Managing data with generative AI

Data management encompasses the organisation, maintenance and governance of data.¹⁹ For agencies using generative AI, this is most relevant to making it faster and easier for public servants to find and retrieve data. Around a quarter of the data projects were in this category.

These solutions often combine generative and non-generative AI capabilities. For example, take a simple tool which retrieves and summarises relevant data for public servants to analyse. The tool will use a non-generative AI method to identify the most relevant information from a given knowledge base (in most cases with text data, this is a semantic search method). Once the relevant information is extracted, the LLM will summarise that content to answer the query.²⁰

In other cases, some tools use LLMs to translate natural language requests into computer code to complete searches of structured and unstructured data. This is similar to some of the code generation tools described in Briefing #3. The data used in these projects includes interview transcripts, data catalogues, supply-chain news, customer experience metrics and maintenance logs. 

The benefit of these applications should be understood in terms of their impact on public servants’ analytical outputs. If generative AI makes it easier and faster for public servants to engage with data, they can do more rigorous and extensive analytical work, which can lead to better insights to support sharper decision-making.

Analysing data with generative AI

Analysis is about turning data into insight. It involves identifying patterns, testing hypotheses and uncovering meaning.²¹ Around a quarter the data projects we identified are designed to help public servants identify patterns from large datasets. These tools accelerate the process of generating insights, enabling faster interpretation and decision-making.

Again, AI-powered analytical capabilities are not new for government agencies,²² but LLMs can offer much greater efficiency and flexibility. For example, an agency needs to evaluate whether submissions for a public consultation are for or against the proposed rule change (a method known as stance detection).²³ The agency could manually train an AI model to detect positive or negative responses by giving it lots of examples. However, that model could only be used in that specific topical context. LLMs, by contrast, can adapt to diverse topical contexts because of their extensive foundational training, making them more flexible.²⁴

As well as analysing these data, generative AI tools can also communicate their insights fluently. For example, the U.S. Social Security Administration uses customer service transcripts to identify bottlenecks in their policies and processes. And the U.S. Department of Veterans Affairs uses AI to evaluate user research transcripts to create personas and user journey stories. 

These tools, however, are not without risks and limitations; they can miss important context or over-generalise and miss subtlety. For instance, in one study reported by The Guardian, an LLM was used to evaluate submissions to a government public consultation on AI policy. The model correctly identified the stance of a submission but missed significant detail about the conditions of that position.²⁵ That places added importance on public servants’ ability to engage with AI generated analysis in a critical way. They need to be able to verify, interrogate and systematically evaluate the quality of the insights.²⁶ This is also true for AI tools that support evaluation and decision-making in government (presented later).

We also see agencies using LLMs to apply a layer of analytical filtering on large volumes of information, so public servants can more easily spot important signals. An example of this is the U.S. Department of Health and Human Services’s NewsScape tool, presented below.

Generative AI-powered research tools

Research tools help public servants improve their awareness and understanding of detailed information related to their work, making the process of discovery and learning more efficient. These tools are predominantly being used in government legal work and legislative drafting. In some cases, these tools are designed to assist public servants in customer service roles, who need to quickly understand the details of official policies and guidance in response to a live query from service users. 

One example of this application design is the U.S. Department of the Interior’s ‘SOCS document analysis’ tool. The solution uses retrieval augmented generation to provide public servants with access to a large portal containing environmental impact statements, environmental assessments, biological assessments, legislative policy documents, environmental guidance and environmental studies. The tool makes this vast store of information easily reusable in new workflows. For example, public servants can easily extract historic environmental impact assessments if they are initiating a review in the same geographic location. These research tools can support public servants to work in a faster, more consistent way while reducing the risk for duplicated effort. 

There were also several applications designed to conduct scientific literature reviews, particularly in the health field. Again, these applications may present governments with a trade-off between speed and quality. A recent UK evaluation study of AI generated literature reviews in government found that they can be completed approximately 23% faster overall compared to human-only reviews. However, the study also highlighted that AI-generated initial drafts were of lower quality and required significantly more revisions than human-produced reviews.²⁷

The emerging role of generative AI in government decision-making 

Apolitical’s analysis found a small number of applications which use generative AI to directly support public servants’ decision-making in areas like project management, service delivery and evaluating applications. None of these use cases automate the actual decision. Instead, AI is used to make recommendations which are presented for human review.

These tools represent the growing presence of AI in how agencies make decisions of all types – from regulatory oversight and programme management to direct service delivery.²⁸ 

• Examples include the U.S. Department of Labor’s AI assistant which is used to evaluate business cases for potential AI use and make recommendations about which programmes are more or less suitable for AI’s capabilities. 

• Another example is the U.S. Department of Health and Human Service’s DECIDE tool, which automates the process of manually reviewing grantees’progress towards achieving their health disparity reduction goals. Instead of public servants manually verifying grantee submissions, the tool automates the extraction of information and data from a range of sources to help speed up the evaluation decision about the grantee’s progress. 

Public servants are often the ‘human-in-the-loop’ of these processes. To fulfill that role meaningfully, governments need to equip them to engage critically with AI outputs, which includes (among other things) understanding sources of data, an insight to how models work and being able to probe outputs for bias and blind spots.²⁹ Data from Apolitical’s AI Readiness Check shows that 47% of public servants lack confidence in their ability to interrogate the data which powers AI systems, demonstrating why building this critical awareness is so important.³⁰

Up next...

The penultimate briefing in this insights report focuses on how governments are using generative AI to respond more quickly to the demands of public servants and citizens. These applications provide an early look at how agentic AI is emerging in government and how this capability could reshape the relationship between governments and citizens.

Generative AI as a responsive capability for government

In a few short years since the release of the first generative AI tools, a lot of interest has focused on how this technology could evolve to work more intelligently and autonomously in government. This ‘agentic’ mode is widely seen as the future of AI innovation in government and industry.³¹ 

That future is not here yet, but Apolitical’s review found that governments are certainly trying to use generative AI as a responsive agent in internal operations and some citizen-facing services, primarily through the interface of a chatbot. What is unique about these projects in our review is that they are not designed into pre-determined workflows. Instead, they function as an ‘on-demand’ capability that can serve unpredictable need, helping governments to close gaps between demand and supply for information, potentially increasing efficiency. These applications could be the foundation for more sophisticated ‘agentic’ tools in the future.

The growth of generative AI chatbots in government

There were 56 “chatbot” projects in our sample, nearly 20% of the total, which shows how this has emerged as the dominant design solution for government AI application in recent years. Most of these tools are geared toward internal use, but more agencies are becoming increasingly ambitious with citizen-facing applications.

Internal applications

The tools developed for internal use are often designed to help public servants quickly access detailed information about internal policies related to HR, operational and administrative procedures, as well as archival records in a few cases.

They create a ‘self-serve’ capacity within the agency, so that individuals and teams can quickly get a good understanding of an issue, without having to do the research themselves or ask a colleague to provide a briefing. The U.S. Federal Emergency Management Agency estimates that these self-service tools can reduce engagement time between teams by 80-90%.³² Likewise, the UK Department of Science, Innovation and Technology’s ‘AskOps’ tool is estimated to solve 90% of the frequently asked questions from colleagues that were previously served by human responses.³³

Internal AI chatbots could help governments in several ways.

• Enhance and accelerate collaboration: Apolitical’s research with public servants consistently shows that siloed working and a lack of collaboration across teams is a major source of frustration.³⁴ A lack of access to information can lead to duplication and misalignment. By reducing friction (including time, cost and outreach), these simple internal tools can grease the wheels of collaboration and quietly elevate internal efficiency within agencies.

• Improve information discovery: Delivering these tools at scale could promote much better coverage of data and information to exponentially improve access and discoverability.³⁵ This could be equivalent to governments having their own internal “AI app store,” copying the decentralised approach of Singapore’s AI Bots programme, which has yielded 12,000 AI tools in just 6 months.³

Citizen facing applications

Overall, only a small proportion (around 10% in our sample) of the generative AI solutions in government currently interact with citizens. Concerns around accuracy and reliability may have slowed this innovation while the technology is still quite new.³⁷ 

The technical design of these projects is consistent. The majority of citizen-facing chatbots use retrieval augment generation [RAG] to help ground the chatbot’s outputs in approved sources. Some agencies apply a more intricate technical design to minimise citizen’s interaction outputs purely generated by AI. For example, the UK Foreign, Commonwealth & Development Office consular assistance tool uses a large language model [LLM] to match user queries with pre-written templates and suggested reading. This means the user does not engage with any AI outputs and instead the technology is used to interpret the intent behind the user’s query and match that to the most relevant template.³⁸

A small number of projects go beyond ‘FAQ-style’ interactions. These use cases use generative AI to automate more spontaneous and relational interactions with users or citizens.³⁹ For example, the UK environment agency’s ‘Hello Lamp Post’ tool enables citizens to get tailored information about local projects, flood resilience and environmental schemes, as well as provide feedback and participate in citizen science initiatives.⁴⁰ Again, to manage the risk of inappropriate or inaccurate responses, the chatbot's responses are limited to pre-approved content; and the agency has used prompt engineering and set strict topic boundaries to manage the LLM’s output. In Singapore, the Ministry of Education has created a generative AI learning assistant for children, hosted in its virtual student learning space.

How chatbots could evolve in government

Informational chatbots have the potential to automatically solve citizen queries before they engage with the government. This can help them to become more informed users of public services, which benefits everyone. The relational tools have the potential to expand capacity, improve access and increase citizen engagement. However, there are open questions about how they could reshape the relationship between government and citizens.

Governments may face pressure to enhance generative AI search and response capabilities for public services. There is growing evidence that generative AI is reshaping internet search behaviours.⁴¹ This is already impacting how people answer questions online, including ones about government policies and procedures.⁴² If citizens increasingly experience search through the filter of generative AI, this will impact their expectations for public services and governments will need to adapt to meet this need. The same has been true for building mobile-friendly digital service experiences over the last decade.⁴³

A glimpse at generative AI agents in government

There is a lot of focus on the future of agentic AI in government. Apolitical’s review shows that public sector agentic AI is at a very early stage, even in highly advanced digital ecosystems, as reflected in wider research.⁴⁴ The most sophisticated examples identified in our sample were prototypes developed by GovTech Singapore, but these had limited usage. The U.S. Department of Energy’s Security CoPilot uses an off-the-shelf tool to support real-time threat monitoring and automated incident responses.⁴⁵ 

What is ‘agentic AI’?

AI agents are specialised software systems that can use autonomous decision-making to complete specific tasks. These could range in complexity and significance from scheduling team meetings to automating passport renewals. When several agents work together to achieve a larger goal, this is defined as an agentic system. Cutting edge agentic systems are capable of working through complex tasks and interacting dynamically with the environment around them using a variety of tools.⁴⁶ 

How could they work in government?

Early agentic systems in government show how these solutions can be designed to create small ‘AI teams’ that work together to complete workflows that might have previously required several public servants.⁴⁷ 

It’s worth noting that, as yet, we do not see any agentic systems integrated into public services or making decisions that would impact citizen rights. This mirrors the broader trend around limiting generative AI to internal administrative operations. The infographic below, adapted from GovTech Singapore’s original case study, provides a detailed illustration of how a small “team” of generative AI-powered agents could work through a multistep workflow, drawing on multiple data sources, to answer a public servant’s query. The prototype was developed for Singapore’s Skills Agency, as a tool to generate data insights from customer relationship management (CRM) records.

The infographic shows seven agents working together to complete one workflow. Business agents fulfill an orchestration role, making judgements about what needs to be done and engaging with the user. These business agents commission the data agents to do analytical tasks, mimicking the delegation of responsibilities in a human team. Each on their own, none of the AI agents have unique capabilities dissimilar from the projects described elsewhere in this report (e.g. filtering questions, writing code, querying databases and summarising information). The added value is their ability to choreograph their actions without human involvement.⁴⁸

Governments around the world are eager to experiment with agentic AI but data readiness, trust and governance are important hurdles to overcome.⁴⁹ There are also important ethical considerations and operational risks linked to managing the accuracy, reliability and explainability of agentic systems. Ensuring human oversight and control is important for retaining accountability of these systems.⁵⁰