A citizen applying for financial assistance experiences artificial intelligence (AI) through what happens to their application.

A faster assessment offers little reassurance if an unexplained flag leaves them waiting, with no clear route to correct a mistake. 

For agencies introducing AI, that journey makes governance a service delivery issue.  

The quality of the recommendation matters, but so does the organisation’s ability to check it and act appropriately. 

In August 2026, Singapore's Ministry of Digital Development and Information reported that almost 70 per cent of public officers regularly used AI tools for tasks including analysis and drafting. It also noted that the depth and maturity of adoption remained uneven across agencies. 

That distinction matters for Southeast Asia’s AI ambitions: usage can spread before an agency has worked out how to demonstrate that a system remains dependable, or whether it improves the service around it. 

The question for public leaders is whether their governance and data foundations are keeping pace with the responsibilities they are giving AI.  

Keep assurance running after deployment 

ASEAN’s expanded AI governance guide addresses generative AI, while Malaysia’s public-sector AI adoption guidelines set out risk categories and responsibilities, alongside requirements covering data quality and organisational readiness. 

Singapore’s Model AI Governance Framework for Agentic AI, launched in January 2026, extends this work to systems that can take actions. 

These frameworks provide a starting point, but their practical value depends on how agencies translate them into decisions about a live service: what to monitor, what level of error is acceptable, and who can intervene. 

Healthcare illustrates why those decisions must be revisited. A model evaluated against one patient population may encounter different patterns when introduced elsewhere.  

Changes in care-seeking behaviour or the data being collected can also affect performance, even when the model itself has not changed. 

In published commentary on AI-enabled real-world evidence, SAS’s Global Head of Real-World Evidence and Epidemiology, Sherrine Eid, highlights how differences in health-seeking behaviour across populations and over time influence datasets.  

She calls for organisations to determine what variation in model performance is acceptable. 

For public providers, the implication is to define review thresholds before deployment and keep checking performance against them. 

Singapore’s health authorities are making this expectation explicit. 

In an October 2026 speech, Senior Minister of State for Health Tan Kiat How described healthcare AI governance as a continuing process encompassing evaluation, monitoring, incident reporting and learning. 

The same reasoning applies to a citizen-service assistant after an eligibility policy changes. 

An approval given before that change cannot establish that its answers remain suitable afterwards. 

Give AI the context behind the data 

Continuous assurance also depends on knowing what information a system is using and what that information means. 

In SAS’s 2025 global research on fraud, waste and abuse, 47 per cent of respondents identified data quality and availability among their leading concerns.  

The survey covered 1,100 senior and mid-level public-sector employees responsible for monitoring these issues worldwide. 

For an agency, improving data quality involves decisions about definitions and ownership as well as correcting incomplete records.  

A system needs to distinguish an application received from an application approved, or an administrative closure from a problem actually resolved. 

GovTech Singapore explored this problem in an October 2026 illustration: an AI tool counting cases marked as resolved could include cases automatically closed after inactivity.  

The database query could run correctly while the answer misrepresented the agency’s work. 

For citizen-experience leaders, that creates a risk of celebrating improvements that residents have not experienced. For transformation teams, it makes agreeing on service definitions part of preparing data for AI. 

An agency should therefore be able to identify who maintains a dataset, when it was updated and which policy definitions apply.  

Where information crosses agency boundaries, teams need agreed meanings and a process for correcting errors at their source. Access should remain limited to the information required for the service. 

These foundations make it possible to investigate an unexpected result and establish whether the problem lies in the data, the model or the way its output was used. 

Give people the authority to act on AI 

Oversight becomes meaningful when staff have the information, time and authority to question a recommendation. 

In SAS’s discussion of public-sector fraud prevention, Principal Industry Consultant Ensley Tan advises agencies to plan both how they will manage data and how they will respond to its findings. 

“Next, figure out how to deal with the decisions and results that come from the data,” he says in the publication. 

This is a useful test of operational readiness. If an AI tool flags a suspicious benefits claim, the agency needs a defined route from the alert to an evidence-based review.  

It should be clear who checks the underlying information, when the applicant can provide clarification and how an incorrect flag is removed. 

The stakes differ between services. Where AI can act, agencies should specify which actions require approval and retain records that allow staff to reconstruct what happened. 

Reviewers need access to relevant source information and the ability to escalate uncertainty. Their corrections should also inform future evaluations, so the same failure does not repeatedly return to the frontline. 

Measure what changes for the public 

The evidence of success should follow the service from beginning to end. 

For a healthcare provider, time saved on documentation is valuable. Evaluation should also examine whether checking and correcting generated notes absorbs that saving, and whether staff gain more time for care.

For teams protecting public funds, the relevant outcomes include confirmed cases and money recovered or losses prevented. Those results should be considered alongside the burden of unnecessary investigations and delays imposed on legitimate applicants. 

Agencies can build these measures into a defined use case from the outset, establish a baseline and assign an owner to review outcomes.  

This gives leaders a stronger basis for deciding whether to expand, modify or stop a deployment. 

It also changes the investment decision. Funding an AI service means funding the data stewardship, staff capability and continuing evaluation needed to keep it useful.  

Shared testing tools and reusable data standards can help agencies spread that work across multiple services. 

Southeast Asia’s governance frameworks can support this shift. Their value will become visible when an agency can show how it detected a problem, corrected a decision and improved the experience of the person relying on it.