Artificial intelligence in social security institutions: The case of intelligent chatbots

Artificial intelligence in social security institutions: The case of intelligent chatbots

Artificial Intelligence (AI) is making rapid inroads into the public sector as agencies pursue greater efficiency, greater quality and more personalized services for their customers. Social security institutions are no exception. While the applications of AI are varied – each with its own far-reaching implications – “conversational AI” or “chatbots” have been leading the way in terms of AI adoption by government agencies.

In a survey of 166 government agencies across the globe, chatbots emerged as the frontrunners with 26 per cent already implementing them, and another 59 per cent planning to implement them within three years (Figure 1). In a review of 230 AI-enabled public services across the European Union, chatbots emerged as the first choice, accounting for over one-fifth of the used cases (European Commission, 2020). The global conversational AI market, including chatbots and intelligent virtual assistants, is expected to to have a Compound annual growth rate (CAGR) of 22 per cent during 2020–2025, reaching almost 14 billion United States dollars (USD) (Deloitte, 2017).

Figure 1. Types of AI adoption across government
Figure 1. Types of AI adoption across government
Source: Gartner, 2021.

A chatbot (or virtual assistant) is an algorithm that conducts a textual or oral conversation. While chatbots are not really new technology – for instance, the first chatbot was already programmed in 1966 in order to discover if humans would be able to find out if they were talking to a person or a machine – the potential of chatbots is now considerably higher due to advances in AI-technologies and changing communication patterns (Van Noordt and Misuraca, 2019).

Chatbots are essentially computer programmes which are able to recognize the input from a user using pattern matching technologies, access information and reply with the information found in the knowledge database. While basic chatbots communicate through pre-programmed answers, the most advanced ones use AI, which enables machines to analyze and process the context of languages better (known as Natural Language Processing or NLP), which in turn allows chatbots to tackle more complex tasks and host more human-like conversations. Chatbots are increasingly being employed by governments to help manage large citizen contact volumes and help citizens navigate complex policies and legislation to ultimately access public services (Henman, 2020).

Driving customer-centric communications in social security institutions: Experiences of ISSA member institutions

Given the importance of client engagement to social security administration, it is no surprise that social security institutions are keenly pursuing chatbots. While chatbots can improve customer satisfaction, and improve operational efficiency, implementing a chatbot is not a linear process. The opportunities, costs and risks associated with chatbots can vary considerably depending on the implementation techniques, means of communication, intended objectives, operational capacity and communication channel (ISSA, 2021). The International Social Security Association (ISSA) promotes the responsible use of chatbots through its guidelines. In the ISSA Guidelines on Communication by Social Security Administrations (ISSA, 2019a), guideline 10 refers to the “strategic use of new communication technology” in all areas related to the use of chatbots in social media and messaging systems, and guideline 14 refers to client-centric information. The ISSA Guidelines on Service Quality (ISSA, 2019b), guideline 5 focuses on “understanding users’ needs and experiences” which equally applies to the development process underlying chatbots. Finally, the ISSA Guidelines on Information and Communication Technology (ISSA, 2022a) refers to implementing a service desk and request fulfilment processes including virtual service desks (guideline 17), and to the potential uses of emerging technology, that contextualize the use of chatbots as an application of artificial intelligence (guideline 96). The ISSA has also been facilitating conversations among member institutions to bring out the complex nuances around chatbots. Notably, a dedicated session on artificial intelligence at the 16th ISSA International Conference on Information and Communication Technology in Social Security highlighted the opportunities and complexities associated with AI-enabled chatbots. This article analyzes these experiences in the same way as an article (ISSA, 2021) that described chatbot implementations in Latin America.

National Employment Office, Belgium

The National Employment Office (Office national de l’emploi – ONEM) in Belgium set up a chatbot to ease contact centre pressures brought on by the unprecedented volumes following the COVID-19 crisis (National Employment Office, 2021 and 2022). The first chatbot, known as Marc, was rolled out on the ONEM website in May 2020. In its initial phase, the chatbot was designed to serve only one type of client request: It gave citizens rapid access to copies of the tax certificates they needed to submit alongside their tax returns. In May 2021, the chatbot’s capabilities were expanded considerably, and a new chatbot named Ori was launched. Based on the analysis of questions posed to Ori by clients, an upgraded version was rolled out in December 2021. It is now able to answer a range of questions relating to unemployment and career breaks. It also helps clients navigate the ONEM’s website with ease. Furthermore, it serves as a promotional tool to encourage use of the e-box, Belgium’s virtual, secure mailbox that enables the authorities to communicate safely with citizens. Most importantly, the chatbot remembers the context in which the customer is situated when she or he asks questions, ensuring that the chatbot can continue supporting the customer regardless of where and how the customer navigates the website. Finally, the themes covered by the chatbot are regularly updated based on the analysis of clients’ questions.

Social Insurance Institution, Finland

The Social Insurance Institution (Kela) in Finland set up two chatbots, Kela-Kelpo and FPA-Folke, to help clients find information about benefits on Kela’s self-service web portal (Social Insurance Institution, 2022a and 2022b). Based on natural language processing, the chatbots speak two languages – Finnish and Swedish – and they also understand English. The Kela initially launched chatbots in 2017, which were augmented with information on a growing set of benefits over 2017 and 2021. The various chatbots were consolidated as Kela-Kelpo/FPA-Folke in 2020 to prevent clients from having to move between multiple chatbots to know about different benefits. These conversational chatbots make it easier to discover and interpret information and to complete benefit applications. Further, the consolidated chatbot provides customized tips based on contextual variables as clients fill out applications for benefits such as parental benefits, social assistance, and so on. During the COVID-19 crisis, a dedicated chatbot was temporarily deployed to address queries on COVID-19-related social assistance.

German Federal Pension Insurance

The German Federal Pension Insurance (Deutsche Rentenversicherung Bund – DRV-Bund) introduced a chatbot to answer questions frequently asked by insured persons (German Federal Pension Insurance, 2021). The main objective was to ensure access to information 24/7, in addition to making a small start towards building a much more extensive chatbot. The chatbot simulates natural language based on artificial intelligence technologies, and has the potential to free up staff time towards more complex matters. The chatbot is in its initial phase of implementation, and as a result, usage remains nascent at 5 per cent. The DRV-Bund aims to extend the chatbot’s capabilities by embedding a form-filling aid into it in the near future.

Employees Provident Fund, Malaysia

Malaysia’s Employees Provident Fund (EPF) introduced ELYA (EPF Loves You Always), a bilingual virtual assistant using Natural Language Processing and supported by Live Chat (Employees Provident Fund, 2021a and 2021b). Although the EPF had a contact centre in place to handle queries, the daily volume of 5,000 calls exceeded its capacity of 4,000 calls per day, leading to 25 per cent of calls being dropped. At the same time, 82 per cent of the enquiries had answers already available on the EPF website, implying inefficient use of the contact centre. A further survey indicated at 55 per cent of the clients found it difficult to navigate EPF’s website. Therefore, ELYA was introduced to support clients to self-discover information and thereby reduce contact centre volumes.

The launch of ELYA was preceded by a detailed analysis of problems and needs in 2017–18. A basic bot was first introduced in 2019–2020, followed by the current conversational bot in 2021–22. The EPF plans to extend ELYA’s capabilities further over 2023–2024 to enable the bot to provide advisory information. ELYA is conversant in both English and Bahasa Malaysia, providing conversational and interactive queries about 30 EPF products and services. It is available 24/7 on the EPF website and is backed by real-time escalation to a human agent during working hours. ELYA draws upon a representative knowledge base that has been carefully consolidated based on inputs provided by customer agents, customer emails and common queries received by the contact centre. The ISSA Guidelines on Communication by Social Security Administrations, in particular Guideline 14. Client-centric information, was referred to in designing ELYA. Further, it champions the four principles in the ISSA Guidelines on Good Governance (ISSA, 2019c), i.e., transparency, predictability, participation and dynamism, and the use of clear, simple language, with a focus on user-centric platforms.


Table 1 summarizes the results these institutions have been able to achieve through chatbots.

Table 1. Results achieved through chatbots
Institution Results obtained
ONEM, Belgium
  • 16,833 people made use of the chatbot, which is equivalent to nearly 9 per cent of the people visiting the website (Sep 2021)
  • 18,616 chats were recorded (Sep 2021)
  • 55,275 messages (questions) posted, of which only 616 of these messages could not be correctly interpreted by the chatbot (Sep 2021)
Kela, Finland
  • 64,372 total conversations with the chatbot (2021)
  • 108,817 questions asked (2021)
  • Conversation quality at 89 per cent and positive feedback at 40 per cent
  • During its existence, the COVID-19 chatbot had 18,678 conversations and answered 31,567 questions.
DRV-Bund, Germany
  • Still in early stages, results awaited
EPF, Malaysia
  • 1.6 million sessions recorded between June 2020–March 2021, with an average 6,100 daily sessions with a conversation duration of 12 minutes, contributing to high client satisfaction (4.1/5) with EPF’s overall digital engagement strategy.

Critical success factors

Translating complex administrative and legal information into conversational content requires rigorous design, training, and testing. The ONEM in Belgium had customer development specialists provide input and continuously test the chatbot to ensure that the language resonated with typical clients. The EPF in Malaysia tested the comprehension and accuracy of the chatbot through 200 testers over alpha and beta stages of product development before releasing the chatbot.  Further, humanizing the bot and giving it a personality helps the user connect with it at an emotional level. In Germany, the DRV-Bund found that many users were put off by the name “chatbot”. Cognizant of this challenge, the EPF in Malaysia designed a specific personality for ELYA that appealed to clients.

Continuous maintenance and improvement are critical to sustained user adoption. Evaluating conversation quality and accuracy against predefined benchmarks is essential. The ONEM in Belgium conducted daily analysis of the answers provided by the chatbot to correct erroneous answers. In the case of Kela in Finland, if the chatbot is unable to answer any questions, the staff publish answers quickly depending on the expected frequency of the question. In Malaysia, the EPF releases additional Natural Language Processing features in batches based on analytics on customer enquiries.

Dedicated staff and cross-team collaboration both underpin successful chatbots. The ONEM in Belgium has a designated chatbot manager who leads development and training. The chatbot manager has a healthy interest in IT, but his/her core expertise is an in-depth knowledge of the services and products offered by the ONEM. The chatbot manager brings together complementary expertise from other teams such as IT, customer development, and so on. In Finland, Kela employs bot whisperers, who are essentially people that create the dialogue, train the AI and maintain quality, in collaboration with customer development specialists. In Germany, the DRV-Bund has an editorial team who are recruited from the administrative staff and given editorial training. In Malaysia, the EPF has deployed a multi-disciplinary team comprising bot developers, conversation designers, bot trainers, system admins and AI analysts.

Infrastructure investments are equally important as staff investments in running chatbots effectively. The DRV-Bund in Germany looked at various options for integrating its chatbot into its website. Existing technology stack and expected scale are key factors influencing technology choice. In DRV-bund’s case, using technology from a major cloud provider cuts costs.

Finally, it is important to stress that chatbots are one among many channels. Effective client engagement by social security institutions requires a suite of digital and physical channels, with each channel bringing in unique and complementary capabilities. For instance, the Kela combines its chatbot with 147 citizen service points, 79 municipalities with teleservice-enabled points, e-services, and a contact centre. Especially in case of the web chatbot, the Kela ensures that it does not merely repeat information on the website, and instead enhances the utility of the website by giving clients additional details and examples. Over time, as chatbots improve, they may reduce traffic towards other channels, which needs to be borne in mind in resource planning and allocation.

Final remarks

According to some estimates, by the end of 2022, on average, people will be talking to bots more often than they talk to their own spouses (Deloitte, 2017). Therefore, chatbots are going to become integral to social security institutions’ overall client communication strategy.

The ISSA members’ experiences outlined in this article highlight several fundamental lessons for social security institutions. First, chatbots are complementary to existing digital and human channels: they may replace some channels, improve others, and in some cases, other channels will co-exist as chatbots may not be desirable due to privacy and legal frameworks. Second, the development cycle associated with chatbots cannot be underestimated. It would serve institutions well to start small, with a restricted scope, as reiterated by ONEM in Belgium and EPF in Malaysia. Thirdly, the sheer complexity of developing algorithms to translate administrative information into conversational content – while taking into account the clients’ context – requires a highly iterative process for training and releasing of new NLP machine learning solutions over short bursts of time, potentially daily. This requires investment in dedicated human resources, usually specialized customer service staff for daily analysis, review and maintenance. Fourth, chatbots need to be built based on the customer perspectives, which means regular engagement with customers for feedback. Finally, chatbots bring forth new legal and ethical concerns for social security institutions (Henman, 2020). For instance, ONEM had to ensure that the bot did not collect any personal data in accordance with strict privacy legislation. As the AI becomes more sophisticated, a key concern is that it could learn harmful behaviours from its interaction with clients (ISSA, 2020).

As the experiences above show, social security agencies that proactively manage for any liability and client protection issues are better placed to be able to fully reap the benefits of intelligent chatbots. The ISSA supports member institutions to successfully adopt chatbot technology and address the implementation challenges, notably those related to AI application, through guidelines (ISSA, 2022a), sharing institutions’ good practices (ISSA, 2020 and 2022b) and organizing meetings.


Deloitte. 2017. Conversational AI - Five vectors of progress. Chatbots. London.

Employees Provident Fund. 2021a. ELYA: The bilingual virtual assistant of the Employees Provident Fund (Good practices in social security). Geneva, International Social Security Association.

Employees Provident Fund. 2021b. EPF Chatbot: A case study in Malaysia (ISSA Webinar: Improving customer services through intelligent chatbots, 8 December). Geneva, International Social Security Association.

European Commission. 2020. AI watch - Artificial intelligence in public services. Luxembourg, Publications Office of the European Union.

Gartner. 2021. Gartner says government organizations are increasing investment in AI, but their workforce remains apprehensive. Stamford, CT, Gartner Inc.

German Federal Pension Insurance. 2021. Using chatbots to improve e-services: What we learned at the ZfA division of DRV-Bund (ISSA Webinar: Improving customer services through intelligent chatbots, 8 December). Geneva, International Social Security Association.

Henman, P. 2020. ”Improving public services using artificial intelligence: possibilities, pitfalls, governance”, in Asia Pacific Journal of Public Administration, Vol. 42, No. 4.

ISSA. 2019a. ISSA Guidelines on communication by social security administrations. Geneva, International Social Security Association.

ISSA. 2019b. ISSA Guidelines on service quality. Geneva, International Social Security Association.

ISSA. 2019c. ISSA Guidelines on good governance. Geneva, International Social Security Association.

ISSA. 2020. Artificial Intelligence in social security: Background and experiences. Geneva, International Social Security Association.

ISSA. 2021. The application of chatbots in social security: Experiences from Latin America. Geneva, International Social Security Association.

ISSA. 2022a. ISSA Guidelines on information and communication technology. Geneva, International Social Security Association.

ISSA. 2022b. ICT response to COVID-19: Leveraging accelerated digital transformation to build better and more resilient social protection systems (ISSA Technical Commission summary report). Geneva, International Social Security Association.

National Employment Office. 2021. Chatbot Ori (ISSA Webinar: Improving customer services through intelligent chatbots, 8 December). Geneva, International Social Security Association.

National Employment Office. 2022. Creation and launch of a chatbot on the National Employment Office website (Good practices in social security). Geneva, International Social Security Association.

Social Insurance Institution. 2022a. Kela chatbot: Bilingual help for online customers, 24/7 (Good practices in social security). Geneva, International Social Security Association.

Social Insurance Institution. 2022b. Kelas bilingual chatbot (Presentation at the 16th International Conference on Information and Communication Technology in Social Security, Estonia). Geneva, International Social Security Association.

Van Noordt, C.; Misuraca, G. 2019. New wine in old bottles: Chatbots in government (Conference paper, 11th International Conference on Electronic Participation (ePart), San Benedetto Del Tronto, September).