What should the UK government do if they truly want to harness AI for policymaking?
www.nesta.org.uk/blog/harnessing-ai-for-policymaking/
www.nesta.org.uk/blog/harnessing-ai-for-policymaking/
In a viral tweet from 2019, former director of design at the Government Digital Service (GDS), Ben Terrett, stated how there should be: “no innovation until everything works”. Our experience working on Policy Atlas has led us to the same conclusion for AI in policymaking: government cannot fully harness the power of AI until we have fixed the foundations.
As IBM and others have asserted, AI is not a magical solution that can be sprinkled on top of a crumbling base. Rather, AI functions like a trojan horse: the promise and opportunity presented by this glittering new technology can be used to ‘smuggle in’ overdue improvement to our existing digital infrastructure. While it’s dangerously easy to be seduced by the shiny promise of AI, we should remember that more often than not, it’s the boring, everyday stuff that can be most important for innovation. For example, while the Cabinet ‘Write Round’ process may appear as a bureaucratic hurdle, it often provides the critical 'catch' that prevents flawed policy. This was the case for the 2018/19 extension of the plastic bag charge, when this process corrected a critical design flaw in the initial proposal (the initial exclusion of small retailers), ensuring the policy achieved its intended scope and impact.
The policymaking process is no different. For AI to be useful, we need to get the basics right.
So how should our government navigate the gulf between the promise of AI and our patchy public digital infrastructure? Our recent workshop - PolicyOS: Harnessing AI to improve policy design - convened policy leaders, AI tool developers, and researchers to map common policymaking scenarios as detailed journeys. This helped us to identify challenges and propose viable opportunities for harnessing AI.
Below we share our five recommendations for what the government should do if they truly want to harness AI to transform policymaking.
Auto industry pioneer Henry Ford allegedly said: “If you ask people what they want, they would say ‘faster horses’.” But Ford understood that a motor car wasn't just an improved version of a carriage; it was a fundamental reimagining of transport.
Advances in AI offer governments a similar, transformative opportunity. Governments must actively avoid the "efficiency trap," where the height of ambition is simply delivering the status quo cheaper and faster. Using AI to take notes or polish emails is useful, but can’t we do better?
AI should be the catalyst to fundamentally reimagine the entire policymaking process.
To realise these new capabilities, the government needs to deploy the same level of imagination and ambition into its digital infrastructure as it does into the AI applications themselves. This means asking radical questions about the how, such as: “what would a whole-of-government ‘brain’ look like?”, or “could we create digital twins of the public sector for dynamic policy testing?”. We need to shed the instinct to ‘make X process faster’ and instead question what the root purpose of the process is - allowing space to achieve the end goal via different means.
For example, across government, key intelligence is moved from expert policy and delivery teams at the front line towards central decision-makers through ‘commissioning processes’ - running from secretariat or private office teams and out through departmental networks. Instead of asking "how can AI make commissioning faster or higher quality?”, AI innovators should first question “what is the purpose of commissioning?”. Following this, we might then ask how to ensure the most important decisions have the best available intelligence behind them, and how AI could help.
The government must be willing, at all levels, to embrace this fundamental change and overcome the historic resistance to transforming how policy is designed, tested, and implemented. Some first steps to achieving this goal might include:
Building on the AI Exemplars programme, the government chief AI officer should be mandated to sponsor a 'zero-base review' of how a key government system (such as the benefits system or the justice system) can be reimagined from the ground up to harness AI. Rather than just adding an AI 'bolt-on' to current workflows, this review should use strategic design and foresight to present a vision for a natively digital system. This should be presented publicly - perhaps through a prestigious lecture - and privately to the prime minister and permanent secretaries within six months.
The GDS recently championed a portfolio approach to public sector innovation, where multiple ideas are initially pursued which are then whittled down through testing to find the most impactful ones. To take this one step further, the government should launch a service design 'challenge prize' to redesign a key government process. This would open up the 'government machine' to outside researchers and developers, ensuring that our digital infrastructure isn't just designed by those already inside the silos, but by the best minds in the wider AI ecosystem.
The 2025 Spending Review identifies 'reforming the state' as a priority, supported by a £3.25 billion transformation fund. To avoid the efficiency trap - where funding is absorbed by patching legacy systems - the government should adopt an 80/20 innovation split. At least 20% of transformation funding should be ring-fenced for high-ambition 'moonshots' that aim to improve performance by at least tenfold. This scale of improvement will likely only be achieved through reimagining systems, rather than incremental add ons. To manage this, the Treasury could adopt a ‘10x impact standard’ inspired by venture capital, and evaluate projects on their potential to achieve orders-of-magnitude shifts.
One of the biggest pain points for policymakers is limited institutional knowledge management across government. Our user research confirmed that policymakers urgently need AI tools to support them to effectively leverage historic internal data, expertise, and archived documents. They want quick answers to questions like: “what’s been tried before?”, “by who?” and “why did it succeed (or fail)?”.
However, when data is fragmented, domain expertise is siloed, and for processes relying on tacit knowledge, an AI tool will simply reflect this, potentially accelerating the wrong policy decisions. In other words, “garbage in, garbage out”. This foundation must be fixed before meaningful AI adoption can succeed.
For example, driving the government's Opportunity mission is complex when child poverty expert policy teams, and various levers for change, sit with the Department for Work and Pensions. But the majority of services and touchpoints with children and their data live in the Department for Education. Enhanced knowledge management and access governed by need, and enabled by AI, could revolutionise how policy is made and how services are delivered.
Therefore, we propose that if the government were to only make one ‘AI’ investment for policymaking, it should be an overhaul of document storage and cataloguing. Knowledge management has much less sparkle than a ‘whizzy AI bot’ but is, unequivocally, the most important foundation of effective AI adoption. The government’s AI Opportunities Action Plan proposes creating a national data library to "responsibly unlock both public and private data sets", as well as a single "AI knowledge hub" to publish best-practice guidance, results, and open-source solutions accessible across the public sector. Both of these solutions could serve as the vehicle for kickstarting this overhaul.
Quick wins to address this challenge might include:
Mapping the people, knowledge and documents in existence would help policymakers quickly find relevant expertise and prevent the loss of institutional knowledge due to high staff churn.
Make data AI-ready by developing common standards for data storage, tagging, and structure across departments, with secure access tailored according to security clearance level. This would further enhance the AI Opportunities Action Plan proposal for infrastructure interoperability, ultimately leading to enhanced “digital plumbing” across government.
With dedicated resources for maintaining, curating and governing knowledge outputs to ensure data quality and accuracy. Skilled knowledge managers are the librarians of the future and hold the keys to unlocking transformative AI.
Many initial efforts to harness AI for policymaking (and other ends) have tackled individual tasks, like editing text or ideation. This isn’t surprising, as these sorts of problems are easier to address and can often be tackled by individuals (as opposed to requiring any wider system changes). Such examples were a great early route to get people in different roles using and learning about the strengths and weaknesses of AI in low-risk contexts. However, to achieve transformative change at scale, governments will need to reimagine the workflows of whole teams, departments or organisations. This is a much tricker task, which involves changing culture, behaviour and power dynamics.
A related barrier to scaling AI in policymaking is earning the confidence of civil servants and citizens. AI-driven policymaking tools that flourish in one setting might be stifled elsewhere by the lack of standards to assess safety, security, and usefulness, or a “not invented here” mindset, while citizens rightly want to know who might be accessing their data. Across the board, there is legitimate concern about LLMs suffering from hallucinations, generating plausible-sounding but inaccurate responses, and personal or sensitive data being sacrificed to model companies and moved across borders.
Ultimately, two conditions must be true for the government to effectively scale AI adoption: policymakers must be able to reimagine their work, and both civil servants and the public must be able to trust the outputs of AI driven tools.
Initial steps that might facilitate scaling AI adoption across government include:
Mandate that any AI tool deployed in policymaking must be accompanied by simple, clear documentation explaining how it arrived at its output, specifically for the civil servant user and the affected citizen. The AI Social Readiness Assessment developed by Nesta’s Centre for Collective Intelligence provides a neat example of this in practice, asking members of the public about the social acceptability of specific AI tools being considered for use in the public sector. The result is an AI Social Readiness Label that helps public sector commissioners make informed decisions about deployment that is in line with public expectations and values. While different contexts will require different degrees of social readiness, starting from a base expectation that social readiness should be considered would be a significant first step towards scaling AI adoption.
The Algorithmic Transparency Recording Standard (ATRS) register is a public database that provides standardised information about how and why UK government departments and public sector organisations use algorithmic tools to support decision-making. Earlier this year The Public Accounts Committee noted that the uptake of this register was slow, indicating that there is more that could be done to enforce this approach. The AI knowledge hub could serve as a helpful launching pad for such efforts.
From time to time, the government creates new institutions to tackle challenges of the day: think the GDS or the Office of Budgetary Responsibility. Why not use the creation of the next new body to launch an AI-native public organisation? It’s much easier to sort out digital foundations when building from the ground up. If we can build robust digital plumbing in one organisation and get AI to work, then this might serve as a beacon to others.
AI already has the potential to revolutionise policymaking - but it hasn’t yet demonstrated transformative power. Since AI has so many potential uses across policymaking, it’s incredibly difficult to tease out precisely what the future will look like.
A useful analogy is electricity. It has so many applications that, if asked which was the most promising, most of us wouldn't know where to start. In the same vein, there are undoubtedly AI applications that pioneers haven't thought of yet - early scientists experimenting with making frogs legs twitch through shocks, would have found it hard to imagine the CERN particle accelerator.
This is why experimentation is vital. Rather than making big claims and commitments up front, government should start small - isolating the specific parts of the policymaking journey, where the biggest opportunities or inefficiencies lie, and iteratively testing critical assumptions early in the real world. AI has been described as having a “jagged frontier”, where it’s not always clear what work should be done by machines and what work should be done by people. A test-and-learn approach to AI in policymaking would help navigate the uncertainties inherent in a powerful new technology. For example, at Nesta we have done small experiments with users that have helped identify evidence synthesis as a fruitful early opportunity for the use of AI in policymaking.
First steps toward encouraging experimentation might include:
Identifying, encouraging and supporting high-impact stakeholders - such as heads of professions or those in the senior civil service - across the civil service to model experimentation.
Historically, governments have not tolerated flops. But if we are to learn how to best use AI for policymaking then calculated risks need to be taken through experiments. Senior civil servants should publicly talk about times when they have failed, and promotion should be possible as a result of a well-planned experiment that yielded a negative result.
Empower policymakers to rapidly test AI applications, while gaining familiarity with the tools. Some suggestions unearthed during our recent workshop include:
The digital transformation required to harness the benefits of AI will reshape power within government. For example, if the government were to move away from departments having their own separate databases and user interfaces, and towards a more consolidated ‘single-entry point’ for wider and more efficient cross-government data access, this might mean digital assets that were previously in the hands of one department end up in the hands of another (or several others), changing power dynamics between institutions.
As noted above in our knowledge management point, if data access were determined by bone fide need for the delivery of public services or improved outcomes, rather than which building a team happens to sit in, citizens would certainly stand to benefit. At the same time, there is a risk that tech companies who might legitimately help governments harness AI for policymaking through improved digital infrastructure, are handed too much power, whereby they get to set the rules. The Horizon Post Office scandal where hundreds of sub-post masters were wrongly prosecuted for theft, fraud and false accounting due to problems with Fujitsu software offers a stark warning of the risks posed by unaccountable corporate power.
Rather than simply juggle power between the state and commercial actors when redesigning the digital architecture of policymaking, the government has the opportunity to empower citizens.
Early steps for government to manage these new power dynamics might include:
Our workshop showed that one early priority for the use of AI in policymaking should be as a means to engage with citizens. For example, semi-autonomous AI agents are increasingly able to conduct and analyse interviews, allowing better understanding of people’s needs than ever before. Rather than replacing traditional qualitative research or quantitative surveys, these new tools offer new ways of engaging with citizens.
As the control of digital assets potentially shifts or expands, from individual departments to consolidated bodies, there is a need for the government to proactively map where this power currently lies and where it might move. This would allow them to better establish checks and balances to ensure all actors (state and commercial) remain accountable to citizens, manage any losses of government power, and tackle unintended consequences.
Just knowing where power lies isn’t enough. If the digital transformation needed to harness AI creates new nexuses of power, then the government needs to ensure that these are accountable to citizens. Checks and balances will also be needed to avoid excessive concentration of power.
The government should prioritise standards that give citizens control over their data, rather than handing power to "black box" vendors. For example, the work by BIT and Yaspa utilised the UK’s Open Banking infrastructure to develop AI tools that identify harmful gambling patterns. Because this relies on "open" standards (rather than a proprietary corporate database), it allows for more accurate, cross-platform protection without a single tech firm "owning" the underlying data. This "open-first" approach is the best defense against the risks of vendor lock-in and the unaccountable corporate power seen in the Horizon scandal.
Ultimately, the power of AI in policymaking will not be unlocked by chasing the shiniest new algorithms. Instead, we need to marshal the political will to sort the basics. From managing institutional knowledge well, to reimaging whole government systems, we need to fix the foundations to properly harness AI for policymaking. The urgency and excitement surrounding AI should be deliberately channelled to smuggle in the necessary, albeit less glamorous, digital and cultural reforms.
The UK government must recognise that investing in AI-ready foundations is not a cost, but a critical strategic capability that will define the quality, efficiency, and democratic legitimacy of policy in the next decade. Failure to act decisively now risks cementing an era of mediocre, bolt-on solutions. If the government truly wants to harness AI for policymaking, we desperately need to sort our digital infrastructure.
