Malaysia AI Governance Bill: What Is Known and What to Do Now

Malaysia's AI Governance Bill is not law yet: on 10 September 2026 the draft was being submitted to Cabinet, with tabling hoped for by early 2027. AI is still regulated in the meantime. Since 30 April 2026, the Personal Data Protection Commissioner's guidelines have expected a formal impact assessment for automated decisions or profiling that have legal or similarly significant effects on a person. If you run AI on Malaysians' personal data, the useful work starts now: an inventory of your AI systems, a named owner for each, and an impact assessment where those decisions happen. Singapore's IMDA and the EU ask for similar artefacts, with very different legal force.

On 10 September 2026, Digital Minister Gobind Singh Deo said: "The draft has already been completed ... we are now in the process of submitting the paper to the Cabinet." On timing he added: "So, hopefully, it will be tabled in Parliament this quarter or, if not, in the first quarter of next year" (Bernama, 10 September 2026). That is a hope, not a deadline. This is general information, not legal advice.

Where does Malaysia's AI Governance Bill stand in September 2026?

Malaysia's AI Governance Bill is a completed draft being submitted to Cabinet, not a law. The Ministry of Digital consulted on it from 10 July to 1 August 2026 through the government's Unified Public Consultation portal, which lists the exercise as "pre-drafting" and shows 66 public comments. On 10 September 2026 the minister said he hoped to table it this quarter or, failing that, in Q1 2027.

With two weeks of the quarter left on 15 September, plan for 2027. After Cabinet, the bill still needs Parliament, Royal Assent, the Gazette and a commencement date, and Malaysia's last data law change shows how long that tail runs. The PDPA amendment received Royal Assent on 9 October 2024, then came into force in three stages: 1 January, 1 April and 1 June 2025 (P.U.(B) 522, 24 December 2024). Our reading: an AI Act passed in 2027 will probably phase in the same way.

Some duties may not wait for the Act. The minister described AI governance as "standards, regulations and laws", with MY-AI Standards being developed with SIRIM Berhad and an enforcement structure under AI Malaysia built in parallel. On 13 August 2026 he also said the government was examining two Acts to create a national data commission with no timeline given (The Star, 13 August 2026).

What will the bill cover, and what is still unknown?

The Ministry of Digital calls the bill "Malaysia's first horizontal legal framework specifically dedicated to AI governance": it cuts across every sector and complements existing laws. Under its risk-based approach, "governance responsibilities will be distributed based on the roles and level of control of the parties involved" (Ministry of Digital, 10 July 2026).

The same statement encourages incident reporting and AI sandboxes (regulator-supervised test environments), and records the minister's position in Parliament on 24 June 2026: the bill will not directly regulate AI output. It is about who answers for an AI system, not what the system says.

The consultation paper sits behind a login; the public submissions on the same page discuss three risk tiers (unacceptable, high and lower), two regulated roles called developer and deployer, a Central AI Authority working with sector regulators, and incident reporting that covers near misses. Several submissions criticised undefined triggers such as "material effects". The Cabinet draft may differ.

The status, item by item:

Which Malaysian laws already apply to AI today?

The Personal Data Protection Act 2010 already governs most business AI in Malaysia, because most business AI runs on personal data. Act A1727 tightened the Act during 2025, and on 30 April 2026 the Personal Data Protection Commissioner published DPIA and automated decision-making guidelines that reach AI directly.

From 1 April 2025, biometric data counts as sensitive personal data, data processors (your AI and cloud vendors among them) must meet the Security Principle in their own right, and breaching the data protection principles can cost up to RM1 million, three years' imprisonment, or both (Act A1727, in force per P.U.(B) 522). From 1 June 2025, controllers and processors must appoint a data protection officer (DPO), and a controller must notify the Commissioner of a breach "as soon as practicable". The Act sets no hour count. The Commissioner's breach notification form does: it asks whether you are reporting "within the 72 hours after becoming aware of the personal data breach".

The April 2026 guidelines are what most coverage of the bill misses. A data protection impact assessment (DPIA) is a documented risk review done before processing starts. The DPIA guideline says a controller "shall carry out a DPIA" where processing is likely to be high risk, and deems two cases high risk outright: more than 20,000 people, or more than 10,000 where sensitive data, including financial information, is involved (PDP DPIA Guideline, para 7.5, April 2026). Findings go to senior management, and a completed DPIA is valid for two years.

On legal status, be precise. The Act itself "does not currently provide specific provisions" on automated decision-making, as the ADMP guideline admits in paragraph 1.1. Both guidelines are issued by the Commissioner under section 48(g) and tied to the DPO's statutory role: regulator guidance in "shall" language, not statute. MOSTI's National Guidelines on AI Governance and Ethics, September 2024 sit a step lower: they call themselves voluntary. Good vocabulary for a policy, no legal force.

When does an AI system need a DPIA in Malaysia?

An AI system in Malaysia needs a DPIA when it processes personal data of more than 20,000 people, or when its automated decisions can legally or significantly affect a person. Both triggers come from the Commissioner's April 2026 guidelines; the headcount drops to 10,000 for sensitive data. Below both lines, the DPO decides using the guideline's qualitative factors.

The companion guideline on automated decision-making and profiling (ADMP: decisions made with little or no human involvement, and automated predictions about a person) states that ADMP "triggers the requirement to carry out a DPIA, regardless of the nature or extent of its intended use" (PDP ADMP Guideline, para 6.1, April 2026). Paragraph 7.1 narrows that: the guideline applies where the outcome may "result in legal effects concerning the data subject" or "significantly affects the data subject". The guideline calls this the ADMP Threshold.

Paragraphs 7.2 to 7.4 list what counts. Use them as a test, one system at a time:

1. Does the system affect someone's money, such as eligibility for credit?
2. Does it give one person a better price than another?
3. Does it gate access to a job, health services or education?
4. Could its output cause reputational harm, or end a contract or entitlement?

One yes means the ADMP Threshold is likely met, so plan a DPIA. Four noes, and a headcount under 20,000, leave it to the DPO's judgment.

The counter-intuitive part is the human in the loop. The guideline counts a process as automated where "human influence is minimal", including where a person merely enters the data and the system decides (para 4.1). A reviewer who clicks approve on every score does not take you out of scope. A reviewer with real authority and the information to overturn a score might. For systems that pass the threshold, the privacy notice must say that ADM or profiling is used (para 9.3), and as a best practice the guideline adds that "AI must not be relied upon as the sole factor" in decisions about a person.

What should a firm operating in Malaysia do before the bill is tabled?

Start with three artefacts: an AI inventory, a named owner per system, and a DPIA wherever the ADMP Threshold is met. The Personal Data Protection Commissioner's April 2026 guidelines already expect the third, and the first two make it possible. None of it is wasted, whatever the final bill says.

The inventory comes first. Count the AI you did not buy as AI: lead scoring inside the CRM, a screening feature in the hiring tool, the chatbot a vendor switched on. For each, record the purpose, the personal data used, the number of people touched, whether it decides or ranks anything, and the supplier. Add a column the bill may need: public submissions used a developer and deployer split, so note whether you built the system or only run it.

Then name owners. The bill signals duties by "level of control". In practice that means two names per system: a business owner who answers for the outcome, usually the head of the function that uses it, and a technical owner who answers for access and logs. The DPO is consulted on both.

Then assess what decides. Run the four-question test above, then a DPIA on every system that passes it, using the template in the Commissioner's guideline, led by the DPO. Where the significant-effect call is borderline, take legal advice before you decide a system is out of scope, not after.

Illustrative example: A Kuala Lumpur insurance agency with 80 staff and 40,000 customer records uses three AI tools. A meeting-notes assistant summarises internal calls. A WhatsApp bot answers policy questions for a few thousand customers a year and decides nothing. A vendor model scores renewal leads and decides which customers are offered a discount. That last one gives some people a better price than others, which paragraph 7.4 lists as a significant effect, and it touches more than 20,000 people, so it needs a DPIA now. The bot stays under both triggers and gets a register entry; the notes assistant gets one line. The decision: one DPIA this quarter, not three, owned by the head of sales with IT as technical owner.

The typical failure runs the other way: a firm assesses its chatbot because it is visibly "AI" and misses the scoring model buried in the CRM. For the MD, the cost is mostly people time and the hard call is ownership. For the CTO, the work is evidence: decision logs, a review step that can overturn a score, and vendor contracts that reflect the processor duties in force since April 2025. For the DPO, it is DPIAs and notices this quarter, then a watch on the Cabinet text.

How does Malaysia compare with Singapore and the EU?

Malaysia sits between Singapore and the EU. Singapore governs AI through voluntary IMDA frameworks on top of its PDPA. The EU has a binding AI Act with fixed dates. Malaysia has binding data protection law, regulator guidelines that already reach AI decisions about people, and an AI bill in draft. All three start from similar artefacts.

The table sets legal status against the artefact each regime expects.

The Singapore column draws on IMDA's Model AI Governance Framework for Agentic AI, version 1.5 (20 May 2026, updated 5 June 2026), which asks that "all agent identities (and their attendant permissions) should be issued from and tracked by a centralised system". It is guidance; our reading of that framework against the EU AI Act explains why that matters. Malaysia's own DPIA guideline (para 5.2) lists Singapore among jurisdictions where DPIAs are recommended best practice rather than mandatory.

The EU asks for far more than three artefacts. Providers of high-risk systems also face conformity assessment, a quality management system and post-market monitoring (AI Act, Articles 43, 17 and 72). The Digital Omnibus moved the Annex III high-risk obligations to apply from 2 December 2027 (Regulation (EU) 2026/1744), and Article 27 requires a fundamental rights impact assessment from public bodies, providers of public services, and deployers of credit-scoring or life and health insurance pricing systems. Our EU AI Act guide for Singapore SMEs covers who is caught. For a firm in all three markets, Malaysia's expectation is the earliest live one.

Where Origin Pi stands

Waiting for the bill is the wrong clock. Whatever tiers the final text sets, the decisions that will draw scrutiny are the ones AI makes about people, and Malaysia's data protection regulator has already said those need an impact assessment once they cross the ADMP Threshold.

We design AI and agent workflows with that scrutiny in mind. Decisions that affect a person pass a confirm step taken by a named reviewer who can overturn them. Agents run with bounded permissions, so a lead-scoring model cannot also send the discount offer on its own. Every action leaves an audit trail that a DPIA or a future incident report can draw on without reconstruction. We build that evidence once rather than per regulation; our AI compliance page describes the approach.

Your first action this week: list every system in your business that scores, ranks, prices or filters people, and run the four questions above on each. Any system with a yes, or touching more than 20,000 customers, gets its DPIA started before the bill reaches Parliament.