There is an assumption embedded in almost every major AI procurement decision being made in the UK right now. It is rarely stated explicitly, but it shapes everything: which vendors get considered, which frameworks get used, which contracts get signed.
The assumption is this: that the revenue of a company is a reliable indicator of the quality of its AI capability.
It is not. And the evidence, much of it on the public record, is increasingly difficult to ignore.
Where the Assumption Comes From
The logic behind it is understandable. Large organisations feel like lower risk. They have established relationships, recognised brands, armies of account managers, and the institutional credibility that comes with scale. If something goes wrong, the decision to engage them is defensible. Nobody got criticised for choosing a household name.
In traditional technology procurement this logic had some merit. Delivering enterprise software at scale genuinely required the kind of organisational infrastructure that only large companies could provide. Revenue and capability were reasonably correlated.
AI has broken that correlation. The frontier of AI capability does not sit inside the largest organisations. It sits inside the most specialised ones: teams with deep domain expertise, genuine implementation experience, and architectures built specifically for the problem at hand rather than adapted from general-purpose enterprise platforms. The organisations best placed to deliver sovereign, secure, production-grade AI for a specific use-case are frequently not the ones with the largest balance sheet. They are the ones who have spent years solving that specific class of problem and nothing else.
The procurement system has not caught up with this reality. And the gap between assumption and evidence is growing.
What the Evidence Actually Shows
Consider the publicly documented record of how large, established vendors have performed against the AI expectations set for them in the UK public sector.
A single US-headquartered data analytics firm has accumulated over £900 million in documented UK public sector contracts spanning the NHS, defence, and multiple government departments. The NHS contract, worth £330 million over seven years, was awarded to build a federated data platform connecting patient data across up to 240 NHS organisations. By the end of 2024, fewer than a quarter of England’s hospital trusts were actively using it. Multiple trusts told NHS England privately that adopting it would cause them to lose functionality rather than gain it. One concluded that the firm’s products did not exceed what the trust already had in place locally.
In December 2025, the same firm received a £240 million Ministry of Defence contract, awarded by direct award, without competitive tender. Thirty-four MPs signed an Early Day Motion expressing concern. A cross-parliamentary scrutiny committee raised questions about governance and accountability that were not directly answered.
This is not an isolated example. It is a pattern. And it is being sustained not because these organisations are delivering exceptional results, but because they are familiar, because they are already embedded in the frameworks, and because choosing them feels like the safe option.
DSIT’s own 2025 review found the public sector is losing over £45 billion in unrealised productivity benefits every year. The money is being spent. The results are not materialising. And the vendors receiving the largest contracts are, in most cases, the same ones they have always been.
What Revenue Actually Predicts
Here is what the data on AI implementation outcomes consistently shows. McKinsey’s 2025 survey found that organisations reporting significant financial returns from AI were twice as likely to have redesigned end-to-end workflows before selecting their modelling techniques. The single biggest driver of successful AI delivery was not the revenue of the vendor. It was the quality of the process understanding brought to the implementation.
UK research on AI adoption found that the top barrier organisations face is not cost, not access to models, and not technical capability. It is the difficulty in identifying activities and business use cases, in other words, not knowing clearly enough what problem they are trying to solve. A large vendor with a generalist platform does not solve this problem. It obscures it, because the platform creates the impression of progress before the underlying question has been answered.
What actually predicts successful AI delivery is specificity. A vendor who has implemented the same class of problem multiple times, in similar environments, and understands the failure modes intimately. A team small enough to be genuinely invested in the outcome rather than the contract value. An architecture built for the specific requirements of the deployment rather than retrofitted from an enterprise product designed for a different purpose entirely.
None of those characteristics correlate with revenue. Several of them are actively hindered by it.
The Risk You Are Not Accounting For
When a procurement decision defaults to a large, established vendor on the basis of perceived safety, it typically accounts for two categories of risk: the risk that the technology does not work, and the risk that the vendor cannot support it at scale.
What it rarely accounts for is a third category: the risk of dependency.
The UK spent £1.9 billion on software licences with a single US-headquartered technology company in 2024/25 alone, locked into an agreement running through the entire current government term. Ninety percent of UK public sector software comes from that same company. Seventy percent of cloud infrastructure is controlled by three foreign-headquartered hyperscalers.
For years, that dependency felt manageable because the relationships underpinning it felt stable. They no longer do. The geopolitical assumptions that made it reasonable to house critical national infrastructure, sensitive defence data, and public sector systems on the servers of foreign-headquartered technology companies are being tested in ways that would have seemed implausible only a few years ago. Alliances that were taken as given are visibly shifting. The strategic calculus that said “our interests are aligned” deserves, at minimum, to be reexamined.
This is not a political observation. It is a procurement risk that is not being adequately priced into decisions that are being made right now, at significant public expense, with long contract terms and deep architectural dependency built in. Choosing a large, foreign-headquartered vendor does not reduce risk. In the current environment, it concentrates it in a place that is increasingly difficult to control.
What Should Actually Drive the Decision
The question an organisation should be asking when evaluating AI capability is not “how large is this company” but “have they solved this specific problem before, in this type of environment, with these constraints, and can they demonstrate it.”
That question changes who is in the room considerably. It brings in organisations that may have fewer employees and smaller marketing budgets but deeper, more directly relevant experience. It surfaces vendors who have built their architectures specifically for sovereign, secure, or operationally complex environments rather than adapted enterprise platforms to approximate those requirements. And it produces better outcomes, because the vendor’s entire capability is focused on the problem at hand rather than divided across a hundred different enterprise product lines.
In 2024, just 5% of MoD direct procurement spending went to SMEs, against a central government average of 11% and a local government average of 35%. The top five suppliers accounted for over 34% of all MoD spending. The frameworks through which most contracts are awarded contain fewer than 20 approved suppliers. The market, as currently structured, does not ask whether a smaller, more specialised organisation might be better placed to deliver. It asks whether the organisation is already on the list.
We are actively pursuing inclusion on the Neutral Vendor Framework, a more open and genuinely competitive model for how the government should be accessing emerging capability. It is a step in the right direction. But changing the framework is only half the answer. The other half is changing the assumption that drives the decision in the first place.
The Practical Implication
If you are a technology decision-maker evaluating AI capability, the most useful shift you can make is to separate the question of revenue from the question of fit. They are different questions, and conflating them is costing UK organisations, in defence, in government, and in enterprise, real capability and real money.
Ask for evidence of delivery, not evidence of scale. Ask for specific implementations in comparable environments, not case studies from adjacent industries dressed up to look relevant. Ask what the vendor’s architecture was built to do, not what it can be configured to approximate. And ask, honestly, what the dependency implications are of the choice you are about to make, because the risk that does not appear in the initial business case is frequently the one that matters most three years into the contract.
The organisations delivering the most impactful AI right now are not always the largest ones in the room. They are the most focused ones. The ones who know exactly what they are doing, have done it before, and are not distracted by a hundred other product lines and account relationships competing for their attention.
Revenue, in AI, is not a proxy for capability. It never was. It is past time that procurement decisions reflected that.
Agentycs is the end-to-end AI orchestration platform built for organisations that need to operate AI in sovereign, secure, and classified environments, on their own infrastructure, under their own control. When Agentycs is deployed, sensitive data never leaves your environment, inference runs under policy you define, and every output is traceable back to its source. Learn more at www.agentycs.com.