Security & Data Governance

PDPA and AI in Malaysia: What Companies Must Know (2026 Guide)

Can staff paste customer data into ChatGPT, Copilot or Claude? What the 2025 PDPA amendments, DPO, breach and cross-border rules mean for your AI policy.

  • By
  • Updated
  • 13 min read
  • Facts checked
Short answer

Yes, the Personal Data Protection Act 2010 applies whenever an AI tool processes personal data in a commercial transaction. Since the 2025 amendments, companies also face fines up to RM1,000,000, breach notification, data protection officer rules and a conditions-based test for sending data abroad. No AI-specific law is in force yet; a draft AI Governance Bill is pending.

On this page
  1. Does the PDPA apply when staff use AI tools?
  2. What are the seven PDPA principles, and how do they apply to AI?
  3. What changed with the 2025 amendments?
  4. Do we need a data protection officer?
  5. How can I make sure my chatbot conversations meet data protection standards?
  6. What rights do people have over data that AI has processed?
  7. What happens if an AI tool causes a data breach?
  8. Can we send personal data to AI services hosted abroad?
  9. Are the free and paid versions of AI tools treated differently?
  10. What should be in a company AI policy?
  11. What is shadow AI, and why does it matter for the PDPA?
  12. What does Malaysia's AI regulation look like in 2026?
  13. How do you vet an AI vendor?
  14. What is a sensible 90-day plan for an SME?
  15. What do AI mistakes with personal data look like in practice?
  16. What do we do at winsym.ai?
  17. Sources
  18. Frequently asked questions

If your team uses ChatGPT, Microsoft Copilot, Claude or Gemini with anything that identifies a customer, patient, applicant or colleague, the Personal Data Protection Act 2010 (PDPA) already applies. Malaysia has no AI-specific law in force yet, so the PDPA is the rule that bites today. This guide explains how it works with AI, what changed in 2025, what to put in writing, and what is coming.

Checked in November 2026 against the Commissioner's own guideline texts on pdp.gov.my (Act 709, DPO, data breach notification and cross-border transfer guidelines) and, for the amendment dates, law-firm summaries of the Gazette notice. General information, not legal advice.

Does the PDPA apply when staff use AI tools?

Yes. The PDPA covers personal data processed in commercial transactions, and it regulates the data, not the technology. Pasting a customer email, a CV, a patient note or a spreadsheet of leads into an AI tool is "processing". The Act does not apply to the Federal and State Governments (section 3; commentators disagree on whether government agencies and government-owned companies count), but it applies to private companies, clinics, schools and agencies.

Personal data means information that relates to a living person who can be identified from it, such as name, IC number, phone, email, address, photo, voice or an account ID that can be linked back to someone. Sensitive personal data (health, political opinions, religious beliefs, criminal records, and since the 2025 amendments, biometric data) needs explicit consent.

A company that decides why and how personal data is processed is the data controller. A company that processes it on the controller's behalf is the data processor. An AI vendor is usually your processor, and since 1 April 2025 processors are directly bound by the security principle, so both sides carry exposure.

What are the seven PDPA principles, and how do they apply to AI?

The Act sets seven principles. Each one has an everyday AI meaning.

Principle What it asks What it means for AI use
General Consent, or another lawful basis, before processing Did customers agree to their data being handled by AI vendors? Check your privacy notice.
Notice and Choice Tell people what you collect, why, and to whom you disclose Name AI tools or categories of tool in the privacy notice.
Disclosure Do not disclose data for purposes beyond what you told people Sending a client list to a public chatbot for convenience is a disclosure.
Security Protect data from loss, misuse and unauthorised access Approved accounts, access control, logs, no personal AI accounts for work.
Retention Keep data no longer than needed Check how long the AI vendor stores prompts and files.
Data Integrity Keep data accurate and up to date A model that invents facts about a person can produce wrong records. Humans must check.
Access People can see and correct their data You must be able to find and fix a person's data if it sits in AI-processed records.

The list comes from the Act itself. The point of the table is practical: for each AI tool, you should be able to answer all seven rows.

What changed with the 2025 amendments?

The Personal Data Protection (Amendment) Act 2024 came into force in three phases, according to the Gazette notice of 24 December 2024 as summarised by law firms.

Date What took effect
1 January 2025 Notices and documents can be served electronically
1 April 2025 "Data user" becomes "data controller"; biometric data counts as sensitive; "personal data breach" is defined; processors are directly liable for security; penalty for breaching the principles rises to up to RM1,000,000 and/or three years' imprisonment (previously RM300,000 and two years); the cross-border whitelist is removed
1 June 2025 Mandatory data protection officer (DPO); mandatory breach notification; right to data portability

For a detailed walk-through of what each change means for AI use, read PDPA amendments: what changes for AI use.

Do we need a data protection officer?

Probably, if you handle personal data at scale. The Commissioner's Guidelines on the Appointment of a Data Protection Officer (issued 25 February 2025, paragraph 4.2) say you must appoint a DPO if you process personal data of more than 20,000 individuals, sensitive personal data (including financial data) of more than 10,000 individuals, or carry out regular and systematic monitoring. The DPO can be an employee or an outsourced provider, does not have to live in Malaysia but must be easily contactable, and must be notified to the Commissioner within 21 days. The Act expects a DPO to be a named, contactable person, so a small company below the thresholds can still appoint one voluntarily. Many SMEs fall below the thresholds. Our short explainer is Do you need a data protection officer in Malaysia?.

How can I make sure my chatbot conversations meet data protection standards?

Treat every chatbot conversation as processing of personal data, because customers type names, phone numbers, order details and sometimes health or financial facts. The Act's seven principles then translate into a short checklist.

PDPA principle What to do for a chatbot
General and Notice and Choice Tell users at the start that they are talking to an AI, what is collected and why, and link the privacy notice
Disclosure Do not pass chats to other tools or vendors beyond what the notice says
Security Restrict who can read transcripts, turn off training on your data where the vendor allows it, log access
Retention Set a deletion period for transcripts and stick to it
Data Integrity Let the bot hand over to a person when it is unsure about identity or account details
Access Be able to find and export one person's chats on request

Two extra points. If the chatbot vendor stores chats abroad, the section 129 conditions above apply. And a vendor that processes chats for you is a processor: the amended Act makes processors directly liable for security, so ask for the contract terms in writing. Our vendor checklist lists the questions, and the AI integration guide covers the build side.

What rights do people have over data that AI has processed?

Individuals can ask to see the personal data you hold about them and to have it corrected (the Access principle), and since 1 June 2025 they have a right to data portability, which lets them ask for their data to be moved in a usable form. They can also withdraw consent for processing that relied on it. For AI this has a practical meaning: you need to know where a person's data sits. If customer records have been copied into an AI tool, a shared drive that an assistant indexes, or a vendor's logs, an access or correction request reaches those copies too.

Three habits make this manageable. Keep AI inputs traceable (a log of what was uploaded, when, by whom). Prefer tools that let you delete uploaded files and conversation history on request. And avoid building permanent copies of personal data inside systems you cannot search, such as free-form chat histories held in personal accounts.

What happens if an AI tool causes a data breach?

You may have to notify the Commissioner within 72 hours. According to the Commissioner's guidelines of 25 February 2025 (sections 12A and 12B, effective 1 June 2025), as summarised by DLA Piper and CMS, two triggers apply. A breach that causes or is likely to cause significant harm (for example financial loss, misuse for illegal purposes, exposure of sensitive data) must be reported to the Commissioner and to the affected individuals. The Commissioner's Data Breach Notification guideline (paragraph 6.1 for the 72 hours) sets the same tests. A breach affecting more than 1,000 individuals must be reported to the Commissioner even without significant harm. The Commissioner must be told as soon as practicable and within 72 hours of the breach occurring or the controller being informed; individuals without unnecessary delay and within seven days after that. If you miss the 72 hours, a written explanation with evidence is needed, and missing details can be supplied in phases within 30 days. Exceptions exist, for example where data was encrypted so that it is unintelligible to others.

An AI-related breach is not exotic. A sales rep pastes a customer list into a free chatbot, a shared link to an AI conversation goes public, or an AI agent with too many permissions emails the wrong file. Our piece on prompt injection and data leaks shows how a hostile document can trigger a leak without any staff error.

Can we send personal data to AI services hosted abroad?

Yes, under conditions. Section 129 of the amended PDPA replaced the Minister's whitelist of countries with a conditions-based test. You need to meet at least one condition, among them: the destination has a law substantially similar to the PDPA or ensures an adequate level of protection; the person gave explicit consent; the transfer is necessary for a contract with the person; or you took reasonable precautions and exercised due diligence so that the data will not be processed in a way that would breach the PDPA.

The Commissioner's Guidelines on Cross Border Personal Data Transfer (29 April 2025) suggest a Transfer Impact Assessment, contractual clauses (the ASEAN Model Contractual Clauses and EU Standard Contractual Clauses are named as examples) or binding corporate rules. An assessment can be relied on for up to three years unless the destination's law changes. We cover it in Sending data to US AI tools and in Data residency and the Malaysian sovereign cloud. If you also handle data of people in the EU, GDPR vs PDPA Malaysia shows where the two rulebooks differ.

Are the free and paid versions of AI tools treated differently?

Yes, and this is where most risk sits. Vendor terms differ between consumer plans and business plans. Anthropic states that its Team, Enterprise and API products are not used for model training by default (privacy.claude.com, page dated 18 August 2026). OpenAI states that business offerings including ChatGPT Enterprise, Business and the API are not used for training by default, and API data may be retained up to 30 days for abuse monitoring. Consumer plans have their own terms that you must read and that can change.

A business plan is not a PDPA compliance certificate. It reduces one risk (training on your data) but you still need a lawful basis, a notice to your customers, a transfer basis if the data goes abroad, and a vendor contract. See Can I put customer data into ChatGPT?.

What should be in a company AI policy?

A short written policy is the quickest practical control. At minimum:

  1. A list of approved tools and the accounts to use (company accounts only).
  2. Data classes that must never go into AI tools, or only into specific tools.
  3. A rule that a human checks any AI output that touches a customer, employee or decision about a person.
  4. How to report a mistake or suspected leak, and to whom.
  5. A named owner who updates the list.

We give a fill-in version in AI policy for a company (template), what separates a policy that works from a PDF nobody reads in What makes a good AI policy, and an employee-facing version in Employee AI usage policy template.

What is shadow AI, and why does it matter for the PDPA?

Shadow AI is the use of AI tools at work without the company's knowledge or approval, such as a personal ChatGPT account used to summarise client emails. It matters because you cannot tell customers, assess transfers or investigate a breach for tools you did not know about. The fix is rarely a ban; it is an approved alternative plus a clear rule. See Shadow AI at work.

What does Malaysia's AI regulation look like in 2026?

Mostly voluntary guidance with a bill on the way. The Ministry of Science, Technology and Innovation published the non-binding National Guidelines on AI Governance and Ethics in September 2024. The National AI Office published a public consultation paper on a proposed AI Governance Bill on 10 July 2026, with feedback closing on 31 July 2026. Per the consultation paper as summarised by Rahmat Lim & Partners, the bill would:

  • cover developers and deployers of AI systems used in Malaysia, with exclusions for personal use and national defence or security;
  • rest on five principles (human dignity, transparency and explainability, accountability, safety and security, responsible data governance);
  • sort systems into three risk tiers (unacceptable, high, low);
  • create a Central AI Authority and require reporting of AI incidents.

The Vibes reported on 21 July 2026 that the bill would be tabled after ministry, agency and Cabinet discussions, with no date; a 10 September 2026 report, which we saw only as a search summary, said the Digital Minister aimed for tabling by Q1 2027. Until it passes, nothing in it is law. A fuller overview is in Malaysia AI regulation overview 2026, and the management side is in What is AI governance?.

How do you vet an AI vendor?

Ask five things in writing before any personal data goes in: where data is stored and processed, whether prompts are used for training, how long data is retained, which sub-processors are involved, and what happens on a breach. Look for independent assurance such as ISO/IEC 27001 or a SOC 2 report. Put the answers in the contract. The checklist is in AI vendor due diligence checklist. If client confidentiality is your main concern (law, accounting, consulting), read Confidentiality clauses and client data in AI tools.

What is a sensible 90-day plan for an SME?

This is our suggested order. It is a practical sequence, not a legal requirement.

Weeks Action Output
1-2 Ask staff which AI tools they use and for what A tool and use-case list
2-4 Classify data: public, internal, personal, sensitive A one-page data classification
3-6 Write the AI policy and pick approved tools Policy signed off by management
4-8 Update privacy notice and vendor contracts Notice and contract addenda
6-10 Train staff on the policy Attendance record
8-12 Test breach response with a simple drill Contact tree and 72-hour checklist

Staff training can be funded by the HRD Corp levy if the course is delivered by an HRD Corp Registered Training Provider and the grant is approved before training starts. See the guide on HRD Corp claimable AI training in Malaysia.

What do AI mistakes with personal data look like in practice?

Three hypothetical cases, not real clients:

  • Clinic in Petaling Jaya: a receptionist pastes appointment notes with diagnoses into a free chatbot to draft reminders. Health data is sensitive personal data, so explicit consent and a lawful route are needed. The fix is a clinic-approved tool and templates with no clinical detail.
  • Property agency in KL: a negotiator uploads a spreadsheet of 3,000 buyer leads to an AI tool to score them. A breach affecting more than 1,000 individuals must be reported to the Commissioner under the 2025 guidelines, so a leak here would trigger notification. The fix is to pseudonymise before upload and use a business plan.
  • Accounting firm in Penang: a partner uses AI to summarise client payroll files. Client confidentiality clauses, not just the PDPA, apply. The fix is a written client-data rule and vendor terms that rule out training.

What do we do at winsym.ai?

We build and train; we do not give legal advice. In our projects we document data flows, classify data, work with approved or private deployments where the data demands it, and tell you plainly when a decision belongs to your lawyer or DPO. Our older note on the engineering side is PDPA and AI integration for Malaysian SMEs; the data-boundary view is in LLM data boundaries and the sovereign stack.

If you want your team to use AI safely and with confidence, we run staff sessions on practical rules for AI at work (see our courses and coaching) and an AI opportunity audit that maps your data flows first. Book a free 30-minute call and bring your questions and, if you have one, your DPO or counsel.

To bring every employee onto the same PDPA-safe habits, many companies run Generative AI at Work in-house, which ends with a usage policy draft.

Sources

  • Personal Data Protection Commissioner, Act 709 and guidelines: DPO guideline, data breach notification guideline, cross-border transfer guideline, all read in full in November 2026. Laws of Malaysia: lom.agc.gov.my. Personal Data Protection Act 2010 (Act 709) and Personal Data Protection (Amendment) Act 2024 (Act A1727), Laws of Malaysia.
  • DFDL, "Malaysia: Implementation of the Personal Data Protection (Amendment) Act 2024", dfdl.com, summarising Gazette notice of 24 December 2024, retrieved 22 October 2026.
  • DLA Piper, "Malaysia: Guidelines issued on data breach notification and data protection officer appointment", privacymatters.dlapiper.com, March 2025, retrieved 22 October 2026.
  • Mayer Brown, "From legislative reform to practical guidance: key amendments to Malaysia's PDPA and the launch of cross border transfer guidelines", July 2025, retrieved 22 October 2026.
  • Rahmat Lim & Partners, "National AI Office issues public consultation paper on proposed AI Governance Bill", 29 July 2026, retrieved 22 October 2026.
  • Anthropic, "Is my data used for model training?", privacy.claude.com, retrieved 22 October 2026. OpenAI business data privacy pages, openai.com, retrieved 22 October 2026 through search summaries.
  • The Vibes, 21 July 2026, on the timeline for the AI bill (read in full). Malay Mail, 10 September 2026 (search summary only; original page blocked).

General information, not legal advice. Have your lawyer or data protection officer confirm how the PDPA applies to your business.

Frequently asked questions

How can I make sure my chatbot conversations meet data protection standards?

Tell users they are talking to an AI and what you collect, limit who can read transcripts, set a deletion period, check where the vendor stores chats (section 129 conditions apply to overseas hosting) and be able to find one person's chats for an access request.

Do we need a data protection officer if we only run a chatbot?

Only if you cross the Commissioner's thresholds: personal data of more than 20,000 people, sensitive data of more than 10,000 people, or regular and systematic monitoring. Below that a DPO is optional but still useful.

Does the PDPA apply to ChatGPT, Copilot and other AI tools?

Yes, if you enter personal data about customers, staff or other people in the course of commercial activity. The PDPA regulates the data and what you do with it, whatever the tool. Typing a customer's name, phone number or medical detail into an AI tool is processing under the Act.

Is there an AI law in Malaysia?

Not yet in force. The 2024 National Guidelines on AI Governance and Ethics are voluntary. The National AI Office ran a consultation on a proposed AI Governance Bill in July 2026, and the Digital Minister said in September 2026 that a draft is complete. Tabling in Parliament is expected in late 2026 or early 2027.

What is the penalty for breaching the PDPA principles after the amendment?

The maximum penalty for breaching the personal data protection principles is a fine of up to RM1,000,000 or imprisonment for up to three years, or both, up from RM300,000 and two years. Other offences carry their own penalties in the Act.

Can a Malaysian company send personal data to AI services hosted overseas?

Yes, if one of the conditions in section 129 of the amended PDPA is met, for example the destination has a substantially similar law, the person has given explicit consent, or you took reasonable precautions and exercised due diligence. Many AI tools are hosted abroad, so check this before staff use them with personal data.

Should a company write an AI policy even if the law does not require one?

Yes. The PDPA already requires you to protect personal data, and staff are using AI tools whether or not you have approved them. A short written policy that says which tools are allowed and which data is off limits is the cheapest control you can add.