Skip to content

Applications for GovTech and the Singapore Government

Current Applications

Agents and Agentic AI systems have already been developed in the Singapore government for some use cases as proof of concepts. Here are some examples of where they have been used successfully:

Gaining Deeper Insights into Customer Relationship Management (CRM) Data through a Multi-Agent Framework

Agency Level: 3

The data science team from GovTech's AI Practice forward-deployed to SkillsFuture Singapore set up a multi-agent system prototype to get deeper insights from CRM data by modelling AI agents as government officers each working on its own specialised task. In the framework, shown below, a user’s question would be taken on by multiple agents which would work together to provide a comprehensive and relevant answer grounded on data in SSG’s CRM systems.
SSG workflow

The set-up consists of two categories of agents, business agents, and data agents, which performs specialised tasks as below, sometimes with the assistance of tools like code executors and graph knowledge base query engines.

  • Business agents

    • deputy_director_agent: Perform final summarisation and consolidation reporting back to user
    • assistant_director_agent: Gatekeeper to prevent irrelevant questions from being processed
    • business_analyst_agent: Break down the question into sub-queries and propose solutions

  • Data agents

    • data_analyst_agent: Select and prepare python code to use tabular data for questions
    • python_executor_agent: Execute python code and return the outputs or any errors encountered
    • graph_analyst_agent: Select and prepare cypher query to use graph data for questions
    • cypher_executor_agent: Execute cypher query and return the outputs or any errors encountered

Agentic capabilities can be seen from how LLMs play the roles of humans and decide which courses of action to take at each stage, instead of following a pre-determined workflow. For example, depending on the nature of the query, the business analyst would choose either the data analyst or the graph analyst (or in some cases, both) to retrieve and analyse the data relevant to the query.

Using this multi-agent prototype, the team was able to glean contextualised and human-validated insights, which would typically involve human interventions by having agents make decisions at the high level, and resolve problems at a lower level. The agency level of this system is level 3 due to the use of multiple agents acting independently and autonomously to generate a high quality report using LLM outputs.

Fact-checking Bot Responses Using Agents

Agency Level: 2

Singapore's Ministry of Education (MOE) recently launched a GPT-powered Learning Assistant (LEA) for students. GovTech's AI Practice was tasked to help evaluate the bots responses for adherence to context and factual consistency. An agentic workflow was used for factual consistency checks as follows, taking reference from the Loki fact verification tool.

Fact-checking workflow for MOE

Given a response from the bot, a planner would assess if it consists of a single claim or multiple claims. If it consists of multiple claims, it would be decomposed into separate claims and each claim would be assessed for check-worthiness (i.e. whether it’s something that can be checked for factuality). Check-worthy claims are then handed off to an evidence assessor which uses web search and/or RAG tools to search the web and/or internal knowledge bases respectively, consolidates the retrieved evidence, and then decides if more searches are needed, until it can give an assessment.

Just like in the SSG application above, the course of action is not pre-determined (e.g. not all claims need to fact-checked) and the next action path is determined by the output of LLMs playing the role of humans. The agency of this system is Level 2 as LLM outputs determine both the control flow (e.g. does the bot's answer need to be decomposed) and which tools to call (e.g. to do a web search or an RAG search on an internal knowledge base).

Effective Web Search using Deep Research Agents

Agency Level: 1

To search and obtain public information for specific business requirements, the data science team from GovTech's AI Practice has introduced an automated workflow using AI Agents. The agent gathers publicly available information using a tiered research process: It begins with the Perplexity Search tool to review online sources, and if the initial search does not yield sufficient details, it will route to another path. This other path would consist of a loop of additional Google Searches and web crawling with modified search terms at each iteration until there is sufficient information. The workflow is as shown below:

Deep research workflow

This system has an agency level of 1 because the use of tools is already pre-determined (use perplexity first, followed by Google Search if there’s not enough information). However, LLMs analyse the search results and determine if there is already enough information gathered, or deeper research needs to be done, influencing the control flow.

Other Possible Applications

Beyond these use cases, the Singapore government can also leverage AI agents to enhance public services across a wider range of domains. Some specific applications are given below:

  • Public Service Automation & Improvement: Handle routine inquiries from citizens, integrating with knowledge bases of government policies and procedures (e.g., "Smart Citizen Assistant"). For example, an agent could autonomously answer queries, or even execute instructions, on application for permits, payment for taxes, or requests for social services, freeing up human officers to focus on more complex cases.
  • Improving Citizen Interactions: Enable chatbots to perform transactions like filing taxes, applying for grants and performing CPF withdrawals through a chat interface. This can be done by enhancing existing chat assistants like VICA and AIBots with agentic capabilities to perform real-time support and transactions. This would provide citizens with a more convenient and efficient way to interact with the government.
  • Job Matching and Placement: Agents could be used to enhance current career advancement platforms (e.g. Career Kaki) to better match job seekers with suitable positions based on their skills and preferences. Tools can be developed to allow such platforms to integrate with job portals and government databases (e.g. MyCareersFuture) through API calls. Further steps can be taken to not just provide job recommendations and placement, but also to provide upskilling guidance for better employment opportunities, depending on the jobseekers’ dynamic requirements. This would help to reduce unemployment, underemployment, and improve workforce productivity.
  • Augmenting Policy Analysis: Multi-agent systems can be built to Identify relevant datasets, analyse trends, and generate reports for policy officers (e.g., "Policy Analytics Agent"). This would provide policymakers with better insights into the impact of government policies and help them make more informed decisions.
  • Cybersecurity and AI Safety: Agentic systems can be designed to probe vulnerabilities in existing systems, or perform safety testing on LLM-powered applications to enhance application safety. For example, red-teaming agents can be used in AI Guardian to autonomously figure out how to hack into a system, and the appropriate measures can be implemented to protect against such hacks. This would help to protect government systems and data from cyberattacks and ensure the safe and responsible use of AI technologies.
  • Simplifying Legal Processes: Multi-agent systems can be used to assist citizens unfamiliar with legal processes by extracting information from legal documents and providing personalised guidance (e.g., "Legal Assistant Agent"). For example, one agent can do translation for non-English speaking users, another agent can use search tools to look up the legal clauses relevant to the user’s query, and a third agent can be used to translate legal jargon into accesible terms. This would improve access to justice and empower citizens to understand their legal rights and obligations.

TL;DR

Protoypes have been successfully developed in the Singapore public service to demonstrate AI agents' abilities to reduce human efforts in various workflows. Agentic AI technology can potentially be extended to other use cases across the service to improve efficiency.