The Baltic republic of Estonia is renowned for being one of the most technologically advanced countries in the world. With just over 1.3 million inhabitants, the small state has over 1,000 start-ups and five unicorns (a start-up that manages to reach a valuation of $1bn), and 5.9% of workers are employed in the ICT sector.
Estonia’s swift technology development began shortly after independence from the Soviet Union. Launched in 2001, the country’s e-Estonia initiative has created a digital ecosystem where 99% of government services are delivered online.
Thanks to what has been described as “the most ambitious project in technological statecraft”, today 95% of tax declarations and 99% of bank transfers in Estonia are done online. The country is entirely covered by broadband internet and practically all companies have an online presence. Digital signatures save the government 2% of GDP per year.
This platform allows the Estonian government to explore digital innovations before most of its peers, and it is currently developing applications of artificial intelligence for public services. One of them is a tool helping civil servants to get unemployed people back to work. Tech Monitor spoke to Ott Velsberg, government chief data officer at the Ministry of Economic Affairs and Communications, to find out how it works.
Estonia’s AI tool to guide people out of unemployment
Two years ago, there were only four AI pilot projects in the Estonian government. Back then, Velsberg says that there was a lot of scepticism towards the technology and limited experience among government employees.
Managing expectations is still an important part of AI implementation, he says. “One of the things I still see is that when you mention AI, people have expectations of a Robocop doing everything,” says Velsberg. “Actually, AI is still a baby that you need to show everything that it is able to do.”
AI is still a baby that you need to show everything that it’s able to do.
Ott Velsberg, CDO, Estonia
To accelerate adoption, Velsberg initiated a “deep dive” to identify challenges that each government organisation or agency faces. Velsberg and his team would typically map out around 30 different problems (in some cases it would be up to 200) and consider potential solutions from a customer-centric point of view. This list would be whittled down to two to five areas where AI could offer a viable solution.
One of the solutions that has emerged is a tool that helps civil servants in the Estonian unemployment office (EUIF) understand and better advise job seekers. A machine-learning system (named “kratt”, after a magical creature in Estonian mythology) was trained using more than 100,000 records to identify training and employment options for job seekers while distributing the workload between civil servants.
The decision-support tool advises civil servants on the best options to help an unemployed applicant to get back to work, such as training, reskilling, or studying a language. The data analysed includes the applicant’s employment record, whether they have received unemployment benefits and their working experience.
“Based on this, the person is [given a calculation] on the likelihood of returning to the job market and finding a job, and what affects it and how could it be done, for example, by learning a language, having an e-mail address, different career path, etc,” adds Velsberg. “This is not an absolute decision, but it assists the government agent in decision making.”
The tool was developed in collaboration with the Center of IT Impact Studies (CITIS), strategic consulting company Nortal and data analysis firm Resta. It has been operational since October 2020 after being tested across EUIF offices for six months, and so far proven “very successful” in helping people back into employment, Velsberg says.
Removing bias from the decision-making process
One of the advantages of the tool, says Velsberg, is that it minimises any bias in the civil servant’s perception of the applicant. “We don’t need to worry on the kind of more emotional aspects, but rationally rely on the data that we have,” he says.
Eliminating bias from the system itself requires careful attention to how the data is collected, its provenance and its quality. This is aided by various practical data management safeguards, such as mandatory privacy impact assessments. “The majority of cases where there has been an issue with bias, they are not related to artificial intelligence,” he says. “They are typically simple rule-based IT developments and what happens is human error.”
The unemployment agency is now working on a parallel project that uses machine learning to understand why people become unemployed and find ways to prevent it. “The logic behind this project is that government should try to avoid unemployment because it’s the most expensive cost for society and the government itself. We want to understand how we can pre-emptively step in.”
