Blog
Building generative AI-powered public services: what have we learned so far?
Last week Faculty and Microsoft brought together a number of senior government leaders and digital & data specialists to share notes about what a year of generative AI implementation has taught us about successfully deploying these rapidly evolving technologies in a public service context.
The launch of ChatGPT over a year ago was a significant moment in the history of AI. The barriers to entry of using powerful AI dropped significantly. These models no longer needed an expert team to train them on one specific use case, but came ready formed to competently manage a bewildering range of general tasks, and they could be accessed through a simple chat interface and, later, an easy to use API.
Public adoption of ChatGPT surged at an unprecedented pace, surpassing 100 million users within months of its launch. The allure of its capabilities and user-friendly interface fueled a hype cycle of remarkable intensity. Despite initial concerns regarding data security, the practical utility became evident but apprehensions lingered about the responsible handling of sensitive data, prompting swift action from Government to offer guidance on secure usage. OpenAI’s connection with Microsoft and adoption into the Azure environment meant models could now be brought to the government’s existing data environments, rather than data being passed overseas.
Practical experiments within public services ensued, guided by internal government teams initiating smaller proof of concepts. And the centre of government has played a role to brigade some of these early examples and share learning.
After more than a year of collective experience with GenAI there are important lessons to be shared. While the barriers to entry of using GenAI have lowered significantly, making these tools work as desired in the real world remains hard. Without continued practical improvement the government risks falling rapidly into the ‘valley of disillusionment’.
At Faculty, our practical experience has shown that teams using GenAI need to:
- Get the engineering right: by adapting to get the most out of the Azure environment and designing services to manage hallucinations and unknown answers
- Getting the most out of model performance: by using specific data and small step-by-step prompts
- Solve an end-to-end problem for users: by ensuring we remember the hard-won lessons of putting users first and applying basic service design principles
We see use cases clustering into three distinct categories:
- One-off personal productivity. e.g. for use cases such as creating a set of interview questions or advising how to improve written material GenAI is now being safely adopted within existing day-to-day office tools such as ChatGPT, Bing Search, GitHub co-pilot and importantly many civil servants are likely to get co-pilot in their existing Microsoft Office365 environment.
- Routine processes and services in an organisation. e.g. GenAI enabled services that work standalone or as part of a wider software system to reduce the complexity, cost and ‘drudge’ for users in repeat administrative processes.
- Next-generation business models. e.g. we are currently in the GenAI equivalent of the ‘Web1’ era where we are making our existing world more efficient, rather than the ‘Web2’ world of the 2010s that introduced radically new business models like Netflix streaming. We can start to see a next generation of GenAI services that will use LLMs as ‘agents’ tasking other LLMs to carry out actions on our behalf, potentially disrupting search engines and even the web as we know it – and this will have significant repercussions for government.
Our overarching lesson for government remains just start. Building small, careful services and then iterating and learning from this is critical to developing the skills and confidence to make the most of this new era. Public service teams need that capability to avoid taking wrong steps in a market where promotion can override expertise. And senior leaders need to help their teams maintain their ambition and momentum – there are plenty of incentives to slow down, from imperfect data, concern about existential threats, and a general fear of change. But vital efficiency savings are on offer and the stakes are too high for government to fall too far behind.
Explore how we’re improving citizens’ lives by using data to its full potential here.
