Home / Blog / AI

PDPA and AI: a practical checklist for Malaysian teams

Note: This article is general guidance from an engineering team, not legal advice. For decisions about your organisation's obligations, speak to your legal adviser or data protection officer.

Most useful AI touches personal data at some point: a customer's message to a chatbot, a face in an eKYC check, a purchase history behind a recommendation. In Malaysia that brings the Personal Data Protection Act 2010 (PDPA) into every AI project. It was significantly strengthened by amendments passed in 2024, which came into force in stages during 2025.

The good news is that building with privacy in mind from the start costs far less than retrofitting it. Here's the checklist we use on our own products and client projects.

What changed with the amendments

At a high level, the amendments that matter most for AI teams include:

For AI, the cross-border and processor changes matter most. Many AI services are hosted outside Malaysia, and the AI vendor is often your data processor.

The checklist

Before you collect or reuse data

  1. Purpose: Was the data collected for a purpose compatible with training or running this AI system? If not, do you need fresh notice or consent?
  2. Minimisation: Does the model actually need names, IC numbers or phone numbers, or can you use pseudonymised IDs?
  3. Sensitive data: Health information and similar categories need explicit consent and extra care. This applies to wellness products like HAI too.

Choosing models and vendors

  1. Where is data processed? Know which countries prompts, files and logs travel to, and check that the transfer is permitted.
  2. Is your data used for training? Check the provider's terms. Enterprise tiers often exclude your data from model training by default.
  3. Processor agreements: Make sure your contracts with AI vendors cover security, breach reporting and deletion.

Building the system

  1. Mask personal data in prompts and logs where possible, and set retention periods for conversation logs.
  2. Respect access rights: an assistant must not show one customer another customer's data, so retrieval must apply the same permissions as your apps.
  3. Explainability: for decisions that affect people (approvals, pricing, eligibility), record which inputs drove the outcome and keep a person in the loop.
  4. Security testing: include prompt-injection and data-leakage tests alongside your usual penetration testing.

Operating it

  1. Breach readiness: make sure your incident response plan covers AI components, including vendor-side incidents.
  2. Access and correction requests: know how you'd find and export or correct an individual's data across training sets, vector stores and logs.
  3. Review regularly: models, vendors and regulations change. Put a PDPA review into each major release.

Privacy as a product feature

We treat consent screens, data retention and access controls as features to design well, not legal boilerplate. Users are more willing to share data with a product they trust, and as we wrote in our piece on proprietary data, good data is what makes AI hard to copy.

Planning an AI project that uses customer data?

Our team can review your design against this checklist and propose a privacy-first architecture before you build. Get in touch →