By Sameer Sharma

Back in November 2018, the Mauritius Working Group on Artificial Intelligence (AI) published its Mauritius Artificial Strategy which essentially gave a broad and high level overview of the potential benefits of AI in enhancing productivity, growth and the quality of life of all Mauritians. From improving the efficiency of ports to harnessing smart Regtech solutions and using machine learning for more efficient power management solutions, AI seems to be getting all the buzz these days and rightly so. Sure, Mauritius needs to increase the level of investment in and improve the stagnating quality of its education system (in a country where the seniors and civil servants seem to be getting an increasing share of central government budget expenses) and improve its Information and Communication Technology infrastructure and Research & Development spending. But where do we start, how do we start and what can all this mean for the future of employment?

Becoming more data driven

Becoming more data driven is hard. While 85% of North American companies are trying to become more data driven, less than 40% of them have achieved meaningful success. For every 10 companies that invested into AI tech, only three have succeeded so far. This is despite the larger pool of talent, better quality tech infrastructure and know-how which exists in the United States and Canada. Many Canadian financial institutions have poured large sums of money in hiring data scientists, but because of a sub-optimal data/information architectures present in many institutions, there has been little meaningful output in terms of production ready applications.

To better understand the problem, readers must understand how an AI ecosystem looks like. Let us for example take a user case where a bank wishes to develop a credit underwriting model which will almost instantly predict whether an applicant borrower will default or not. Such models will use cleaned data from various sources such as credit bureau reports, application data, alternative data, which all may in their raw formats be structured or unstructured. The raw data must be pre-processed automatically, variables must be created and fed into the model which then will send a prediction with reasons for decline or approval to a front end API end end-user. The ecosystem must have the ability to also learn from new data and models need then to be redeployed.

Over the last few years, data scientists have handled model training and research, data engineers have handled the export, transform and load or export load and transform (of the raw data) process and model deployment while data architects and engineers have worked on DevOps (building the entire infrastructure/data pipeline behind it all). The saying goes, there is no AI without the proper information architecture (IA). The future of the industry is likely to see a more unified approach in terms of job descriptions and tools. Very often there is a misconception that one must first collect the data before moving onto defining architectures and into machine learning. The problem with this approach is that those who are working on the collection and storing of the data may not know exactly what kind of data may be needed and whether the way they are storing it is optimal in a deployment setting. The typical limiting factors can be summed up as follows:

1) The lack of a data strategy which corresponds to the overall business strategy. The need to have appropriate data governance policies has also tended to be under-appreciated.

2) Many organisations still use legacy data infrastructures and have a culture of having silo people, processes and systems. There is currently a focus on getting data from various sources into a single and efficient data lake that can cater to current and future needs be it on premise, in the cloud or both.

3) There are so many database and analytics tools available in organisations that it becomes hard to integrate it all. Data architects, data engineers and data scientists need to filter what they need based on the data strategy.

4) Some firms also face significant technological and skills gaps to implement it all.

5) Existing data environments with legacy systems can be quite complex and for those with on premise architectures, scalability (the more data you get, the more computing power you need to train models) becomes an issue. Even the most reluctant and largest banks today are choosing to move parts of their data architecture into the cloud via a hybrid setup for this reason.

There must be a clear approach to tackle data related issues.

6) There is a need to ensure that there are model validation teams as part of the risk management function for certain AI solutions such as fraud detection and credit underwriting. For example, to manage model risk and beyond model risk, one must cater for risk which goes beyond model risk (the model has learnt to discriminate indirectly on the basis of race, sex, age). There is a lot of effort now to ensure that complex non linear models’ output are interpretable. A black box approach does not gain traction with stakeholders and regulators. There must be a clear approach to tackle data related issues.

In sum, Mauritian companies seeking to become more data driven must ensure that they properly define strategies, policies, optimal data architectures and standards and enforce compliance. There must be a clear approach to tackle data related issues. Today cloud providers such as Google Cloud, Microsoft Azure and Amazon AWS (the rest do not even come close) offer an optimal and highly scalable ecosystem for storage, data processing, machine learning and deployment. A hybrid approach to the cloud would be appropriate for Mauritius where sensitive data would be kept on premise.

Optimising costs

Over the next decade from banking to retail, firms which are able to offer customized solutions to their clients will win out. In retail for example, firms that can harness data to build more intelligent recommender systems via machine learning could be able to optimise marketing costs and sales. Machine learning in the retail space can be used to optimise supply chain management, consumer up-selling, consumer cross-selling and sales forecasting.

On the insurance side, Canadian auto insurers today offer the driver the option to share his data for a decent discount. In turn the insurance company is able to build machine learning models which are better able to price risk. In the offshore sector, a lot of repetitive jobs when it comes to Know Your Client or anti-money laundering, can be automated today. In the accounting and legal fields, natural language processing can today make for more efficient lawyers and auditors.

Within the banking space, regtech solutions such as fraud detection models can provide near real time detections and massive out performance when compared to traditional rules based approaches to detecting fraud. The ability to use machine learning within the lending space in North America has not only optimised costs but has allowed lenders to massively increase access to previously under served borrowers to loans without increasing their risk because models that are well trained are better able to classify good from bad future borrowers within seconds. Both the Financial Services Commission and the Bank of Mauritius today collect a lot of data but have yet to implement a roadmap to get to supervision tech. In Canada, machine learning has recently shown good promise at optimising crop yields in agriculture.

While the opportunities are clear, the lack of skills and perceived high costs are often pointed to as being limiting factors. On the infrastructure side, AWS, Azure and Google Cloud today have drastically lowered the barriers to entry when it comes to storing, training and “productionizing” end to end and highly scalable machine learning pipelines. Software tools such as Python and R today are completely free and open sourced with all the machine learning libraries one can think of. When it comes to data preparation and beyond the much cheaper SQL type databases found on the cloud today, efficient data prep tools such as Trifacta Wrangler (with Spark at the back) are reducing data prep time to hours or days from weeks with full audit trails. It has hence never been easier and cheaper to become more data driven even for Mauritian companies assuming one gets the data indestructible right.

When it comes to skills, while there is no doubt that Mauritian firms will need to hire consultants (especially for formulating the data strategy and appropriate data architecture) and domain experts from abroad initially, policymakers must ensure that the labour force becomes increasingly fluid. Data science is a field where one is constantly learning and improving. Coding must be taught from early years while on premise Bootcamps and on line educational platforms such as Datacamp or Coursera offer a good entry point. From deep learning to training a basic linear regression, it has never been cheaper and easier to learn on-line. Even the necessary math skills, which are no longer taught to a proper degree in local schools, can be learnt on line.

Firms which are able to offer customized solutions to their clients will win out.

In Canada and in the United States, a more developed venture cap ecosystem and a culture of welcoming the participation of start-ups has been key to the development of the AI industry. In Firms which are able to offer customized solutions to their clients will win out. Mauritius, as I have often argued, local private and public pension funds along with policymakers have not developed the capital markets and alternative lending solutions enough. Pension funds all around the world which are professionally managed invest in their local venture cap ecosystems. In Mauritius, despite low bond yields and sub-optimal asset allocations, there has been little movement in this direction.

Building a better start-up ecosystem

This will perhaps be the biggest challenge in Mauritius. The government must itself lead the way by allowing start-ups to help it design appropriate data architectures and AI ecosystems which allow it to optimise costs and offer more efficient services from health care services to education. It is easier to obtain meetings with large corporate US and Canadian CEOs and decision makers as a start-up than it is to get the ball rolling in Mauritius. The Mauritian market is much more closed and while it is easy to set up a business, it is something else to break the many barriers which still exist in the country. This includes working with the public sector and with the regulators and public owned companies.

In Canada and in the US, larger firms are incentivised to work with start-ups. In Mauritius we indirectly encourage everyone to become employees out of the lack of access and support. Building a better start-up ecosystem along with an appropriate immigration policy remains key to AI innovation, and for Mauritius to become this fintech platform we often like to talk about. AI is complicated and our regulators too must be well trained. Very often those who obtain sandbox licenses complain that regulators are simply behind the curve or take too long when it comes to getting licenses and understanding pros and cons of what they propose.

When it comes to employment, with wage growth in Mauritius remaining above productivity growth, policymakers should be very mindful of the fact that in a country with a large pool of semiskilled labour, many repetitive jobs will be eliminated in the coming decades. While demand side economics with a focus on consumption may be politically sellable and win elections, policymakers must put more resources behind improving the deteriorating quality of education and re-skilling programmes. Not everyone can become a data scientist of course but there will be many jobs which simply cannot easily be replaced by AI.

Jobs that require human interactions and touch will remain. The tourism industry, if it can maintain its Mauritian hospitality brand, will continue to employ people. The entertainment industry too will do well in the era of the fourth industrial revolution. Many jobs which require human to human interactions will continue to flourish. Our education system must hence be able to direct the right people at the right places. If we are not careful, automation and AI will lead to job losses and social unrest, but if we play our cards well, a small country like Mauritius has more to gain than to lose if it is able to harness the fruits of Artificial Intelligence.

Sameer SharmaAI Specialist
Sameer Sharma, a chartered alternative investment analyst and a certified financial risk manager, is an AI Specialist for Princeton based Creative Business Decisions, a leading regtech AI provider.