Data protection for AI chat in Malaysia and Singapore: a practical checklist

An AI chat that collects a name, a phone number or a health question is collecting personal data, and Malaysia and Singapore expect the same basics: notice, a stated purpose, limited retention, security, control of vendors and overseas hosting, and reporting of serious breaches. This is orientation for operators, not legal advice. Have a qualified adviser review your set-up.

How current is this, and what does it cover?

Every rule below was checked on 22 October 2026 against primary sources. For Malaysia, those are the documents of the Personal Data Protection Department (JPDP) (opens in a new tab)^[9], including the Act 709 reprint (opens in a new tab)^[3] reprinted as at 15 September 2026. For Singapore, they are Singapore Statutes Online (opens in a new tab)^[10] (current version as at 21 October 2026) and the Personal Data Protection Commission (PDPC).

JPDP says Malaysia's Act covers personal data processed in commercial transactions. Whether either law reaches your business depends on where you are established and where data is collected or processed. Sector rules and other Southeast Asian countries are not covered. AI replies can be wrong. Keep a person involved in financial, legal, medical and other consequential decisions.

How do Malaysia and Singapore compare?

Which parts of Malaysia's amended Act are in force, and when?

The Personal Data Protection (Amendment) Act 2024 (opens in a new tab)^[1] took effect in three steps under a Gazette notice published on 24 December 2024 (opens in a new tab)^[2]. Administrative sections began on 1 January 2025. On 1 April 2025 came biometric data as sensitive personal data, a definition of "personal data breach", a direct Security Principle duty for processors, higher penalties for breaching the principles and a rewrite of the cross-border rule. On 1 June 2025 came the data protection officer duty, mandatory breach notification and a right to data portability.

Is Malaysian breach notification 72 hours or "as soon as practicable"?

Both appear, in different documents. Section 12B of the Act requires a controller with reason to believe a breach has occurred to notify the Commissioner "as soon as practicable", with no fixed hours. JPDP's Data Breach Notification Guideline (opens in a new tab)^[4], issued 25 February 2025, adds an outer limit of 72 hours from the occurrence of the breach.

Its worked examples start that clock when the controller is told of a loss, realises it sent data in error or confirms a compromise, or, for a processor's breach, when the processor tells the controller or the controller has clear evidence, whichever is earlier. A late notice needs written reasons and evidence.

Notice is required where a breach causes or is likely to cause significant harm, which the guideline says includes sensitive personal data, identity-fraud risk and more than 1,000 people affected. People facing significant harm must be told within seven days of the report to the Commissioner.

Who must appoint a data protection officer in Malaysia?

Section 12A says controllers and processors must appoint one or more data protection officers. JPDP's DPO Guideline (opens in a new tab)^[5], also issued 25 February 2025, applies that where processing involves more than 20,000 data subjects, sensitive personal data including financial information of more than 10,000, or regular and systematic monitoring. Its monitoring examples include a retail website that tracks searches and purchases to offer recommendations, so ask whether profiling visitors in chat could count.

What changed for processors and overseas transfers in Malaysia?

Since 1 April 2025, a processor is directly bound by the Security Principle, commits an offence if it breaches it and, under section 9(2), must give sufficient guarantees about its security measures. For transfers abroad, JPDP's Cross Border Personal Data Transfer Guideline (opens in a new tab)^[6], issued 29 April 2025, describes a transfer impact assessment whose findings stay valid for up to three years, and expects the transfer to be disclosed in your notice, a record of each recipient, country, data type and purpose, and security clauses in any processor contract.

Is there Malaysian guidance specific to AI?

JPDP's Automated Decision-Making and Profiling guideline (opens in a new tab)^[7], which carries a 2026 copyright notice, says Act 709 has no provisions specific to automated decisions and profiling, but its principles apply, and it suggests telling people about AI use in the privacy notice. If your Mind sorts people automatically, for example through automatic Stages, ask your adviser whether that counts as profiling and meets the guideline's threshold of legal or significant effects on a person. JPDP also opened Public Consultation Paper No. 1/2026 (opens in a new tab)^[8] on 5 October 2026, closing 23 October 2026, on a proposed AI and Personal Data Protection Framework. It is a draft and not in force.

What does Singapore add for AI chat?

Singapore's Act does not mention artificial intelligence, so the PDPC applies its general rules through guidance. The AI Recommendation and Decision Systems guidelines (opens in a new tab)^[14], issued 1 March 2024, encourage organisations to say what an AI feature does and which personal data it uses and why, and suggest layering the information, for example a pop-up that links to a fuller policy.

The Generative AI guidelines (opens in a new tab)^[15], issued 20 July 2026, say deployers bear primary responsibility for the systems they choose, should treat prompts, inputs and outputs as data they must safeguard, and should tell users which categories of personal data to enter. Both sets are advisory: the PDPC says they are not legally binding. A vendor processing data for you under a written contract carries only protection and retention duties and a duty to tell you about breaches, while you keep the same obligations as if you processed the data yourself (s4 of the PDPA (opens in a new tab)^[10]).

What should an AI chat checklist cover?

Work through these nine items before a Mind goes public, and again when you add a channel.

1. Notice at the first message. Malaysia expects written notice, in the national language and English, when you first ask for or collect data; Singapore expects purposes notified on or before collection unless deemed consent or another exception applies. Keep it short, link to a fuller policy and say the chat is AI (see AI disclosure and the EU AI Act). Check what your channel can show before the first reply; where it cannot, put the notice on the page that hosts the chat. On messaging channels the first message can arrive before any page does; see the WhatsApp AI agent guide.

2. A stated purpose, and a Mind kept to it. Both Acts tie collection to a purpose, and the PDPC's generative AI guidelines tell deployers to define the purpose and the amount of personal data it needs. In Plandora, Data Collections define the details your Mind can collect, tagged in the Playbook where it should ask, and Plandora's guidance is to ask only for details you will use.

3. What the Mind stores, and for how long. Plandora's Privacy Policy (effective 16 October 2026) lists messages, contact details, intent signals, tags, notes and interaction history. It says retention lasts as long as reasonably necessary and varies by data type and workspace configuration, with no fixed period. Set your own rule for conversations and contacts and review it on a schedule. Contacts can be exported, archived or deleted; archiving keeps history, so it is not deletion.

4. Deletion and correction requests. Neither Act, in the text read, creates a standalone right to erasure. Requests arrive as withdrawal of consent (Malaysia s38, Singapore s16), as correction requests (which in Malaysia include deletion) and through the retention rules. The PDPC's generative AI guidelines say access and correction duties cover personal data collected to deploy an AI system, and suggest agreeing where reasonable, such as data held in a retrieval-augmented database. Plandora's FAQs say to contact the business running a Mind first; for information Plandora controls, the address is privacy@plandora.com, and deletion can depend on identity checks and records Plandora must keep. Decide who receives requests, and confirm what deleting a contact covers before promising full erasure.

5. Vendors and processors. Ask each vendor in writing who its sub-processors are, whether chat data is retained, reused or used for model training, where it is stored and how fast it will tell you about a breach. Plandora's privacy policy puts instructions and notices with the business deploying a Mind, and says Plandora processes information to provide the service. Plandora's FAQs add that the public terms make no blanket promise covering every underlying AI provider or model, and point to sales for commitments on training, isolation, retention and residency.

6. Cross-border hosting. Find out where chat data is stored, processed and accessed. Malaysia expects you to disclose transfers in your notice and keep records; Singapore expects you to confirm the recipient is legally bound to a comparable standard. Plandora's public privacy policy and FAQs name no hosting regions or sub-processors, and say Enterprise customers can request a review of data residency before rollout (see Enterprise).

7. Children. The PDPC's children's guidelines (opens in a new tab)^[16], issued 28 March 2024, cover online services likely to be accessed by children: the PDPC considers that a child aged 13 to 17 may give valid consent when the notice is readily understandable, while consent for a child under 13 comes from a parent or guardian. Malaysia's Act makes a parent or guardian the "relevant person" for access and correction requests about a child under 18, and the draft AI framework would require consent from the authorised person for under-18s. Plandora's privacy policy says the service is not directed to children. If minors may use your chat, decide how before launch.

8. Sensitive data. Malaysia's sensitive personal data includes health, political opinions, religious beliefs, offences and, since 1 April 2025, biometric data, and needs explicit consent or another listed condition (s40). Singapore's Act has no separate category, but its breach regulations treat certain health and financial details, combined with an identifier, as causing significant harm. Plandora's FAQs advise keeping sensitive and regulated personal information out of the platform unless your agreement and workspace configuration expressly support it. Tell visitors not to share sensitive details, set Boundaries for the limits you want, and give a route to a person (see human handoff for AI conversations).

9. Staff access and breach readiness. Plandora's privacy policy says authorised organisation members may be able to review conversations and audience records according to their assigned access, so give each person the least the job needs. The PDPC's Key Concepts guidelines (opens in a new tab)^[17], revised 29 April 2026, list access controls and need-to-know restriction of access among their examples of security measures. Name who decides whether an incident is a breach, who reports it and who acts as data protection officer or contact. The clocks are short: three calendar days from assessment in Singapore, per the PDPC's breach reporting guide (opens in a new tab)^[18], and 72 hours under JPDP's guideline in Malaysia.