There has been much talk of how organisations can leverage AI to improve productivity and effectiveness, especially with the emergence of generative capabilities through Large Language Model (LLMs) and related technologies.
At StormID, we have looked at this opportunity through the lens of government organisations given that there has been good recent progress in evaluating the potential productivity impacts of LLMs within the public sector. And it’s why we are also so excited to have recently launched the Public Sector AI Challenge with Futurescot.
Undoubtedly, there are common data challenges across different sectors, such as health, pharmaceuticals, legal and financial services – and the public sector is no different in many ways, save perhaps for the word ‘public’ and the need for abundant caution when it comes to spending taxpayers’ money and the overarching requirement for good governance and responsible use of modern technology.
Nonetheless, we believe that the UK public sector faces an inflection point on the use of LLMs and we are at the beginning of a new wave of technology developments that could have broader implications for organisations than the internet itself.
For the purpose of this article, it would be useful to actually describe what an ‘LLM’ is and how organisations can use them to boost productivity. Firstly, the word ‘large’ self-evidently means a great deal of underpinning data being used for training purposes; ‘language’ refers to the fact that the model is about understanding and producing natural language and ‘model’ is the abstracted, mathematical process upon which the algorithm is developed.
LLMs are based on artificial intelligence techniques derived from neural networks, initially proposed by Google in 2017, and subsequently brought to the mainstream by OpenAI and its ChatGPT tool.
They are a type of general-purpose AI designed to learn relationships between pieces of data and predict sequences from a large knowledge-base of pre-set content. We can interact with this process by providing a “prompt” as an input and then gain remarkably good natural – and increasingly video and graphical – output.
Even at this early stage in the story, LLMs are already performing a wide range of economically useful office-based functions such as real-time information retrieval, auto generation of documents, reviewing large amounts of unstructured content to detect patterns and insights, synthesising data and interacting in a chatbot setting via natural language.
The public sector productivity puzzle
Recent UK Office for National Statistics (ONS) data has highlighted that total public service productivity only grew by an average of 0.2 per cent per year between 1997 and 2019, with several service areas completely static or seeing diminishing growth. Despite significant investments in digital transformation, the public sector has not taken much advantage of the changes in technology in terms of translating into enhanced productivity.
Productivity is about achieving more output with the same inputs and as AI technology has improved it has raised expectations of what government agencies can do. Recent research shows the average civil servant spends up to 30 per cent of their time on documenting information and other basic administrative tasks.
In its 2023 Autumn statement, the UK government announced that the potential productivity benefits from applying AI to routine tasks across the public sector were estimated to be worth billions.
This ambition is aligned to numerous studies in recent years which have argued that AI has the potential to generate significant economic impacts, with predictions of increases in labour productivity of up to 40 per cent.
How can LLMs support increased productivity?
The last 20 years of digital transformation have moved paper-based business and citizen processes into software-based solutions. However there remains large numbers of repetitive administrative tasks, which have until now, required human processing.
Such repetitive bureaucratic steps are usually based on understanding language and applying rules to unstructured data which makes them ideal candidates for LLM automation unlike traditional automation methods which require predefined rules and structured data formats.
The natural language abilities of LLMs make them highly compatible with this process, as they excel on deriving insight from large amounts of unstructured data. In addition, LLMs offer the potential to interact with databases, applications, and other software systems using natural language, enabling end-to-end automation of complex processes.
To illustrate the productivity saving potential, a 2024 Alan Turing Institute study estimates that the UK central government handles approximately one billion citizen-facing transactions each year across around 400 services. Among these transactions, approximately 143 million are complex repetitive transactions and 84 per cent are highly automatable using AI. Even saving just one minute per complex transaction could result in the equivalent of approximately 1,200 person-years of work saved annually, researchers found.
LLMs offer significant benefits for the public sector and its users, if government concerns around ethics, security, and safety are addressed, and pro-innovation regulation is established. Indications on this are positive and we believe that LLMs can reduce substantial administrative burdens on general civil service work, as well as many sector specific opportunities in education, healthcare and criminal justice.
It is also important for organisations to build trust in AI with their users and develop user-centric AI solutions by understanding their needs and pain points and follow responsible AI principles – which are outlined in Scotland in a National AI Strategy and AI Playbook. Ministers have further committed to transparency here in a welcome public sector AI Register, outlining each and every use case of AI by government and its agencies.
How can organisations see value quickly?
We’ve found that organisations can get quick value out of LLMs by identifying duplicative or time-consuming tasks within existing processes. In the short-term organisations should focus on low stakes processes, and where human intervention is easy but necessary. In productivity terms, these are examples of AI performing the automation of time-consuming tasks which frees up labour to be allocated elsewhere to perform higher value activity.
Examples of LLM use cases we’ve already introduced to government organisations include:
- Removing administrative burden – many civil servants have high caseloads and their jobs can be admin-heavy. We’ve applied LLMs to support in retrieving and summarising large amounts of organisational information in faster timescales, helping staff to focus on high value activity.
- Building evidence base for decision making – we’ve introduced LLMs which can perform data analysis cost-effectively, summarising insights from unstructured data or creating structured data.
- Targeting inefficiencies in existing business processes – we’ve introduced AI into existing workflows. In one example we built an AI tool which automates the forms creation process which previously involved a lot of manual work.
In addition, progress can be accelerated by utilising existing Microsoft365 licencing, Azure infrastructure, data governance and Data Processing Agreements already in place so organisations can quickly take advantage of a wide range of specialist and general purpose AI tools that can be cost-effectively used within Azure and Power Platform, including Open AI.
Whilst we expect a range of LLM options to be generally available to organisations in the coming years, Open AI accessed via Microsoft Azure is proving to be a compelling route for organisations to get started with LLMs. Already, 65 per cent of US fortune 500 are using Azure Open AI service.
Organisations have a large amount of data, often unstructured, within the Microsoft suite including SharePoint, SQL Server and Azure Storage. This is often in varying formats such as PDF, Excel, CSV, Word, PowerPoint, and image formats, such as scanned and handwritten documents and forms.
LLMs, using architecture patterns such as Retrieval Augmented Generation (RAG) with full access to relevant corporate data can interact with it like a chatbot and efficiently help search, synthesize, report on and generate content quicker than a human would otherwise do.
How productivity will increase over time
Even the most conservative application of LLMs is likely to create productivity benefits over the next several years with the automation of repeatable, time-consuming tasks that frees up labour time to be allocated elsewhere.
Longer term, as the general purpose LLM technology improves and evolves with custom applications for specific tasks or processes, they will inevitably become part and parcel of the way organisations operate. We believe AI assistants will eventually be able to reliably interact with end users and line of business systems without much human intervention, automating or partially automating a wide range of complex business processes and decision-making tasks. This could bring about a larger shift in how labour is allocated across public sector.
It will likely take several years before LLMs become mainstream within the public sector and given the diversity of enterprise and legacy systems, the adoption will likely be incremental, starting off with niche use cases where there are the right conditions for high levels of autonomy.
Improvement in general purpose LLM technology
The pace of improvement in AI models is also subject to much speculation. We anticipate that models will improve along the following lines:
- Multi-modal – Text-based AI has demonstrated high capability, but we’re seeing rapid developments in other media such as code, images, audio, and video input. For example, OpenAI’s Sora is already producing startling video generation results.
- Larger data analysis capabilities – LLMs will be able to analyse much larger amounts of data at a time, referred to as a ‘context window’. What this means in practice is that AI will be able to extract, analyse and summarise huge amounts of corporate data to spot trends, make predictions and provide insights humans cannot easily perform.
- Reasoning and explanation – LLMs are expected to improve around their ability to reason and therefore provide explanations for their responses which could prove highly valuable in decision support.
- Personalisation – We will likely see development of LLMs which can be easily customised for specific situations to provide personalised experience and learning based on your interactions with them.
Autonomous agents
We believe new innovations will enable creation of more sophisticated applications or ‘AI agents’ capable of handling intricate set of tasks rather than one off prompts, recently coined as agentic workflows. This lets organisations build in memory and context to allow for greater reasoning for end-to-end automation of complex processes which may involve interaction with several databases, applications, and other software systems allowing the AI agent to perform complex tasks with little human intervention. In this scenario, the human becomes the conductor, providing the AI agent with instructions and overseeing the outputs.
Decision support AI
Decision support AI is where AI is able to effectively support decision-making by predicting, recommending, or prioritising outputs based on a set of inputs. As LLMs are released with greater reasoning and explainability, we expect user confidence will greatly increase to allow for further decision-making support. We would still see a role for professional judgement, working with AI, in many circumstances.
Multi AI agent collaboration
Multi AI agent collaboration refers to having multiple language models or agents collaborate through interaction to complete complex tasks. For example, collaborating together in specialised roles, each AI trained on specific datasets could offer a powerful approach for automating complex tasks that cut across different specialist domains within or between organisations.
Smaller LLMs and edge AI
We are seeing the rise of smaller LLMs with a suitability for specific or niche applications, in particular domains or for tasks that differ from large, general-purpose models.
Some smaller LLMs can be downloaded as standalone models and then contained in phones or laptops. To save a round trip into the cloud, so called “edge AI” moves all the capabilities to the edge e.g. the user’s device. We are starting to see the technology used in this way in the latest round of competition between the big smartphone manufacturers.
This means that systems can be used offline, embedded, be air-gapped and for other circumstances where reliability on external LLMs is not desired such as in areas where high levels of sensitivity are required or for frontline workers. In addition, we are seeing developments in this area allowing models to be executed natively in a user’s web browser, allowing for specific LLM use cases to be embedded directly into existing web applications.
Generative UI
Generative UI offers the potential of generating user interfaces in real time by understanding user intents, needs, and context enabling more personalised, accessible and usable experiences.
In summary
LLMs have already moved very quickly from the fanfare of Open AI’s ChatGPT launch in late 2022 to widespread use and adoption across multiple sectors. The public sector has been understandably cautious in its approach to the technology, given well-documented constraints around consent for the responsible and ethical use of public data.
However, there are signs that some governments are moving ahead with its adoption. In Ireland, the Department of Agriculture and the Department of Transport have both experimented with LLMs and Romania launched its first government adviser service using AI last year.
There is little doubt that governments around the world will continue to embrace the technology, albeit at different rates according to their own particular socio-cultural and political contexts. With a rich abundance of technology talent and skills in Scotland, and a history of innovation, it would be nice to see us at the head of the adoption curve – willing to be bold, take risks and to take the public with us on this exciting and era-defining journey.
We’re very excited about its potential and would encourage any public sector organisation considering exploring AI to take a look at the Public Sector AI Challenge we have launched with Futurescot. If you can think of a suitable use case for the technology in your organisation, please do get in touch.