An infrastructure-first approach as key to scaling public sector AI

Governments around the world are increasingly deploying artificial intelligence (AI) across their public services, but few have done so as deliberately or visibly as Singapore.

About 80 per cent of Singapore’s 150,000 public sector workers had used AI tools in the workplace, according to Singapore’s Ministry of Digital Development and Information in November last year.

Deputy Prime Minister Gan Kim Yong said that Singapore cannot compete on the size of its AI models or data centres, but can leverage its government, industry and research institutions.

Singapore’s approach to AI in the public sector has been one focused on developing institutional processes and sovereign infrastructure, which are aspects which may deserve credit for enabling rapid and coordinated civil service adoption.

Centralising governance of AI tools

Singapore takes an approach of institutional centralisation, that is, consolidating the oversight over its multiple AI deployments to a single body.

In 2016, Singapore established the Government Technology Agency (GovTech Singapore), a specialised institution in charge of driving digital transformation in the public sector.

Chatbot assistant Pair Chat was born out of GovTech Singapore’s annual hackathon competition. Pair is used regularly by a third of all civil servants to help streamline writing, coding and administrative work.

The agency also created Singapore’s second-most widely adopted AI product for civil servants, AIBots.

Pair and AIBots illustrate how one agency can build and scale tools across the whole- of- government without requiring other organisations to navigate independent procurement processes.

As of 2025, the US government reported 3,611 individual AI use cases.

Despite the large scale, each agency implements government-wide AI policy according to its own capacity and maturity, leading to inconsistent practice on the ground.

This challenge may have been compounded by the elimination of the 18F, a team within the US General Services Administration in charge of building shared digital infrastructure across federal agencies in 2025.

Rather than each agency independently creating AI tools, GovTech Singapore builds shared platforms that any agency can use or build on.

This centralises engineering capability in one institution, while distributing the benefits across the government. 

Risk-based sequencing in deployment

Singapore's National AI Strategy 2.0 committed the government to “risk-based” interventions for AI, which means regulating each use case at a level determined by its risk of causing harm.

This principle guided the sequence in which Singapore built and released its AI tools.

Deploying internally first uses lower-stakes environments as a testing ground before tools are trusted with higher-stakes interactions.

Shared infrastructure came before agency-level applications. Singapore prioritised shared infrastructure such as MAESTRO, GovText, and Government on Commercial Cloud before scaling AI deployment across agencies.   

Civil servant-facing tools came before public-facing ones.

Pair Chat was deployed to government officers before comparable public-facing government AI assistants were broadly deployed.

Built on the same shared infrastructure, AIBots then allowed individual officers to build and deploy their own custom chatbots without needing independent engineering capacity.

France’s experience with the Albert chatbot illustrates the risks of deploying generative AI into citizen-service environments before systems are fully mature.

Although Albert was developed by the French state digital agency, DINUM, the tool was piloted in France Services centres while still producing unreliable answers.

The government later decided not to generalise the system in its original form.

Sovereign AI tools over commercial platforms  

Sovereign AI tools are hosted on government-controlled infrastructure, rather than third-party commercial clouds.

While Singapore’s Pair Chat is built on Claude Sonnet 4, the country controls data boundaries.

This way, data stays within government systems, enabling deployment of sensitive use cases that commercial platforms cannot legally handle.

GovTech Singapore also retains the ability to swap the underlying model when better options become available, without rebuilding the product.

The UAE government is widely cited as a leading adopter of sovereign cloud and national AI infrastructure strategies.

It uses Jais 2, a large language model celebrated by one of its developers as a “blueprint for sovereign AI” that can support public institutions with a system that is contextually grounded in the Arabic language, culture, law, and social norms.  

Meanwhile, in the US, the General Services Administration establishes procurement frameworks for agencies to purchase services from pre-approved commercial AI vendors such as OpenAI, Google, and Anthropic.

The decision to rely on commercial vendors, however, might just be a rational response to having to engineer at a much larger scale.  

Key takeaways 

An important caveat is that Singapore’s approach is hard to replicate.  

Its incumbent political party’s unusually strong continuity and central governance were what allowed government technology to survive across electoral cycles, something many politically fragmented systems struggle to sustain.  

While Singapore cannot export these, its logic is transferable: build infrastructure first, test internally before going public, and govern by risk.