Akhilesh Tuteja, Partner and Head, Digital Consulting, KPMG inaugurated the session by saying that he is privileged to be a part of the discussion on a topic that everybody is more and more interested to talk about these days. While AI could mean different things to different people, for this discussion, he chose to use AI in generic terms talking about the range of technology that makes AI happen. He set the context of the discussion by asking the questions what do we learn from implementing large scale AI system. Dr Fredrik, Associate Professor of Computer Science, Linköping University, Sweden & President, Sweden AI Society, opined that the term AI is very confusing and the basic question is how do one make computers do things that people can do. One of the major advantages in this respect is scale. He believes that due to automation, we can scale up things that previously needed manual work which aids in cost minimization. He cited an example from Sweden where they used AI to screen children with reading problems in all classes within an hour as compared to previous manual intervention. The second thing regarding key learning lesson which he cited is the need of people with technical competence who not only have deep understanding of their profession but also understand the power of using AI in their particular field.  

Ott Velsberg, Govt Chief Data Officer, Republic of Estonia, took the opportunity to showcase how Estonia went on from becoming a government with a couple of AI systems in 2018 to more than 100 use cases in 2021. The obvious question is how was that possible. Ott believes that the critical aspect is business understanding. Typically, the tiny nation does not start by saying that they are going to use ML to solve a concrete problem but instead started from understanding the business needs and how they can be solved using different processes. He gave the example of their unemployment agency providing recommendations on how to increase employability. He also provided another learning example where due to AI taskforce that they have created, there was a raised awareness about governance. He said that “previously not may organizations considered data quality or providing metadata as a crucial aspect, but due to the work that the government has been doing, more and more people are thinking about how they gather, guard, disseminate, visualize and understand the data.’’ 

Mr. Tuteja then steered the discussion around the need of having less but high-quality data rather than more low-quality data. To this, Dr. Fredrik stated that the today many of the successful techniques are highly data-driven as we rely more on gathering data and annotating them. He gave the example of using computer graphics in order to train object recognition for automated driving which he believes is a fascinating area. In his words, ‘’data is a very important topic”. 

The discussion then shifted towards understanding how to select use-cases for AI in the government and how to solve for issues like transparency in AI. To this, Mr. Ott Velsberg first touched upon the aspect of expert knowledge on data pointing and annotation and its crucial role in the entire project. He reflected on the fact that industry people expect the work of data annotation and labelling should be done automatically using AI. However, the task is not easy and requires tremendous amount of effort. On the use case part, the government of Estonia has a framework to consider. They consider efficiency as one of the parameters. However, efficiency does not imply replacing people but how to simplify tasks which would otherwise be time consuming. Also, for every project, they try to have some measurable metrics and this is typically done through end-user satisfaction with the service or reducing the time span using the system. An interesting fact shared by him was that the measure of success for unemployment recommendation for increasing employability was to see if the person stuck to the recommended job 6 months after joining it. Interestingly, the recommendation provided by the AI system had a success rate of 80% as compared to 58% of human. However, there is no universal evaluation metrics that apply to all projects as it varies from one project to the other and depends on the use-case. In another interesting move, they are working on a government virtual assistant with the idea that people are able to use any government service using voice-based or text-based channel. He signed off by saying that on the privacy side, they have different rules that must be followed. For instance, they have put in place “ethical impact or privacy assessment, cyber security rules, formal and informal guidelines” on how to safeguard privacy but yet carry out the work.