AI reaches Mauritius through the cloud: what you still have to decide

July 21, 2026 · 4 min read · SOVALYX Technologies

The allmyt summit on 16 and 17 July marked a clear acceleration: an agreement between Mauritius Telecom and Amazon Web Services, consumer AI platforms, and a target of 50,000 Mauritians trained within a year. For a local business, the announcement changes the backdrop but not the responsibilities: five decisions remain entirely yours.

What was announced, and what it means

Mauritius Telecom signed a memorandum of understanding with Amazon Web Services, presented on stage by AWS's general manager for Sub-Saharan Africa, with the Prime Minister in attendance. The operator launched several AI services alongside it, including an education platform rolled out to Grade 9 students and their secondary school teachers, Rodrigues included.

The economic subtext is explicit. On the figures presented at the summit, the ICT sector's contribution to Mauritian GDP fell from 7.4% in 2014 to 5.4% in 2024, across more than 35,000 jobs. AI is framed as the way to close that gap. The ICT minister summed up the precondition in a line worth taking seriously: « Without the trust in data, we can forget about AI ».

That is exactly where the work starts for a business. A deal between an operator and a hyperscaler brings capacity, pricing and skills. It answers none of the questions that carry your legal and operational liability.

First decision: which data is allowed to leave

Before choosing a platform, you need to know what you are about to put on it. The problem categories are well known: customer and employee personal data, health records, financial data, trade secrets, contracts. The Mauritian Data Protection Act governs transfers off the island and requires you to demonstrate that the recipient offers an adequate level of protection — an obligation that stays yours, whoever sits in the middle.

Mapping is a short and usually instructive exercise. Most organisations find that only a minority of their data is genuinely sensitive, and that this minority circulates far more freely than they assumed. That minority is what justifies local private hosting; the rest can probably move to a hyperscaler without losing sleep.

Second decision: what the AI does with your documents

An enterprise AI assistant is only useful if it reads your documents. So the questions to put to the vendor are the ones you would put to any processor: where is the data processed, is it retained, does it train a model, who can access it on the provider's side, and what happens when the contract ends?

When the answers do not work — law and accounting firms, healthcare, HR, finance functions — a private language model hosted on your own infrastructure removes the question at the root: the documents never leave your perimeter. This is not an ideological choice but a trade-off between cost, performance and confidentiality, settled case by case.

Third decision: continuity stays with you

This is the most common blind spot. Moving a workload to a hyperscaler does not transfer the risk of interruption: it relocates it to a supplier you have no leverage over. Large platforms do go down, often through internal configuration errors, and their service agreements refund credits — never a lost day of revenue.

So the questions remain open: how long can your business run without that platform, how do you operate in degraded mode, and by what route do you come back if the provider becomes unavailable or too expensive? A written and tested recovery plan does not become optional because the infrastructure sits elsewhere. It becomes harder to write, and therefore more necessary.

Fourth and fifth decisions: skills and the way out

The national target of 50,000 people trained in AI within a year is good news, but mass training does not produce the judgement a board needs this year: which use cases deserve funding, which are gadgets, and which data must stay in Mauritius. That skill is hard to recruit in an already tight market — one reason a fractional IT leadership arrangement makes sense for a mid-sized organisation.

Then there is reversibility. An AI platform gets chosen today on price and availability that will change. Knowing what format your data comes back in, at what cost and how quickly, is a clause to negotiate at signature — never at the moment you want to leave.

Questions to ask before signing

• Map it: which data categories does this project involve, and which are sensitive or personal?
• Locate it: in which country is the data stored and processed, and on what legal basis if it leaves Mauritius?
• Check training: does your content feed a model? Get it in writing, not in a meeting.
• Price the outage: what does a day without this platform cost, and what is the degraded mode?
• Write the recovery plan and test it, including the "provider is unavailable" scenario.
• Negotiate the exit: export format, timeframe, cost, migration support.
• Name an internal owner for the file, with the authority to reject a use case.
• Decide case by case between public platform and local private infrastructure, project by project.

🧰 The companion tool: Are you ready for the AI Act? — free · 2 minutes.

Reviewed and optimised by AI.