Analysis

AI undoubtedly offers impressive potential for government and public services. This view is shared by politicians, civil servants and private sector companies alike, as well as Labour-aligned think tanks likely to be influential on the new government.

But there is a need to temper the hype somewhat. Figures suggesting AI could save government as much as £200bn over five years should be treated with caution. Our interviewees were sceptical of such huge numbers and noted that public sector transformation around AI will be complex and time-consuming. And, as this report has shown, we’re starting from a relatively limited base within government.

In fact, a common theme throughout our research was the need to do the hard yards first. Away from the hype about the most cutting-edge AI applications, there’s much work needed to lay the groundwork for this transformative tech. As one interviewee put it: “There’s a lot of hype about using AI without doing the fundamentals. [But] we need to do the fundamentals first.”

Three fundamentals

In that spirit, our research suggests three core issues that require particular attention:

Data – government must get serious about improving data interoperability and quality across Whitehall. There are excellent reasons to do this even aside from AI (as Sir Robert Chote set out in the article referenced above), but it’s even more essential now. It is a non-negotiable part of the AI transition.

AI tools are only as good as the data used to develop and refine them. Many proposed use cases – e.g. an AI assistant to help citizens use government services – can only reach their true potential if they can use data held across different parts of the state. To ensure reliability and to keep improving, these models require a constant, high-quality data flow.

Funding – as well as realism about what the headline potential benefits are (whether cashable or not), government needs more clarity about what ‘investing in AI’ really means.

The upfront cost of buying or developing an AI tool is only the first step. Monitoring the outputs of the first iteration of that algorithm and then refining it will entail a repeating cycle of additional investment. Having that more rounded understanding of cost will make it easier to estimate the true returns AI can offer.

Government should also do more to establish common AI model training architectures to improve effectiveness. At present, many services (inside and outside government worldwide) take general large language models and train them in specialised domains or functions. This increases the risk of hallucination and inaccuracy problems (more on this below).

The central government units we engaged with have developed thoughtful architectures for training AI models on specific tasks, but some departments may not have done so – even as they push forward with their own use cases.

The new AI hub should be given funding to publish details of the architectures used by AI units. This would allow departments to build on or adopt them wholesale themselves. In the long run, funding for this priority could improve effectiveness and reduce hallucination – a wise, much-needed investment that could avert future controversies.

Regulation – balancing the accelerator and the brake is one of the key themes of this report. Labour’s AI plans appear to include some further regulation, though the full details of this are yet to emerge. This follows on from the previous government’s decision, in February 2024, to expand
the algorithmic transparency recording standard to cover all central government departments.

There may be merit to each individual decision to enhance regulation and we fully accept that AI safety is extremely important. But Whitehall must not fall into working in one of its comfort zones – risk management – without doing enough to hit the accelerator. The new AI hub within DSIT must make sure it’s also focusing on implementing and scaling up new AI tools.

These two priorities aren’t necessarily opposed. In some areas, proper risk management work can lay the foundations for improving adoption of AI.

For example, we heard that one central government unit improved the accuracy of one AI tool from an initial confidence level of 78% to 85%. DSIT’s AI hub could develop a framework that explains what confidence levels should be achieved before AI tools of differing risk profiles should be launched. And they could also create the mechanisms needed to measure accuracy (and in turn confidence) in these systems.

That would help those in government understand what’s needed to take an AI product from a trial to a deployed use case. It might also help Whitehall departments independently determine
whether the more over-hyped private sector AI products – some claiming accuracy of nearly 100% – really do make the grade. This kind of approach to regulation can balance risks and ease the way to implementation.

A roadmap for AI in government

As well as fixing these three fundamentals, government needs a clear, overarching AI roadmap. As things stand, we didn’t find – on any of the 7 Lenses – that government is yet in the right place to fully enable its AI transition.

The decision to consolidate GDS, CDDO and i.AI within DSIT is an excellent first step. As we’ve already argued, the lack of clear leadership around the AI agenda so far has hampered progress.

With a new single voice established, this is the perfect moment to set out a clear, coherent and influential AI roadmap – we recommend the following as a starting point for this. Our recommendations align closely with previous successful transformation programmes that accelerated digital capabilities in government.

A roadmap for the UK Government’s AI transition

Vision: finalise and publish a vision for AI use across the whole government by the end of 2024,
building on the existing work done by the CDDO to develop it. This should include plans to develop world-class AI use-cases for government, an improved training and education offer, better use of data, and identification of the key pieces of architecture (technical and regulatory) needed to underpin scale-up of AI. Consider GDS’ 2017 Government Transformation Strategy as a useful model for this type of vision.

Design: introduce a £100m funding pot administered by DSIT to identify and scale up promising AI innovations for use in government. Run a competition, using a similar methodology to the previous £20m GovTech Fund, to identify these exemplars. Repeat this process every two years to capture and develop new innovations.

Plan: following the publication of the AI vision in 2024, introduce a fully fledged AI plan for government – building on the departmental AI plans currently being developed. This whole-of-government plan should be launched in early 2025 and run for 12-24 months.

Collaboration: build on the limited formal and informal networks for collaboration around AI, by tasking DSIT’s AI hub with developing an engaged community across Whitehall on this topic. Learn from what’s already taken place by developing both formal and informal networks to share ideas and encourage cross-system working.

Accountability: capitalise on the creation of the single AI hub that is now fully accountable to DSIT ministers and top officials. Create a new set of cross-government digital missions, following the expiration (in 2025) of those included in the CDDO’s existing roadmap for digital and data, to drive improvements in areas such as data sharing and interoperability.

Transformation leadership: supplement existing efforts to upskill leaders around AI by investing heavily in bespoke training for permanent secretaries, project senior responsible owners, and other senior leaders. Draw on private sector, academic and other expertise to develop a well-rounded programme that is tailored to the needs of those at each tier of leadership.

People: urgently update the Government Digital and Data Profession Capability Framework to include AI roles with defined competencies. Invest heavily in training aligned to this new framework (from highly technical to softer skills like prompt engineering), with ambitions akin to GDS’ upskilling of 10,000 civil servants around digital.

This roadmap doesn’t offer all the answers. The transition to AI is a programme of change that will take years, with the potential to far exceed the impact of the digital transformation of government. And its also true – as even the digital leaders we spoke to acknowledge – that the full suite of opportunities AI offers is far from settled.

Time to hit the accelerator

The new government has inherited an extremely difficult legacy, with profound challenges and significant constraints on what it can do to resolve them. But this is also a moment where the system seems to recognise that reform, not new spending, is the only plausible path to change. Delivering on Labour’s ambitious missions means embracing frugal innovation and the potential
AI offers. Fixing public services now inescapably means modernising and transforming government.

Placing CDDO, GDS and i.AI within DSIT under Rt Hon Peter Kyle MP’s leadership was a wise decision. Our research shows that it will be strongly welcomed by digital leaders inside the system who felt that direction around AI has been lacking for some time.

With the machinery of government changes complete, there is now a chance to fully reset the government’s approach to and plan for AI. Measures like the new AI Opportunities Action Plan, led by the chair of the Advanced Research And Invention Agency (ARIA) Matt Clifford, suggest there is a clear commitment to drive wider adoption across the UK economy. Our observation is that DSIT does seem to be grasping this essential agenda.

The government must keep up this momentum. It must seize this moment and embrace the roadmap we’ve presented here. For too long, we’ve been stamping down on the brake while
– for all the excited rhetoric – failing to take the practical steps towards implementing AI.

A new government, with a new PM and a landslide majority, has the chance to be bold and commit to making the most of AI. It’s time to press the accelerator.