The Policy Innovation Lab has been actively trying to understand the opportunities for using AI in the public sector. As part of this effort, we curated a Public Sector Tech Watch catalogue of digital tools used in European governments earlier last year. This catalogue proved useful in several ways:
- It helped us see what AI tools are currently used by various governments in Europe.
- It helped us develop a tool that could be useful for other South Africans by curating a catalogue of international tools for South African relevancy, removing those items that we deemed irrelevant or impractical for South Africa and that were not AI-related.
- It helps inspire us to develop certain AI tools ourselves.
We published a report identifying the different categories of tools and highlighting those that had the greatest impact or that had the greatest potential for South Africa.
In December 2024, we updated our AI tools catalogue to keep up-to-date with this rapidly changing technological environment. Some of the new AI-based tools included come from the November Public Sector Tech Watch catalogue of European tools, but we also included new tools from the website of AI use cases in the USA public sector and some African AI tools that we think are noteworthy. This resulted in about fifty new tools being added to the catalogue.
A trend that we identified in this update was an increased usage of Large Language Models (LLMs) or AI tools that rely on LLMs. This technology is also the magic ingredient to most of the Policy Innovation Lab’s tools developed during our first year. We use LLMs to collect citizen-generated data, summarize research, compare policy documents and draft policy briefs. Our catalogue update includes a range of new LLM-based tools, such as in Spain where CIDOBOT, AI4Justice, and others are used to retrieve, summarise and effectively communicate complex legislative and government documents. While LLMs currently work best in English and Chinese, governments everywhere are investing in developing their own LLMs to support translation and access to services by having them available in local languages, including African countries such as Kenya, Eswatini and Ghana. In South Africa, private organisations such as Lelapa AI are developing South African language models, and other existing commercial products such as ChatGPT can communicate in some South African languages.
LLMs are rapidly advancing, with new models emerging almost weekly, each outperforming its predecessors in speed, accuracy, or task-specific capabilities. These tasks include text generation, summarization, and analysis. Remarkable strides are also being made in areas like reasoning, logic, and code generation by prompting chain-of-thought processes with LLMs. Models such as DeepSeek (designed for code and mathematical reasoning) and OpenAI’s o1 reasoning can solve complex instructions by breaking them down into simpler sub-instructions. One particularly impactful application of these advancements is in agentic systems. These systems use a wide range of improved text-generation and reasoning models to coordinate and plan complex tasks by breaking them into smaller, manageable components, each handled by specialized AI systems. A real-world example is self-driving cars, where different AI subsystems manage tasks like object detection, route planning, and decision-making, all working together to enable safe autonomous driving. While agentic systems are only as good as their weakest link, we expect that in the next catalogue update, they will be used more frequently.
At the Policy Innovation Lab, we continue using LLMs and other AI tools to develop digital technologies to support policymaking in South Africa. We are currently using LLMs to create a media analysis tool that will be able to identify local and international responses to policy proposals. Furthermore, we are going to make a platform for accessing our suite of AI tools securely, providing appropriate access based on log-in credentials, while tools such as the AI tools catalogue will remain accessible to the public.
The last couple of years have been groundbreaking in the uptake of AI tools, in particular, generative AI. While it is difficult to know whether 2025 will continue to see the same sort of improvements in the tools themselves, or in the way that we use those tools, we will be sure to keep you in the loop with how these technologies are used in the public sphere with our future catalogue updates.