Michael Padfield says the Civil Service must undertake transformation, not passively undergo it
A friend turning up to court with an AI assistant instead of a lawyer; a father trusting a chatbot over his family doctor; a mother confiding to an algorithm, not a therapist – what three years ago would have been the stuff of dystopian fiction is, for many, the Britain of 2026.
Outside of Whitehall, the rush to adopt AI is challenging the long-held relationship between citizens and the state. Large Language Models (LLMs) are providing credible alternatives to government services; they are enabling mass objection to government projects with floods of (AI generated) letters and, by threatening seismic changes to the labour market, may even be rewriting the social contract.
And yet in government, against this backdrop of extraordinary change, many discussions around AI have a narrow focus on getting officials to adopt commercial tools to achieve efficiencies.
This framing around “AI adoption” ignores the upheaval taking place in the real world and condones a future in which officials are passive recipients of technologies that offer only pockets of productivity in an otherwise unchanged system of government.
As the ground moves beneath our feet, government must undertake, and not undergo, this transformation. In the short term this will require providing system-wide access to the best tools and capabilities. In the medium term, government will need to shape the development of technology to meet citizens where they are, not where we want them to be. And in the long term, we need to build public institutions that serve a society transformed by abundant intelligence.
Access to tools and capabilities
Adoption focuses us on the user of a tool, conjuring up civil servants refusing to engage or lacking the training to make change. But in many cases officials don’t even have access in the first place. The AI Security Institute handles the most powerful models in the world before their public release and the Incubator for AI (i.AI) builds products with the latest tools; at the other end of the spectrum, meanwhile, prison officers are managing cell occupants on whiteboards and planning officers pore over hand-drawn maps to decide planning applications.
These are of course extremes, but even in central departments different technology estates prevent the simplest interoperability. Officials cannot collaborate on live documents, let alone access an AI tool built by another team. Transforming the state will be impossible while this is the case.
Part of the solution is building robust platforms for AI services. The Government Digital Service’s new internal Single Sign-On process allows users to log into tools from different departments. For the Incubator, this has enabled us to now offer tools like Minute (transcription) and Parlex (parliamentary analysis) across the public sector, triggering a significant uplift in usage. Consult (consultation analysis) will soon join them, having already been tested on more than twenty-five consultations and saved over twenty thousand hours of work.
Conceiving of “adoption” as a platform problem, as well as a tooling and training problem, is necessary if we are to harness the full potential of AI. Estonia is building towards an “agentic” state, in which AI agents can take actions on behalf of citizens and officials, meaning services adjust dynamically to user needs and policies adapt to real-world outcomes. This will only be feasible with a secure platform that can authenticate, authorise and audit the work of agents within clear legal frameworks.
While we can argue whether a truly agentic state is desirable, narrowly focussing on tools only available to officials will leave citizens having to navigate digitally fractured public services in which their needs are partially understood and the solutions inadequate.
Mass adoption, and the expected benefits that follow, cannot be delivered without building out the infrastructure to make tools available, scalable and interoperable, and will not happen in full without considering how we build securely for AIs as well as officials. These aspects must come into the discussion on AI adoption.
Shaping the development of technology
Though some remain sceptics, vast swathes of the public are already rushing to use AI. The Lady Chief Justice has commended how AI is enabling individuals to better represent themselves in proceedings. In health, OpenAI data shows that millions are turning to AI to diagnose issues. In education, a whole host of AI tutors are on the market.
This changes the debate from what problems should citizens solve with AI (they have already decided), to what level of capability should the state provide and how to shape what is on the market to best serve citizens. Adapting to the changing needs of the public, of course, is something the state has always contended with – digital services are still relatively new – but never on this scale and at this rate of change.
In high-risk areas like health, equipping expert doctors with the latest models may be more desirable than releasing a public-facing tool directly advising citizens. But lower risk areas may be different. In May 2026, the Government Digital Service released GOV.UK Chat to help the public navigate information on GOV.UK, whether that be understanding childcare entitlements and finding apprenticeships, working out first‑home schemes or retirement benefits.
More interesting is what the Government can do to shape the technology. Here are three ways. Firstly and most simply, we should make sure that every live government page is as up to date as possible, so AIs can call on the best information.
Secondly, if we can do this securely, we should make government data more available, so that companies can build tools better suited to UK needs. This may mean sharing anonymised data from court proceedings and planning decisions. We have had some success helping startups by putting UK law and the parliamentary record into machine-readable forms. The better data models can use, the more aligned they will be to UK interests.
Thirdly, we can set standards for providers to compete against. For example, in education, the Incubator is not only co-designing and testing AI tutors with schools, but is also building technical benchmarks, alongside teachers, so that companies can ‘hill-climb’ against government-set standards on safety and pedagogy.
Undergoing transformation is letting unproven tools loose on the general public. Undertaking it requires updating content, publishing data and stating our expectations to help the industry build better things.
What kind of society do we want to build?
A world of abundant intelligence changes everything. But a narrow focus on “adoption” risks hollowing out a process without ever asking what it was for.
We might build tools to analyse consultation responses, the responses themselves written by AI, without thinking about what democratic engagement was meant to achieve in the first place.
We might run AI decision engines over planning applications themselves drafted by AI and forget to consider the role of taste and aesthetics in architecture and design.
We might field AI tutors to coach students through AI-set and AI-marked exams, for careers in knowledge work made redundant by technological advances, without questioning what the classroom is for. Are we optimising for a career well started, or a life well lived?
None of this is an argument against building such tools. Indeed we are building them in i.AI. They solve problems and, done well, could pay dividends for years. But this cannot be “job-done”. We must keep asking what the process is for, rather than seeking automation for automation’s sake.
And there is a graver failure than hollowing out a process: failing to protect what should never be automated at all. We already observe certain ethical red lines, where we do not adopt a technology even if it is capable of the task, for example in autonomous weapons and in some surveillance technology. But as the technology progresses, the red lines may erode, the Overton window may shift and other areas may fall into scope. Does this apply, for example, when AI can decide a case more consistently than any judge?
You may think this is far-fetched, but experts suggest the length of tasks AI can complete autonomously is doubling roughly every four months. Even if this is wrong, we can confidently say that the services we design today will meet a very different world in five years.
Closing thoughts
Adopting tools without critical thinking will not lead us to where we want to be. In the near term, our efforts need less onus on adoption at the individual official level and more onus on platforms – making technology available, scalable and interoperable for use by humans and AI.
In the medium term, we should meet the market where it is heading and shape it, rather than buy what is off the shelf, or else we will fail to adapt to how citizens are using the tech in the real world.
In the long term (whatever that means in this ever-accelerating world), we as civil servants should support ministers to be opinionated about what public services are for and build our institutions to deliver on that.
None of this is free or easy: it asks for real investment, and for some institutions built for an age of scarce intelligence to change. Politicians will have to lead this. But the Civil Service must meet this moment with seriousness and radicalism, and be willing to discuss the deeply technical systemic changes required. In this way we can undertake and not undergo transformation – and provide the best possible outcomes for citizens.
Michael Padfield is Head of Strategy in the Incubator for Artificial Intelligence, the UK Government’s Applied AI team.
