AI in Malaysia · Public sector

Government AI in Malaysia has moved from pilots to contract terms. Read them before you bid

In two years the public sector has gone from a 270-officer pilot to AI tools for 445,000 officers and an AI agent in the MyGOV app. Along the way it has written down what it expects of anyone who builds for it.

Pikabyte Engineering 7 min read

445,000

public officers to be given generative AI tools under the programme announced in February 2025

On 20 October 2026 the Digital Minister answered a parliamentary question in writing. As reported by The Edge, the reply set out how the government handles its data when cloud and AI services are involved.

  • Highly classified public-sector data may not be stored in cloud environments at all.
  • Confidential and restricted information may be stored only in domestic private or hybrid cloud.
  • Only open, non-sensitive data may go to cloud environments overseas.
  • Government data must be encrypted at rest and in transit, and for classified systems the encryption keys are owned and controlled by the government.
  • Vendors are security screened. Their access is role-based, uses multi-factor authentication and is logged.
  • AI procurement contracts bar providers from training models on government or public data, and require the data to be deleted after processing.

The reply added that the ministry is drafting an AI Governance Bill.

For anyone who builds software for agencies, statutory bodies or government-linked companies, that is the closest thing to a published specification. It did not appear from nowhere.

What has actually been rolled out

December 2024. The National AI Office was launched on 12 December. A pilot called AI at Work put generative AI in front of 270 public officers.

February 2025. On 5 February the Digital Ministry and Google Cloud launched AI at Work 2.0, with some 445,000 public officers to get access to Gemini in Google Workspace. At its 100-day mark, the National AI Office said more than 400,000 civil servants across 23 ministries, 201 agencies and three state governments had access.

February 2025. On 27 February the National Digital Department, working with MDEC, launched the Public Sector AI Adaptation Guidelines, known by the Malay acronym GPAISA. Utusan Malaysia reported a 158-page document covering ethics, roles and responsibilities, risk management and a self-assessment template for agencies.

April 2025. The Data Sharing Act 2025 came into effect on 28 April. It gives agencies a formal route to request data from each other, with a response due within 14 days. An officer who misuses shared data faces a fine of up to RM1 million, up to five years in prison, or both.

July 2026. AI Malaysia Berhad was launched on 28 July as the national lead AI agency under the Ministry of Digital, formed by institutionalising the National AI Office. Its functions include AI governance and the Malaysia AI Safety Institute.

August 2026. An AI agent entered beta testing in the MyGOV Malaysia app. Bernama reported that, as of July, the app had more than 2.9 million users and offered 52 services from 19 agencies.

October 2026. MyDigital ID had 14.2 million registrations and 82.75 million logins since 2024, as of 7 October, and supported more than 200 applications and services.

November 2026. AWS announced a programme to make its workplace AI tool available to up to 445,000 officers through the National Digital Department’s AI-as-a-service platform, and a developer platform for GovTech Malaysia starting with 500 developers in 2026.

What one chatbot taught

The MyGOV app had an AI chatbot once before. In August 2025, during the app’s beta, SoyaCincau reported that the feature was disabled after it gave wrong answers. Asked about the RM100 SARA aid and the RM1.99 RON95 petrol price, both then recent announcements, it said they “have commenced since 1st July 2022”. It named the wrong ministers for several portfolios. The explanation reported at the time was that it used a local language model still undergoing training. The chat box was replaced with “Akan Datang”.

It would be a mistake to treat this as one team’s problem. Any language model answering from memory will state out-of-date facts with confidence. The fix is in the design, not the model:

  • Answer only from current, approved source documents, and show which one.
  • Say “I do not know” when the source is silent.
  • Hand over to an officer when the question is about a person’s own case.

The 2026 version takes on a harder job. The ministry’s example is a user who checks a police summons and is guided through to payment. That needs the system to know who is asking and what they are allowed to do, which is what a national digital identity is for. The ministry has said the rollout will be “in stages and responsibly”, with safeguards covering security, data privacy, accuracy, transparency and human oversight.

Four questions a vendor will be asked

1. Where does each class of data go? The three tiers in the October reply decide the architecture before any code is written. Classify every field at design time. An overseas model API is an option only for open data. Restricted information needs a model running in a domestic private or hybrid environment. Highly classified material stays off the cloud. The guidelines also call for data to be minimised, masked or anonymised before AI processing. We cover the hosting options in where your AI runs.

2. What does the model provider keep? No training on the data, and deletion after processing, are now contract terms. You need the same promise from your provider in writing, and logs that let you show it was kept.

3. Does it work in Bahasa Malaysia? Circulars, forms and letters are in the national language. Citizens write in to agencies the way they speak. Test on real documents and real messages, not on English examples translated for the demonstration. See how well language models handle Bahasa Malaysia and Manglish.

4. Who is accountable for the decision? At the launch of the guidelines, the minister said AI must not become a “black box” that makes decisions without robust checks and balances. For a decision about a citizen, such as eligibility for aid, a licence or a summons, that means a named officer, a record of what the AI was shown and what it produced, and a route to have the decision looked at again.

The pattern we use in regulated work fits here. AI reads and drafts. Rules that decide eligibility or compute amounts are ordinary code, tested, giving the same answer every time. An officer approves. We describe it in human in the loop is a design problem and, for a regulator with similar expectations, in what Bank Negara’s paper means for vendors.

One more follows from the access rules. Role-based access, multi-factor sign-in and logging are hard to honour if the system reaches its AI provider with a shared, permanent key. Retire those keys before the security review asks.

If you are a GLC or statutory body

Not every instrument that binds a ministry binds you. Check which apply. But your board, your auditors and your regulator read the same documents, and the three-tier rule is a sound default for any organisation holding data about the public.

What to do before you bid

  1. Write the data map first. Every field, its classification, where it is stored and which of them reach a model.
  2. Decide where the model runs for each class, and be ready to show the agreement that covers it.
  3. Ground every citizen-facing answer in a source document, and keep the reference with the answer.
  4. Keep decisions in rules and with officers. Use AI for reading, drafting and routing.
  5. Log every AI output with its input, the model and the officer who accepted it.
  6. Run the agency’s self-assessment template yourself before the agency does.

The direction of travel is clear enough: more AI in government, under terms that are being written into contracts and, soon, into law. A system built so that every output has a source, a record and a person answerable for it will meet those terms as they arrive.

Sources

This article is general information and is not legal advice.

Written by

Pikabyte Engineering

Pikabyte builds custom business systems, applied AI and secure cloud infrastructure from Malaysia. Talk to us.

Start a conversation

Want this applied to your business?

Tell us what is slowing the business down. You will talk to an engineer, not a salesperson, and get a straight answer on whether software is the right fix.