The UK is in the middle of shaping the public sector’s artificial intelligence (AI) capability for decades to come. The visible part, comprising Copilot rollouts, foundation model partnerships and the headline contracts, is already well advanced.
The less visible part is whether, in the course of those deployments, we also build the buying capability, cultivate a plural supply base, and establish the shared standards that will let us adapt as the technology evolves.
We have made a strong start on the first half of the task. The second half is the one that will determine whether the first half pays off.
Silent lock-in
We can call this the “silent lock-in” trap. It is the accumulation of AI capability on top of infrastructure, management practices and governance approaches that are individually defined, poorly coordinated and mismatched to the pace at which the technology is changing.
Despite the hard work of individuals and teams to procure and experiment with AI’s emerging capabilities, the pieces are not adding up the way they should.
What can we do to learn from the last decade’s digital transformation experiences to accelerate the UK’s AI adoption? That is the subject of Making AI work for Britain, published by London Publishing Partnership and available for download at FutureOfAI.uk under an open-access licence. The book draws on several years of research into the UK’s AI strategy and ecosystem, and on over a decade working with UK government on digital transformation.
In the book, I set out a framework for AI success based on a simple strategy – consolidate demand, diversify supply. The short extract that follows is drawn from the final chapter and summarises three of the key recommendations flowing from this analysis: Build buyers who can push back; pool demand that is already shared; and keep the supply side plural.
Build buyers who can push back
The smart-buyer problem is easy to describe and difficult to solve. AI suppliers, particularly the larger ones, now routinely make claims that require significant technical capability to evaluate – claims about training data provenance, model behaviour under distribution shift, security properties of the fine-tuning pipeline, interoperability with alternative providers.
Most procurement functions were designed to assess claims like, “this system meets this specification” and “this supplier has these references.” They were not designed to assess claims like, “this model will remain useful as underlying capabilities change.”
“None of this requires new powers or new money. It requires the decision to treat AI procurement as a capability we are actively building rather than a series of individual deals”
Alan Brown
The smart-buyer model does not mean building deep AI expertise in every department. It means, in each organisation spending meaningfully on AI, having a small core of people who can sit opposite a vendor and know what they are looking at. That core needs three things – the authority to say no, the technical depth to justify it, and enough exposure to current practice to recognise when claims have quietly drifted from their evidence.
Where that capability exists, suppliers behave differently. Where it does not, they behave as suppliers to an unsophisticated market always have.
Pool the demand that is already shared
Many of the AI problems public sector organisations are solving are the same problem. Case summarisation. Triage. Translation between policy language and operational systems. Document extraction. These requirements do not vary meaningfully between one department and the next, and the money spent separately working them out is considerable.
Consolidating demand is the less celebrated half of the lesson we took from digital government reform. When the specification of a shared requirement is done well once, as a common evaluation framework, a reference architecture or a shared procurement vehicle, the supplier market responds to it. Three or four suppliers quickly learn what “good” looks like, and they compete on it.
The aim is not to buy the same system everywhere – that was the mistake of an earlier generation. The aim is to agree on what the shared requirement is, measure it consistently, and let departments make the local calls within that frame.
Keep the supply side plural
No market stays plural on its own. Left to itself, enterprise AI will concentrate, because the economics of foundation models favour scale and because the switching costs of deeply integrated AI services are high. That concentration is not inevitable, but avoiding it requires active stewardship rather than hope.
In practice, stewardship means three kinds of move. First, treat open source and open-weight models as first-class options in public sector procurement, with evaluation criteria that credit them for the strategic flexibility they preserve.
Second, use the UK’s research base and AI ecosystem as suppliers as well as subjects of study, which means procurement vehicles that smaller providers can actually clear and contract durations that give them a realistic shot at building capability.
Third, treat the AI Security Institute and the UK’s sovereign compute investments as part of the operational supply map available to departments, not as national prestige projects standing apart from day-to-day procurement.
These three recommendations illustrate the way forward for AI in the UK. Over the next few months, most large public sector organisations will sign AI contracts that shape what they can do with the technology well into the next decade. Each is a chance to build smart-buyer capability, to consolidate shared requirements, and to keep a seat at the table for new suppliers.
None of this requires new powers or new money. It requires the decision to treat AI procurement as a capability we are actively building rather than a series of individual deals.
Alan W. Brown is the author of Making AI work for Britain, published by LPP. He is a professor in digital economy, an experienced business executive and a strategic advisor. He has spent more than 30 years in the US, Europe and the UK driving large-scale software-driven programmes with commercial high-tech companies, leading R&D teams, building state-of-the-art solutions and improving software product delivery approaches.