At a glance

This article is general information, not legal advice. For a specific deployment, speak to your data protection officer or a Singapore lawyer.

Which rules apply to a chatbot

The PDPA is the law. PDPC lists eleven obligations on its Data Protection Obligations page, and a chatbot touches almost all of them: consent, notification, purpose limitation, accuracy, protection, retention limitation, transfer limitation, access and correction, data breach notification and accountability.

The advisory guidelines explain how PDPC reads those obligations for AI. They are not legally binding and do not override the Act, but they are the clearest statement of what the regulator expects. The Advisory Guidelines on Use of Personal Data in Generative AI, issued on 20 July 2026, are the ones written for large language model chatbots. They build on the 2024 guidelines for AI recommendation and decision systems, which still matter if your bot recommends products or decides eligibility.

Know your role

The generative AI guidelines name three roles:

• Model Providers develop and make available the underlying models.
• System Providers develop and make available the chatbot system or application.
• System Deployers use, or enable the use of, a generative AI system under their authority.

If you are an SME launching a bot on a vendor platform, you are the System Deployer. PDPC says deployers "bear primary responsibility" for making sure the system meets PDPA obligations. A vendor that processes chat data on your behalf is usually a data intermediary, and that does not move your accountability to them.

Part 1: Settle the data before you build

1. Write down the purpose and the minimum data

Purpose limitation means using personal data only for purposes a reasonable person would consider appropriate. The guidelines ask deployers to specify the purposes and the amount of data needed. A delivery-tracking bot needs an order number and perhaps a mobile number. It does not need a date of birth. If you think you need an NRIC number, read PDPC's guidance on national identification numbers before you design that step.

2. Treat prompts and outputs as new data sources

Users type things you never asked for: medical details, account numbers, a colleague's name. PDPC says deployers should track, and assign responsibility for, "end-user prompts, inputs and generated outputs" as new data sources, and safeguard them. Its HR chatbot example shows good practice: one team owns prompts and outputs, staff get rules on what personal data may go into prompts, prompts are scanned for common identifiers, and access to prompt logs is limited.

For a public bot, add a short line at the start ("Please do not share your NRIC, card or bank details here") and mask obvious identifiers before logs are stored.

3. Decide now whether transcripts will train any model

This is the question most teams leave too late. If chat data will be used to train or fine-tune a model, and no exception applies, PDPC says a general notice such as "new product development" is not enough. You must give an AI-specific notification, and PDPC encourages it to cover, as far as practicable:

• the model functions that need personal data
• the types of personal data used
• how the data will train or fine-tune the model
• how people can decline or withdraw consent

You may layer the notice so the key points come first, and you cannot make training consent a condition of service beyond what is reasonable. If you plan to rely on the business improvement exception instead, the 2024 guidelines set two conditions: the purpose cannot reasonably be achieved without identifiable data, and a reasonable person would consider the use appropriate. Record that assessment.

Then ask your vendors the same question. A model provider that collects end-user prompts for its own model development is an organisation under the PDPA for that data, with every obligation that brings.

Part 2: Settle what users see

4. Notify at first use, and publish a chatbot info card

The notification obligation requires you to tell people your purposes on or before collection. IMDA's voluntary Transparency Guidelines for Generative AI Chatbots add a practical format: a chatbot info card covering four areas, with at least one substantive disclosure in each.

1. What the chatbot can be used for, including limits and when to seek human help.
2. How reliable and safe it is, with the safeguards and remaining risks.
3. How user data is used and protected, including training use, retention and user controls.
4. How users can report issues and what happens next.

IMDA encourages a high-level safety statement and a clear link to the card at first use, with the card reachable from inside the chat afterwards. The guidelines cover external-facing chatbots; tools used only by your own staff are outside their scope. Make sure the data section of the card says the same thing as your privacy notice. Mismatches are easy to spot and hard to explain.

5. Offer a human and a complaints route

Accountability under the PDPA includes a complaints process and a data protection officer whose business contact details are public. Put both on the info card, and give the bot a clear handover: "I can't help with that, here is how to reach our team." Test the handover as carefully as the answers.

6. Check marketing follow-ups

Singapore's Do Not Call Registry lets people opt out of telemarketing messages to their Singapore telephone numbers. If the bot collects a mobile number and you later want to send promotions to it, check the Do Not Call rules first, and keep service messages separate from marketing consent.

Part 3: Settle vendors, security and retention

7. Classify every vendor and ask for evidence

PDPC says deployers must make sure they have enough information about upstream safeguards when they procure a system. Its example of a model provider's safeguards lists role-based access, encryption and a retention policy that states when inference data is deleted.

8. Test for prompt injection and data leakage

The guidelines name prompt injection as a risk, and encourage system providers to share their input and output filters, test results such as the likelihood of data leakage, and incident response procedures with deployers. Ask for them. Give the bot access only to the data its purpose needs: if it answers order questions, it should not be able to query the whole customer table. Our explainer on prompt injection shows the attack patterns to test.

9. Decide where the data may go

The transfer limitation obligation allows overseas transfers only with protection comparable to the PDPA. Check each platform's settings. In Microsoft Copilot Studio, for example, an environment in the Singapore geography stores data in the Southeast Asia region, but if an admin allows data movement across regions, prompts and outputs for generative features may be processed in region or in the United States, and Bing search is processed in the United States. Microsoft also notes that some Copilot Studio models can process data outside your region even when cross-region movement is restricted. Make these choices deliberately and record them.

10. Set a retention period for transcripts

The retention limitation obligation means you stop keeping personal data once it no longer serves a business or legal purpose. Pick a period for transcripts, configure deletion in the platform, and state it on the info card.

11. Prepare for access requests and breaches

PDPC confirms that access and correction rights still apply when personal data has been used to develop or deploy generative AI. Make sure you can find one person's transcripts quickly.

For breaches, PDPC's breach management guide asks you to assess whether a breach is notifiable within 30 calendar days and, if it is, notify PDPC no later than three calendar days after that assessment. A breach involving 500 or more individuals counts as significant scale. Data intermediaries must tell you about breaches they detect, so put a prompt-notice clause in every vendor contract.

12. Write it down

PDPC encourages deployers to develop clear written policies and make them available, for example on their website, and says they should regularly review whether their safeguards are still sufficient. That matters most if your bot takes actions, such as booking or refunding, and the guidelines point to IMDA's framework for agentic AI. Our guide to Singapore's Model AI Governance Framework for Generative AI covers the wider governance picture.

The one-page pre-launch checklist

How CertTulen Academy can help

CertTulen Academy runs live, MCT-led training with hands-on labs. If your team is building its first bot, AI Chatbot Creation (CT-AI-702) covers design, testing and safe handover, and Create Agents in Microsoft Copilot Studio (PL-7008) covers building on Microsoft's platform, including knowledge sources and data settings. We also run private cohorts for companies around their own use case. Contact us or WhatsApp +60 16-241 8686 for a batch that suits your time zone.