AI in the Mauritian Public Sector: Where It Works, and Where It Should Not

With AI named the first pillar of the Budget 2026/27 and a National AI Strategy explicitly covering citizen services, Mauritian public bodies are under real pressure to deploy something. The Budget funded that pressure directly: 5,000 public officers are to receive AI training over the year, and measures include AI chatbots in public licensing systems and a Healthcare Innovation and AI Unit to evaluate clinical deployment. The Budget 2025/26 had already put Rs 25 million toward equipping ministries with AI tools and established a dedicated AI Unit at MITCI.

That pressure produces both good projects and bad ones, and the difference is usually visible at the design stage. This is a framework for telling them apart, written from the perspective of building the systems rather than commissioning them.

The distinction that matters: assistance versus determination

Almost every useful question about public-sector AI resolves to one thing: does the system help someone decide, or does it decide?

Assistance is where the strong public-sector cases live. Determination is where the reputational and legal exposure lives, because a citizen affected by an automated determination has interests that a productivity case does not address.

Strong candidates

• Triage and routing. Directing a citizen enquiry to the right department is a classification problem with a cheap failure mode: the enquiry gets re-routed. Nobody loses an entitlement because a form went to the wrong desk first.
• Document processing at scale. Extracting structured data from forms and records is genuine drudgery, and errors surface quickly when a human reads the result.
• Translation and plain-language rewriting. In a jurisdiction operating across English, French and Kreol, making official information comprehensible is a real access-to-services problem, and the output is reviewable before publication.
• Backlog analysis. Finding the patterns in why cases stall is analysis, not decision-making, and it tends to produce process improvements rather than software dependencies.
• Drafting support for officials. A first draft that a named officer edits and signs keeps accountability exactly where it was.

Where we would not deploy

We would advise against automated determination in any process where the output affects an individual's entitlement, liberty, livelihood or legal status without a person exercising judgement. That includes benefit eligibility, licensing refusals, enforcement targeting against named individuals, and anything feeding a decision that a person has a right to appeal.

The reason is not that models are inaccurate. It is that these decisions carry a duty to give reasons, and a statistical system that cannot explain a specific outcome in terms a citizen can contest is a poor fit for a process built on contestability. A system that is right 97% of the time is also a system that is wrong about a great many people at national scale.

The data protection dimension

The Data Protection Act 2017 applies to public bodies, and it requires impact assessments for high-risk processing including profiling and automated decision-making. For a public-sector AI project touching personal data, the assessment is not a formality to complete after the design is fixed; done properly it is the exercise that determines what the design should be. We cover this in the Data Protection Act and AI in Mauritius.

Three questions worth asking before commissioning

1. What happens to the person the system is wrong about? If the answer involves them noticing and appealing, ask how they would know. People rarely appeal decisions whose basis they cannot see.
2. Who can explain an individual output? Not the model in general: this decision, about this person, on this date. If nobody can, the process has lost a property it previously had.
3. What is the fallback? Public services cannot pause. Any system in a citizen-facing path needs a defined manual route for when it is unavailable or plainly wrong.

If you are on the supply side rather than the commissioning side, the mechanics of selling into government are covered in AI and public procurement in Mauritius.

Common questions

Can a Mauritian public body use AI to make decisions about citizens?

There is no blanket prohibition, but automated decision-making about individuals is high-risk processing under the Data Protection Act 2017 and requires an impact assessment. Beyond the legal position, decisions carrying a duty to give reasons sit awkwardly with systems that cannot explain a specific outcome in contestable terms. Our view is that AI should assist the decision-maker rather than replace them wherever an entitlement, licence or legal status is at stake.

What are the strongest public-sector AI use cases in Mauritius?

Enquiry triage and routing, document processing at scale, translation and plain-language rewriting across English, French and Kreol, backlog analysis, and drafting support that a named officer edits and signs. What these share is a cheap, visible failure mode and a human who remains accountable for the output.

Does the Data Protection Act apply to government bodies?

Yes. The Act applies to public sector bodies including ministries, departments and statutory bodies, and the Data Protection Office operates independently in overseeing compliance.

What should a public body require from an AI supplier?

At minimum: the ability to explain an individual output after the fact, a defined manual fallback for when the system is unavailable, clarity on where data is processed and stored, notification when an underlying model changes, and confirmation of what the body owns at the end of the contract.

How does the National AI Strategy affect public sector projects?

The strategy explicitly names citizen services as a target area and the 2026/27 Budget made AI its first pillar, so there is policy support for deployment. The FAIR guidelines set lifecycle expectations covering monitoring and decommissioning as well as design, which is a useful structure for framing a project's governance from the outset.

Related: Public procurement

Talk to the people who build it →