Tracking Progress on the UK’s AI Opportunities Action Plan: Implementation is what counts for growth

The UK has seen four national strategy-style documents in the past four years¹. Talk is cheap, implementation matters. The UK government’s latest announcement that they will be ‘taking forward’ all 50 recommendations from Matthew Clifford’s AI Opportunities Action Plan could bring about exceptional results for AI development and adoption in the UK. 

The focus of these recommendations is to: (1) invest in the foundations of AI; (2) push hard on cross-economy AI adoption; and (3) position the UK to be an AI maker. For each, the government has promised updates on commitments throughout 2025 and we have created a tracker to bring together existing progress, and keep up with work on the Plan.

The government is not starting from scratch. There is significant progress to build on, especially in areas relating to:

• Opening up government data safely, through data sharing programmes like the UK Data Service for researchers, and trialling the use of synthetic data by the NHS
• Supporting AI skills growth, in academia through Turing AI Fellowships and in the public sector through the No10 Innovation Fellowship, 
• Piloting AI use cases across government departments, some of which can be tracked in DSIT’s Algorithmic Transparency Register. 

Relevant initiatives for each recommendation are recorded in our progress tracker. Much of the current work is fragmented. Arguably, the newest contributions from the Plan come from the infrastructure commitments, which hinge on what will be included in the Long-term Compute Strategy to be published later this year, and the launch of a new ‘UK Sovereign AI’ unit, whose success will likely depend on the powers and remit it is given within the already crowded UK government AI unit/taskforce/incubator/institute/policy directorate space.

Delivering the Plan will take a cross-government effort. Recommendations for (2) are heavily focused on public sector adoption of AI. The role of HM Treasury will be critical for the “Scan > Pilot > Scale” approach to work. Firstly, funding needs to be available for cost-risky ideas and there must be a clear path to continued funding for turning successful pilots into long-term initiatives. Otherwise, good, novel use cases will not be tested. 

Secondly, many pilots will require collaboration and need joint funding opportunities for departments to apply for together. The treasury will need to support departments to distribute ownership of the benefits of collaborations on their balance sheets fairly, in a way that incentivises participation.

The UK must prioritise the role it wants AI to play, and the role it wants to play in AI. Recommendations for (1) and (3) aim to attract infrastructure investment and talent to the UK, and encourage frontier AI development domestically. Each of these goals puts the UK in a global competition for limited human and financial capital. 

Recognising this demands nuance to the ambition for the UK to become an AI superpower. We need to define what it looks like for the UK to be an AI superpower. It will not look the same as the US’ definition, and PM Keir Starmer has pushed regulation into the future in contrast to the EU’s approach. 

The Plan starts to sketch out this definition by suggesting the UK aims to grow companies supplying targeted parts of the AI stack, encouraging mission-focused use cases in government, and prioritising the sharing of high value datasets. The Plan also includes some suggested priority application areas of its own. Targeting AI for science is a good example that builds on the UK’s higher education reputation and infrastructure. The UK should lay out a set of progress indicators that capture its own definition of leading the world in AI.

Becoming a leader in AI is a moving target and these commitments need to be nimble. The prerequisites for encouraging AI adoption and development are changing as the fast moving industry develops. The scale of commitments may need to change; for example, compute requirements predicted today may be obsolete in 5 years. 

The shape of commitments may also need to change. For example, as progress in different AI techniques shapes what data is valuable and in what format. How will data need to be labelled (if at all)? Will unstructured data be useful? Will new tools allow us to make sensitive data accessible?