We interviewed 85 individuals, reviewed 1021 documents, and conducted 12 observations.
A detailed description of the dataset is provided in Table 1.
Table 1 Overview of the dataset
Table 2 provides an overview of emerging findings, which we will explore in more detail in the subsequent paragraphs.
Table 2 Overview of findingsaAn ambitious experimental initiative in a complex ecosystem
The AI Lab represents a unique and ambitious initiative, intervening across several overlapping ecosystems, including vendors (and the associated anticipated development of an innovation ecosystem stimulating the formation and growth of start-ups), regulators, the NHS, and policy. As such, it faced challenges in balancing multiple, sometimes conflicting, objectives, such as focusing on both the development and commercialisation of early-stage AI innovation and the large-scale deployment of established AI technologies. This required compromises that, in some instances, meant neither objective was fully met. The UK Government’s priorities included innovation and scaling of technologies, but this often conflicted with NHS provider structures.
The question of how to accelerate adoption and scaling within the NHS is a recurring topic among Number 10 [10 Downing Street, the official residence and executive office of the Prime Minister of the United Kingdom] and the Secretary of State. Leveraging the NHS’s spending power and streamlining the process for companies and innovative ideas to scale across the system are major challenges. The NHS’s structure is not optimised for innovation, which can be especially difficult for start-ups or ideas that lack evidence or change frequently.
AI Lab Board 24, Minutes, May 2023
Government’s priorities were also not always viewed as reflecting system-level priorities, as the following quote illustrates.
…the original high level objective is about testing and accelerating the use of AI in in health and care, but I don’t know from my perspective it felt like well surely then we should be looking at the system and looking at where the problems are and wondering where could AI feasibly help, whereas it didn’t feel like that’s what we were being driven by. I was never really clear what it was being driven by.
Interview 5, DHSC
The AI Lab was also experimental by nature, operating in a rapidly evolving area with little precedent to guide its activities. There were, for instance, different priorities at different times. At the beginning, there was a focus on proof-of-concept projects through the Skunkworks, which was closed mid-way through the programme. Responding to the COVID-19 pandemic, the AI Lab developed an image repository to develop and test diagnostic tools: the National COVID-19 Chest Imaging Database (NCCID). Experiences from this fed into a vision to develop a national AI Deployment Platform to support the validation and deployment of AI tools at scale. A pilot implementation of a commercial product was launched in two imaging networks towards the end of the programme. There were also various attempts to create a knowledge ecosystem, such as setting up the AI Lab Virtual Hub community, but these dissipated over time.
Without prior knowledge, and given the perception that urgent action was needed, it was necessary to adopt an accelerated experimental learning approach: learning-by-doing and learning across different contexts. One perhaps inevitable consequence was uneven success across different programme components and different projects within those components. The following quote illustrates that the AI Lab was initially envisaged as a series of experiments.
…think about it like it ran a series of experiments looking at you know, can we stand up a national imaging platform which is centralising health data which we know how bad that went the first time. So now we’re going to try it again, but in a different way. You know, we’re going to stimulate industry through targeted investments in new technologies.
Interview 43, Previous AI Lab Member
The AI Lab’s accelerated launch limited opportunities to establish robust processes, project plans, and baselines, while the COVID-19 pandemic presented a serious disruption. The pandemic shifted strategic priorities and diverted resources, but it also created new opportunities for AI, such as the development of the NCCID. More widely, COVID-19 also focused the NHS on data, automated interpretation of data and the benefits of sharing infrastructure in enabling the response of the health system. This shifted attitudes towards digitalisation of health and care delivery, coupled with rapid product development by vendors, and opened the minds of many to the potential benefits of the work of the AI Lab. As evidenced in the following quote, many resources were diverted during the pandemic, including those of provider organisations and those within the AI Lab itself.
And then COVID came. And COVID caused a multitude of problems…people were pivoted to work on COVID programmes, it was the COVID kind of data vaccine database… All sorts of things have been set up where people were called from teams. And so, the team in the AI Lab shrunk hugely, … the director… was moved to work on other urgent programmes. And we pivoted some of the work of the Lab because it felt wrong to continue on the same track. We continued with some of the programmes we’d already set up, continued with the regulatory aspects, but pivoted to set up a national COVID chest imaging database.
Interview 54, Previous AI Lab Member
The AI Lab has made considerable progress in coordinating and facilitating conversations across stakeholder groups, which have become more mature and better informed about the opportunities and challenges of deploying AI in health and care. The overall structure of the AI Lab was rational and coherent, but, in operation, too often individual projects within the AI Lab were siloed and not integrated effectively with wider activity.
The AI Lab has also contributed to national regulatory guidance, helped identify existing gaps and needs in the healthcare AI ecosystem, begun to understand AI markets, and facilitated learning on how to develop, implement, and evaluate AI technologies. The following excerpt shows how the AI Lab may be viewed as a series of natural experiments that allow learning and contribute to the empirical evidence base surrounding AI implementation and adoption processes.
Every minute we spend talking about something shines a little bit of light on the subject… And even actually you learn a lot more from your failures than these successes. You know, from my perspective, that’s a tremendous value. And doing a lot of things under if not controlled conditions at least you know being able to you know have the opportunity to look most NHS hospitals wouldn’t talk to you about their failures at all. So having a link into…lots of similar projects in lots of different and relatively similar organisations and understanding why they fail for me is a tremendous opportunity. It’s a huge opportunity to be able to work out a really powerful and evidence-based logical model for complex technical change in complex organisations.
Interview 18, DHSC
As such, the AI Lab has begun to establish the foundations for safe and effective adoption of AI in UK health and care. Initial over-hyped optimism about the readiness and profitability of AI has given way to a more realistic understanding amongst some (but not all) stakeholders of the contexts and challenges associated with development, regulatory approval, and procurement. The extensive exploration of data stewardship models highlighted the tensions between various approaches to stewardship and the importance of Patient and Public Involvement and Engagement (PPIE). This work deepened understanding of the challenges facing a data-driven NHS. The quote below highlights the important contribution of the AI Lab in relation to regulation.
…we need to have public and patient involvement…we work with the health research authority as well, and obviously through their ethics lens, that’s a big part of things. One of the big things that came out of the AI Lab is definitely the digital and AI regulation service…that’s genuinely good guidance on how to go about generating an AI product and that does advocate for clinical and public engagement. And we sponsored a piece of work looking at patient stewardship, data stewardship.
Interviewee 85, Previous AI Lab Member
However, there remains a risk that valuable lessons, especially those emerging from the AI Lab’s challenges, may be overlooked in favour of focusing selectively on successes revealed in immediate outcomes. There is also a risk that lessons identified are not taken forward to inform future strategy and that longer-term developments are not sufficiently followed up. The excerpt below captures the difficulties encountered with the AI Deployment Platform, which was initially planned to be rolled out nationally, but this did not materialise. Nevertheless, the work provided important learning opportunities.
…in terms of like the AI Deployment Platform for instance, the way that that’s been procured with a view to you know this would be a national roll out after the pilot if successful, what we’re seeing is that actually a national rollout might not be the most appropriate route, that’s a huge benefit because it saves what could have been. …you know, a disaster; we don’t know yet. The pilot’s still in full swing at the moment… Early conversations are saying potentially this might not be the way to go moving forward. …So, although it’s a bit of a, you know, a loss from our side, overall, it’s a really big win because it gives you an opportunity to actually see, right, that wasn’t the right way to do it. This is another way that we could be doing this from the lessons learned that we’re taking from this particular project. So that’s a that’s a massive benefit.
Interview 9, DHSC
Turbulent macro-environmental influences and political drivers
The AI Lab’s strategy and delivery were heavily influenced by a turbulent macro-environment and political drivers. The programme was established in part to stimulate the UK economy and position the UK as an AI superpower driven by ministerial enthusiasm for innovative technologies (particularly around diagnostics). However, despite these ambitious goals, the budget was modest compared to other large government programmes. As evidenced in the following quote, the drivers underlying the establishment of the AI Lab at the Government level were to position the UK at the forefront of using AI in health and care.
The work was done and ultimately with Boris Johnson [the then Prime Minister] and Matt Hancock’s [the then Health Minister] sponsorship, the 240 million was put together to run the AI Award and the AI Lab to move all of that forward with the launch in…early 2019 with a view to…putting England at the forefront of using artificial intelligence for the benefit of health and care, so that just reflecting back to you, that’s kind of the value idea. Wow, that’s hard, really hard. I mean, really, really hard for all sorts of different reasons. I look at that and I, the cynic in me can understand….it was really clear that it was—this halo of…trying to position England at the forefront of using AI for healthcare.
Interview 7, NHSE
The turbulent political landscape has significantly shaped the AI Lab’s activities. Since its inception, there have been six Health Ministers and four Prime Ministers, each with different priorities and changing governance arrangements. In addition, regions have acquired greater responsibility in planning and funding local services since 2022, which further contributed to changes in stakeholder priorities. NHSX, the initial delivery unit for the AI Lab, was dissolved in 2022. This led to staff changes, uncertainty, and restructuring within the programme. Departures and associated disruptions meant that painfully acquired expertise and knowledge dissipated. The effects on the ground of changes at the centre are illustrated by the following quote.
You’ve had all of these, sort of, mergers and demergers at the centre. You’ve had a political system, and you’ve had how many health secretaries have you had? I mean, honestly, working in the NHS right now is so hard. It’s so hard. And part of it is you don’t know if you’re… guaranteed that you won’t have consistent leadership, consistent funding. Consistent strategies and things that would have gone…one month get blocked the next.
Interview 21, ASHN
Changes in government were at the core of changes in the AI Lab’s strategic direction.
…you have a hiatus with political change, you also have periods where people are aware, you know, often weeks, if not months in advance about when a government is reaching the end of its period and there’s going to be an election. So that means that decisions are slowed even then, so it’s like, “well, the election’s coming, let’s not make any buying decisions until after that point”. So, you you’re kind of floating along through those layers.
Interview 56, Supplier
Additionally, a DHSC spending review in 2022 resulted in a ~£107 million reduction in the AI Lab’s funding and drove a shift away from experimentation toward a focus on evidencing the delivery of tangible benefits. Ring-fenced funding was reallocated, and planned programmes were cut.
The AI Lab receiving a challenging spending review settlement, which removed its ring-fenced funding for AI-related activities and the departmental reprioritisation reducing the AI Lab’s future years remaining Capital budget down from £180 m down to £75 m for 22/23–24/25. This required pausing and/or stopping a number of programmes.
AI Lab Board 21, Minutes
We observed that, associated with this, AI Lab staff spent significant efforts on progress reporting, trying to track and account for programme impacts and benefits. Reporting activity was escalated by requests from multiple disparate stakeholders with different reporting requirements and timeframes across departmental boundaries, driven by the need to demonstrate the value of investment. As a major government programme, the AI Lab was subject to high accountability requirements supported by reported evidence.
I mean, again, for the wider team probably did affect quite a lot because there’s a lot of governance that’s attached to being a government major programme. There’s a lot of reporting, rightly so. There are gateway reviews, there’s the whole kit and caboodle of things that you have to do if you’re a government major programme because you are given some money to go and do stuff and you have to see how you’re delivering against that stuff.
Interview 5, DHSC
The AI Lab was unique in its ability to connect policy, technology development, evidence generation, and healthcare services. This position created significant opportunities but also posed challenges in aligning the diverse needs and priorities of stakeholders. For example, participants noted a widespread overestimation of the maturity and capabilities of AI and its supply chain, leading to inflated expectations about progress. Additionally, the lengthy timelines required to achieve and demonstrate outcomes often clashed with the shorter-term nature of funding cycles. Government funding models with fixed timelines and costs potentially inhibit innovations that require more flexible models.
It’s a problem with like how government funding models work, because it’s between each financial year and then you know it creates the wrong behaviours of like how we’re trying to deliver and do things. And it’s just it doesn’t work for innovation and needs to be changed. … There needs to be more flexibility because you don’t want to have a funding constraint in terms of you need to spend, you know your budget is only per financial year driving like how quickly you stand things up because otherwise you do things incorrectly to kind of meet a timeline, whereas actually it’s just you know it’s extremely difficult to plan these things anyway. Then you add on all these constraints of like you know outcome based milestones that you can only get paid once you’ve delivered and then it’s just it just creates an unnecessary set of challenges to try and deal with on top of the actual challenge of trying to deliver. An innovation project anyway, which is already difficult.
Interview 32, Previous AI Lab Member
The AI Lab’s distributed structure and governance further complicated delivery. Rather than being a single, established organisational entity, the AI Lab was an evolving network of actors and collaborations that changed over time. Its joint governance between NHSE and DHSC added complexity, sometimes resulting in misaligned objectives, strategic directions, and blurred boundaries between the two organisations. Organisations collaborating in the delivery of the AI Lab were often competing:
… all of those organisations, all of whom are constantly competing for their own right to exist and their own money, and none of them have the same objectives, and yet they are supposed to somehow bring about…the same end goal now when you’re thinking about AI, that’s even more complicated
Interview 13, Previous AI Lab Member
While bringing together organisations from across the ecosystem—including policy makers, providers, suppliers, and regulators—was seen as a strength, there was a noticeable lack of well-established communities of practice at the frontline of service delivery. Early efforts to build these communities were made, but they were not sustained due to limited resources and lack of forward planning for a sustainable community. The strain on the NHS, particularly during the COVID-19 pandemic, further diverted capacity away from these initiatives.
Generating real-world evidence to support market authorisation and large-scale adoption
Phase 4 Awards aimed at supporting technologies with market authorisation but insufficient evidence for large-scale commissioning or deployment. Supplementary Table 1 and Table 3 summarise data extraction for the Phase 4 Award projects.
Table 3 Benefits realised for completed Phase 4 projects
Whilst the potential for impact was not limited to Phase 4 projects, the maturity of the projects and focus on evidence generation meant that a systematic analysis of impact and benefit was possible. There were 10 Phase 4 projects with final reports available at the time of writing, which were included in the analysis. One project stopped without delivering a final evaluation report. The overall funding provided by the AI Awards for the 10 projects included in the analysis amounted to £22.5 million.
We found that while all included projects correctly identified clinical contexts and pathways, as would be expected from technologies with market authorisation, some failed to identify important bottlenecks/pain-points that could be targeted to deliver improvements in efficiency or effectiveness (3/10). There were also instances (2/10) where technologies supported by the award were successfully deployed in care activities but did not demonstrate a clear improvement in health or economic terms when compared to existing best-practice care pathways. For example, some compared less favourably with other technologies or a varying staff skill mix. One project (1/10) delivered a retrospective study only, limited by timeline and budget constraints, which impacted the opportunity to complete the originally planned prospective study. However, the latter is required to provide evidence suitable for regulatory purposes. This highlights challenges with aligning multiple stakeholders to deliver prospective evaluations within the AI Award timeframes in the absence of extensive prior engagement efforts.
Two projects (2/10) were able to demonstrate significant improvements in effectiveness and efficiency. For example, in Project 1 “treatment rates rose to 5.7% at [technology name] hospitals compared to the national average of 3.6% […] highest performing hospitals taking part in the evaluation reached mechanical thrombectomy rates over 10%, the target set out in the NHS Long Term Plan baselined at 1% in 2019.” This supports the potential of the technology to introduce efficiencies that support the delivery of best practice treatment targets. Furthermore, in Project 6: “The use of [technology name] impacted management recommendations in […] 7.9% of cases where there were nodules detected [of which 69%] would otherwise be dismissed by the reader and instigated a follow-up recommendation, constituting a major change in management.” This highlights the potential of the evaluated technology to improve effectiveness in radiology by demonstrating a superior adherence to radiology guidelines by less experienced clinicians.
Phase 4 Awards provided the greatest scope for delivering measurable Return on Investment (RoI) but we found significant heterogeneity in the level of evidence and approach to measurement. We could only value the RoI and the associated economic benefits for three completed projects (Table 3).
Where we could evaluate RoI, we observed heterogeneity in approach and scope of evaluation. In some cases, evaluations were not designed to capture long-term impacts. Many technologies did not fulfil some stakeholders’ prior expectations of large-scale adoption, but nevertheless reported some substantial impacts. These are discussed below.
Project 1 implemented a diagnostic tool in a non-elective care setting across a range of regional networks within the NHS. The technology provided a set of decision support tools that aided frontline clinicians to make time-critical treatment decisions. Increased rate of optimal treatment was hypothesised to improve patient outcomes and reduce associated costs to health and social care.
Considering short-term care efficiency, there was a reported increase in optimal treatment, leading to a modelled average of £110 increase in cost of care per patient. This was offset by efficiencies in longer-term care and improved patient outcomes modelled as a five-year (discounted) saving to social care and an increase in quality-adjusted life years (QALYs)29. The project valued this close to £400 per patient, leading to a significant cost-saving estimate of nearly £44 million across the approximately 150,000 patient cohort.
The evaluation was not able to conclude whether the technology provided point of delivery efficiency. Whilst this is not likely to be a significant value compared to the five-year projected benefit, understanding of implications to delivery of care is critical for implementation and adoption.
Project 6 deployed a diagnostic AI platform, integrated into existing software systems, to optimise oncology pathways in secondary care (elective and non-elective). The evaluation examined the impact on the quality of patient management recommendations, including decisions such as discharge, scheduling a re-scan at a later date, or pursuing further diagnostic workup or review. This technology, unlike Project 1, was reported to save clinician time at the point of use. Patient quality of life benefit from earlier diagnosis was not modelled due to a lack of external evidence for the impact of early diagnosis on disease progression and patient outcomes.
Project 7 deployed diagnostic AI-software to support pathologists in identifying and diagnosing tumours in secondary care. It anticipated that lab-based testing activity would be reduced by increasing accuracy of initial imaging review. The evaluation calculated short-term care efficiencies of £36 per patient. However, the analysis focused on 90 days histology impact and not clinician time at the point of delivery. The project could not draw conclusions about long-term care efficiency and patient outcomes.
Only one evaluation conclusively addressed the impact of the use of AI on patient quality of life. Here, the associated benefit was reported to be substantial. This highlights a core challenge in assessing the impact of AI technologies: the most significant benefits or disbenefits often emerge over extended timeframes, exceeding the typical duration of procurement and evaluation periods. Similarly, only one evaluation addressed the immediate impact on service delivery—understanding of this is crucial for service planning to support adoption of new technology.
Projects 1, 6 and 7 illustrate some challenges. These arose from omissions in data collection, which resulted in key impacts and benefits not being measured, and gaps in the evaluation design, where the connection between the data gathered and the expected outcomes was not clearly articulated. Additionally, some impacts and benefits were inherently difficult to quantify, making their measurement complex. These issues were further compounded by ambitious evaluation goals that could often not be realised within the constraints of the available project budget.
These projects did, however, demonstrate broadly appropriate evaluation strategies and, coupled with their maturity and alignment within their deployment setting, have provided promising indications of the capability of AI technologies to realise a RoI.
For the Phase 4 projects examined, those demonstrating potential RoI tested a diagnostic or screening tool in a clinical setting. Administrative tools in operational settings were included in Phase 3 projects (discussed below). In clinical contexts, the most significant return was based on early diagnosis to avoid downstream patient care costs and impacts on quality of life. An overarching factor for “success” was therefore that projects aligned to identified national health service and whole system priorities and targets for improvement, underpinned by the level of burden and unmet need with significant monetary implications and patient outcomes.
Ultimately, our analysis indicates that a project’s success in generating evidence for large-scale commissioning or deployment depended both on technology maturity/penetration and appropriateness of the evaluation programme. Supplementary Table 1 summarises the dimensions of appropriateness of the evaluation programme, technology maturity and alignment for all Phase 4 projects under study.
Whilst benefits were reported in relation to AI Award funding, which we assumed covered all costs of technology and implementation, reports did not always robustly analyse RoIs based on the cost of the technology to the healthcare provider when deployed as part of normal service delivery. Extrapolating both cost and patient benefits is challenging due to various local factors, including digital infrastructure, a healthcare provider’s capacity to act on the availability of earlier diagnosis, and their ability to deliver preventative care. We discuss this further in the Limitations section.
Having reviewed the outputs from Phase 4 Awards, we will now describe findings from Phase 3 Awards. Supplementary Table 2 summarises data extraction for the Phase 3 Award projects that had been completed at the time of writing. Again, we observed a great deal of heterogeneity in evaluation approaches, likely influenced by technology maturity, innovation stage, and degree of NHS penetration. Approaches ranged from assessing the feasibility of technology integration with existing clinical pathways (Project 5) to randomised controlled trials aimed at establishing the effectiveness of interventions compared to standard care (Project 3). The latter matched the scope of a Phase 4 Award, which underscores the difficulty in clearly distinguishing between the two phases.
A considerable proportion of Phase 3 Award projects (7/10) reported sufficient evidence of having achieved the main objective of this phase, which was to support the first real-world testing in health and social care settings, including evidence for routes to implementation. This suggests that they could be advanced to the next AI Award phase, irrespective of whether the innovators pursued this option or not. Interestingly, most technologies (4/7) were already generating revenue, some (3/4) in an international context, despite great variability in the level of self-reported maturity (ranging from TRL 3 to being CE marked and having Food and Drug Administration clearance). A higher maturity level was not necessarily associated with success in Phase 3, which highlights that mature technologies can also encounter implementation challenges. The three projects (3/10) that did not make enough progress to generate significant outcome evidence pointed to the following confounders:
a) changes in clinical context (e.g. diagnostic test guidelines) that invalidated the main use case for the innovation;
b) focus on the effectiveness of the technology without addressing integration in existing (or new) clinical pathways;
c) innovators and/or project managers lacking expertise in integrating systems with local information technology infrastructures.
Nine out of 10 technologies provided evidence of successful integration into an existing clinical pathway, eight out of 10 addressed efficiencies in the pathway, and five out of 10 addressed cost-effectiveness. Notably, project 10 was able to quantify significant savings in administrative tasks arising from the implementation of the technology as part of the evaluation. Finally, half of the technologies (5/10) demonstrated some degree of adoption in new settings arising from the AI Award. However, this varied (from sites involved in the AI Award continuing to use the technology after the project to double-digit new site deployments contracted during the project, see Supplementary Table 2). We cannot ascertain medium-to-long-term sustainability. There were also challenges to gaining buy-in to support procurement. For example, one project found that the specialty team deploying the technology and achieving a reduction in emergency admissions did not directly benefit, as budgets remained unchanged. This disconnect reflects the fragmented NHS budget system, where savings in one area (e.g., emergency admissions) do not translate to budget reallocations for responsible teams. More generally, Phase 3 projects do not appear to have followed any standard approach to procurement.
Scaling of technologies and pathways to procurement, reimbursement, deployment, and operation
The AI Lab was a research and development programme and not a delivery programme. We observed significant uncertainty surrounding the procurement pathways of technologies developed and implemented as part of the AI Lab. For example, we found that many sites participating in the AI Awards had no transition plans to procure the implemented technologies after completion of the AI Awards, as they lacked evidence to support clinical or economic cases at the time that the AI Award funding was due to run out.
This uncertainty was seen as a threat to the sustainability of these technologies, as participating sites might face financial challenges in maintaining the systems. The situation was further complicated by coordination difficulties stemming from national procurement guidance and the inherent complexity of NHS procurement, which is managed independently by hospitals with diverse needs. The following statement exemplifies the convoluted procurement processes in provider organisations.
…it seems ludicrous if you’ve got 12 people doing a procurement for a specific type of thing that you don’t turn around to them in the centre and go “here’s a template that you can all use for doing it”. … Because to me, why would I have 12 people go off and write their own template? Makes no sense whatsoever. It’s the same thing for all of the stuff that’s kind of going on here. We’re saying, well, if…they’ve got the approval to deploy their AI, why don’t we get their clinical safety approval and share it with everyone else? … Because otherwise…I’m going to get 12 of them or all of the [hospitals] are going to do their own thing, because the best role that NHS England in the centre can play is to connect the people together and go: “You’re doing what this person over here did it six months ago? Here’s all the stuff that they used to do it”.
Interview 29, NHSE
Suppliers highlighted the lack of a route to procurement and the impact on suppliers.
And the trouble is, you just fall off the cliff edge at the end of it. So you do, you know, you implement something, you evaluate it, you show that it’s, you know, you could, if you’re lucky, within the small time that you’ve got and you’re able to show clinical effectiveness and you can cost effectiveness, which is hard to do within a time scale. You know, even with all of that, you build a business case, and you’ve got the evidence, there is no route to procure procurement.
Interview 23, Supplier
Many suppliers were primarily focused on the United States of America (USA) market, and some reported that the complexity of UK procurement pathways discouraged companies from scaling their technologies within the country. The existing NHS procurement processes were viewed to be ill-equipped to accommodate the dynamic nature of innovative technologies, prompting some companies to withdraw from the UK market altogether.
The procurement within the NHS in general is lagging very far behind the technology that exists and could be available. So, if we get a tender specification from NHS supply chain it is…for older technology and it is actually, and I’m talking about technology that has been around for 60 years and hasn’t changed…and it doesn’t even give you the option of explaining how your technology could benefit this hospital. Because it doesn’t ask the right questions in the tender specification.
Interview 57, Supplier
In addition, although the AI Lab worked with the UK Medicines and Healthcare products Regulatory Agency (MHRA) on adapting regulatory processes to take account of AI and with the National Institute for Health and Care Excellence (NICE) on adapting the Evidence Standards Framework to establish a quality and cost-effectiveness framework including AI, in the UK, regulatory complexity was reported to be greater than in other countries (such as the USA).
Significant public funding for mature product development, implementation, and rollout mobilised expectations about RoI for the taxpayer. This presented novel challenges for NHSE/DHSC, which had to develop benefit sharing agreements, based on licensing or fee discounts or share ownership, on an ad-hoc, case-by-case basis. Efforts to establish a stable framework encountered intractable and not fully resolved issues. For example, some stakeholders surfaced concerns that fee discounts or shareholding might distort the market and commit the NHS to products that would become obsolete. This participant outlines the perceived lack of a robust commercialisation process.
… it wasn’t clear what the [commercialisation] process was, and they were kind of building that process and filling that process on the fly.
Interview 56, Supplier
AI Awards for less mature products presented fewer challenges than more mature products as investments were lower and concrete benefits were further from being realised. The National Institute for Health and Care Research (NIHR), which managed these projects, had mechanisms for research funding that involved fluid arrangements in relation to measuring and recovering benefits. Early-stage approaches to benefit sharing avoided applying commercial terms, as the value of these technologies was not yet known.
Intellectual property (IP) considerations added another layer of complexity. Established companies, with a strong background in IP, were better able to get their products adopted by the NHS, while those developing foreground IP offered better commercial value. This tension underscored the challenge of aligning commercialisation strategies with both market dynamics and public-sector goals. Careful balance is needed to avoid market distortions that could stifle competition or unfairly advantage specific suppliers. This participant involved in contracting described the delicate balance between government intervention to promote a thriving supplier ecosystem and adversely influencing the market.
So, we usually blend the model, so we’ll say, you know, if you do that, we’ll do, we will take X percentage, but if you go for the more commercial route, we’ll take Y and that sort of gives the company a bit of flexibility. But yeah, I think the principle is one that is sound, but you have to be sort of careful about how you apply it to ensure that you’re not adversely influencing the market or making the funding you know unattractive because that’s not what anyone’s here to do.
Interview 52, External Contractor
The AI Lab had many stakeholder groups, but some were more actively involved in strategy and delivery than others. Participants mentioned that frontline health and care service delivery staff, in particular, had limited involvement, resulting in activities in some instances not being sufficiently aligned with system needs. For example, some observed that the range of applications supported by the AI Awards was influenced by supplier interest, rather than emerging from a bottom–up process of service delivery staff to establish health system requirements. Similarly, some suppliers identified an opportunity to secure funding for AI products without ensuring their products were tailored to service delivery and patient priorities. We also observed limited development of operational models and associated infrastructure (i.e., application of AI to improve efficiency of service delivery), with most projects focusing on clinical applications. Establishing a need before developing technological solutions is essential.
… you start with the problem with the clinical need and for whom. So rather than you start with a solution which is AI and who’s regulating it… So, I think that it’s basic we just start there. Unfortunately, what they did was to…discover that they were in some cases solving the wrong problems.
Interview 14, Academic
Delivery of AI Award projects was hindered by conflicting service priorities and a lack of active service delivery involvement. As alluded to above, we observed cases where, after successful evaluation of projects (both in AI Awards and proof-of-concept work), adopter organisations were unable to sustain the system beyond its initial pilot. This was due to a lack of evidence needed to establish a clear business and safety case for adopting the application into routine organisational practice. The following excerpt illustrates the challenges surrounding evidence and contextual variations.
For certain products, in certain situations, the stumbling block came that the evidence, clinical safety and value people weren’t sure what to do with it outside of its initial context, people didn’t know if it was translatable fundamentally […] both on the financial perspective and on the safety net accuracy perspectives.
Interview 48, Previous AI Lab Member
In addition to these challenges, adopter organisations lacked the capacity, IT skills, and resources to support implementation and evaluation. This was further exacerbated by the pandemic, which shifted attention to COVID-19-related priorities and delayed non-COVID-19 projects.
We are still very much reeling from the COVID epidemic, we’re really feeling the effects of a reduced workforce and huge waiting lists and so actually whereas you know, when we designed the programme. … It was pretty fair to imagine a world in which workforce had a bit of capacity to engage with this. The reality that we’re faced with now is that. …You know even. … Staff that want to see improvements that have a real interest in research and AI simply do not have the capacity to engage in what is, you know, exploratory research based.
Interview 12, DHSC