Malaysia's AI Governance Bill: What Businesses Need to Know

Malaysia's proposed AI Governance Bill would set national rules for how AI systems are built and used in the country. It takes a risk-based approach: possible prohibitions for unacceptable risk, stronger safeguards for high-risk AI, and baseline principles for low-risk systems. The proposed scope is broad enough to matter to businesses that use AI as well as those that build it.

The National AI Office (NAIO) published a summary of the proposed Bill for public consultation in July 2026. This article explains those consultation proposals, which may differ from the final legislation.

Status as of 2 November 2026: the Bill has not been enacted. On 29 October, the Digital Minister said it had been completed and would be submitted to Cabinet, with tabling in Parliament planned for early 2027. That is a planned tabling date, not an enforcement date. Read the reported parliamentary update.

Existing laws can still apply to AI use. The absence of an enacted AI Governance Act does not mean businesses have no legal responsibilities.

Full disclosure: my company, Kognosys, builds enterprise AI, so these proposals are relevant to our work too. The policy summary below is separate from my practical recommendations and product priorities.

What the Malaysia AI Governance Bill covers

The proposal covers the AI system lifecycle, from inception and development through operation, monitoring and retirement. ISO/IEC 5338:2023 provides a related framework of AI system lifecycle processes.

Under the consultation proposal, an AI system would fall within scope if it is:

• placed on the market or put into service in Malaysia, or
• designed, developed or used in Malaysia, or
• used by a Deployer established in Malaysia, no matter where the system is hosted.

That last point matters. Hosting AI in Singapore or the US would not, by itself, remove a Malaysian Deployer from the proposed scope.

The proposal includes exemptions for personal use (an individual using AI for personal, family or household matters) and systems used only for national defence or security.

Developers and Deployers

The consultation identifies two roles. An organisation could perform both.

• Developer: anyone who materially shapes what the AI can do, how well it works, what limits it has, or what risks it creates.
• Deployer: anyone who puts the AI to work in the real world. They may not have built it, but they decide where, how and under what conditions it is used, and they can monitor, suspend or withdraw it.

Example: a software firm builds an AI that screens job applications (Developer). A Penang logistics company switches it on for its hiring (Deployer).

Who would enforce it

The summary proposes one Central AI Authority as the main national body. It could appoint Sectoral Leads from existing institutions with suitable legal authority, expertise and governance capacity to support sector-specific implementation.

The Central AI Authority would have three core jobs:

1. AI safety: keep a risk framework, oversee assessments, support testing and run incident reporting.
2. Investigation and enforcement: find out what happened when AI incidents occur, and who was involved.
3. Enablement: give guidance, templates and training to the public sector and regulated firms, and run AI sandboxes.

During the consultation period, NAIO transitioned into AI Malaysia Berhad, launched on 28 July 2026. That institutional change should not be taken as confirmation of who will exercise the proposed Authority's statutory powers.

The five AI governance principles

The proposal would require Developers and Deployers to have due regard to five governance principles:

1. Human dignity and agency. AI should preserve meaningful human control and keep important decisions open to human review.
2. Transparency and explainability, in proportion to risk and impact.
3. Clear accountability through traceability and redress. Responsibility should remain attributable to identifiable persons, not "the AI did it", backed by a traceable chain of accountability.
4. Safe and secure use. Spot and deal with risks before they cause harm.
5. Responsible data governance. Data quality, integrity, provenance (where it came from) and security all affect how fair and reliable an AI is.

The Bill would make this a statutory duty: Developers and Deployers would have to give "due regard" for these principles throughout the system's life. "Due regard" means actually applying them, scaled to the system's nature, context and impact.

The three risk tiers

The proposed risk framework is anchored on four kinds of harm: death, bodily injury, unlawful loss of fundamental liberties under the Federal Constitution, and breaking any written law.

Proposed tier	What it means	Possible regulatory response
Unacceptable risk	AI developed or deployed with the intention of causing specified harm	Could be prohibited or subject to immediate regulatory intervention
High risk	No intention to cause harm, but a risk of specified harm occurring	Could face additional risk assessment, documentation, internal control, human oversight, monitoring and mitigation requirements
Low risk	No foreseeable material AI harm	Baseline governance principles, ordinarily without the additional obligations contemplated for high-risk systems

The consultation envisages requirements set by class of AI system (rules for the technology itself), by sector (industry codes of practice), or by domain of deployment (rules for a specific use case, such as AI in hiring).

A staff-information chatbot and an AI hiring tool can create different risks. Classification would depend on their actual use, potential harm and applicable requirements, not simply whether they are everyday tools.

AI incident reporting

The consultation proposes a reporting system for AI-related events, failures, weaknesses, misuse or unexpected effects that cause or may cause harm. Near misses may also be relevant.

Reports could come from Developers and Deployers through set channels, or from the public through the Authority. Relevant report information could include:

• what happened
• the foreseeable harm
• what was done to contain it
• the root cause, if known
• remediation actions taken or proposed

Where a Sectoral Lead already runs a similar reporting system, the proposed framework could build on it.

The AI sandbox

The consultation proposes an AI sandbox: a controlled environment for testing AI in its intended context, including its limitations, safeguards, human oversight and real-world effects.

For smaller firms, this could offer an opportunity to test with regulatory oversight. Access, participation conditions and operating arrangements remain to be settled; the proposal does not guarantee admission or regulatory approval.

What to do now

These are my practical recommendations, not a statement that every step is already required by the proposed Bill:

1. List your AI. Write down every AI tool you build or use, who owns it, and what decisions it touches.
2. Know your role. For each one, are you the Developer, the Deployer, or both?
3. Flag likely high-risk uses. Hiring, lending, health and anything that affects people's rights deserve a closer look.
4. Keep a paper trail. Can you show where an AI answer came from, who approved the action, and when? Keep records proportionate to the system's use and risk.
5. Put a human in the loop for important decisions.
6. Write a simple incident playbook using the reporting topics above as a starting point.

For a broader starting point, see my AI governance guide, the enterprise AI roadmap and what sovereign AI means for Malaysia.

How this lines up with what we build

The proposals overlap with the principles guiding our work on SenangAI at Kognosys: source-backed answers, access controls, human approvals, auditability and customer control over deployment.

These are product priorities, not a claim of compliance with the proposed Bill. My view is that businesses should be able to check the evidence behind AI-assisted work and identify who is responsible for consequential actions.

Quick answers

Is the Malaysia AI Governance Bill law yet?

No. As of 2 November 2026 it has not been enacted. The Minister's 29 October update indicated plans to table it in early 2027, following Cabinet consideration. The July consultation proposals may differ from the final legislation.

Does the Bill apply if my AI is hosted overseas?

Under the proposal, overseas hosting alone would not exclude you. The proposed scope includes AI used by a Deployer established in Malaysia, regardless of where the system is physically hosted.

Does it apply to personal use of ChatGPT?

The consultation proposes an exemption for an individual's personal, family or household use. Using AI for company work should not be assumed to fall within that exemption.

What counts as high-risk AI?

The consultation describes AI developed or deployed without an intention to cause harm, but which creates a risk of specified harm occurring. Classification and applicable safeguards would depend on the system, its context and future regulatory instruments.