HTN was joined for a deep dive into AI strategy, implementation, adoption, and opportunities by Neill Crump, group associate director of innovation & partnerships at The Dudley Group and Sandwell and West Birmingham, and Pip Hodgson, group digital transformation specialist at University Hospitals of Leicester (UHL) and Northamptonshire (UHN).
Our panel discussed their organisation’s approaches to AI and AI strategy, best practices in AI strategy development, Ambient Voice Technology and successful implementation, and the opportunities likely to be ahead with the next wave of AI.
Starting out with some introductions, Neill shared a bit around his trusts’ approach to AI, including a focus on solving frontline challenges, and splitting benefits into different areas across productivity, safety, experience, and so on. “We’ve adopted a framework called the MITRE framework, which allows us to take a closer look at maturity and guides how that is developed over time,” he said. “The second pillar is then the governance aspect, making sure we’re responsible and have oversight, and the third pillar is around workforce capability – we’re rolling out Multiverse, which is a data academy helping clinical and operational roles get the most out of AI, because our goal isn’t just to deploy pilots; it’s to build that organisational capability.”
Pip told us about her role, taking responsibility for delivering a range of AI and automation projects across UHL and UHN, as well as the groups’ plans to build a Centre of Excellence for Applied AI. “Part of that will be how we use AI or emerging technologies that use AI, in a safe way,” she shared. “Using a sandbox environment or simulated environment, seeing what works best, and then taking that forward through live rapid deployment. Some of that is being brave, like we’ve been doing recently with Ambient Voice Technology (AVT) and going out with Europe’s largest procurement tender with 19,000 licences to be delivered over three years.”
Procurement and funding
“I actually think the procurement commercial models are a major determinant of whether AI adoption is going to succeed in the NHS,” Neill considered, “and we need to be thinking about how we work together in order to do that – a lot of current frameworks focus on traditional software which doesn’t change much over time; AI is evolving rapidly, so we need that continuous evaluation.” Dudley and Sandwell have just completed a regional procurement for AVT, he continued, which allowed greater testing of the market, a comparison of suppliers, and better cost efficiencies. “It’s important to evaluate multiple suppliers so you can benchmark their performance,” he suggested, “and if we coordinate, rather than having hundreds of separate buyers within the NHS, we can negotiate better and reduce duplication, which will allow for standardisation, sharing learning, and evaluation.”
Pip talked about going straight from paper to AVT in some scenarios, and needing to take a step back to look at what was realistic in terms of what UHL and UHN needed from the technology. “I really encourage people to think about that early on,” she said. “What capabilities do you need? Is it a standalone system or will it be integrated? We’re obsessed with open standards in the NHS, but the reality is that just looking across my own patch we have multiple, multiple different systems.” Drawing up a heat-map of different solutions and suppliers, as well as what they offer and what the organisations’ needs are, before moving on to think about scale, is key, she went on. “AVT for us is anywhere that generates a conversation output, and predominantly the business case focused on outpatients.”
On average, the trusts have 8,000 users per year using the technology, Pip explained, meaning that by 2028 the majority of those responsible for any kind of clinical documentation should have access to it. The pilots tested four different suppliers, which were all slightly different in functionality, and which all produced different results, she said. “Other things to come out of the pilots were questions about risk management, ongoing governance and monitoring, AI degradation and putting in place an acceptable percentage threshold, and how to support staff with using it. The path of resistance can be high, and the pilots will capture that for you, which makes going to tender much easier.”
AI strategy and deployment
Although Dudley and Sandwell do not yet have an AI strategy, Neill shared that they have created a three-year costed and funded digital plan that covers AI as one of its components. That plan is intended to focus on what can be achieved over the next three years to make a difference, offering a solid plan for implementation, and looking at what work can be done to demonstrate improvements. First year focuses include developing solid foundations, to prepare for years two and three which are more concerned with experimentation, he outlined. “From that point, we’ll probably craft out a specific AI plan so everyone in the organisation and our system partners can understand it, and we’ll have something tangible on the way we’ve gone about things like AVT,” he noted.
Sharing further insight into the use of AVT at the trusts to date, Neill told us how his team took the approach of rolling out first in specialties like rheumatology, making sure to understand what benefits existed before rolling out further. “We had a big letter backlog we wanted to resolve, for example,” he said, “and we wanted to be able to prove we could do that. We then looked at it end-to-end for outpatient journeys, and got some great clinical benefits off the back of that. The third part has been the actual clinical coding and working out how we can embed that within the software, so rather than having a separate way of doing it we’re co-developing that with a partner and putting that into the solution itself.”
Balancing innovation and evaluation
Pip offered some insights around balancing innovation and evaluation, pointing to rapid evaluation and user testing as good places to start. “Evaluation for me is around whether it is any more or less dangerous than what we’re doing already – when we look at outpatients, some people are waiting up to ten weeks to get a letter, so we didn’t have a great place to start from in terms of those metrics,” she reflected. “We also did some time in motion metrics to understand how much time is spent on admin or clinician tasks, and what was interesting was that whilst it might not save loads of clinical time, the evaluation showed us that AVT reduced the number of our clinical workforce staying late to finish clinical documentation from 78 percent to 12 percent.” Other impacts observed included reducing or removing the backlog altogether for letters.
“We have to think about what is tangible to achieve with the rollout, what metrics are important, and how we can validate,” Pip went on. “Then look at how you can assure it is as safe as what we have today, not less safe, because that’s always what comes back – what if the AI is less accurate? The answer to that is we should have a human-in-the-loop process, with someone who is trained understand and mitigate when things go wrong, and we should have a fallback mechanism of what our safe parameters are. For me, evaluation needs to be real world, evidence-based, and adaptable.”
Neill described taking a similar approach, adding that it’s also important to approach validation through small, controlled deployments to begin with. “As well as rheumatology, we also did a pilot in same day emergency care which gave us very practical evidence around reductions in transcription, backlog, faster letter turnaround, relieved burden on clinicians and secretaries,” he shared. “With having a clinician in the loop, it’s almost like an augmentation model, which then brings the risk down – I think AI deployments are the ones that augment clinicians rather than actually automate those clinical decisions.”
Talking about the MITRE framework, Neill explained its six pillars and five assessment areas, adding that he and his team have taken time to link to What Good Looks Like. In general, it has helped to understand how capabilities are being developed over time and to frame the next steps, to keep progress going on an organisational level, not just on an individual project level, he concluded.
Governance
UHL and UHN have signed up to the Trustworthy and Responsible AI Network (TRAIN) and set up an AI governance office, according to Pip, offering the ability to centrally manage, monitor, and report on risks when deploying AI. That work has included creating an AI and automation asset register to keep track of what is in use across the organisations, who the service owners are, functionality, and things like data use. “The government are also looking at introducing an AI algorithm register to allow people to publish what algorithms they ae using and what data is being shared,” she continued, “so we can do more to be open and transparent there.”
An AI board has also been established at UHL and UHN to build understanding of risks and how to manage them, considering AI degradation, where systems are not working as intended, and setting guardrails for accepted levels of inaccuracies or tolerance. “When delivering things like AVT, there can be infrastructure challenges like WiFi black spots and putting more load on an already struggling system isn’t going to be useful, so we try to measure our latency speed, which looks at how quickly we can return the information we need,” she said. “We put things on cloud and it improves somewhat, but we need it to get better, because clinicians can sit there waiting up to eight minutes some days to generate a letter.”
“We’re also still learning on this,” Neill acknowledged. “We’ve got frameworks like DTAC and clinical safety safeguards, but those are more focused on digital solutions in general, not necessarily AI. With AI, it’s different because it’s continually moving and changing, so it needs to be monitored differently to normal digital solutions.” Dudley and Sandwell have worked with the ICS to collectively develop a set of joint SOPs, which has had the benefit of bringing in different expert advice that might not exist within the organisation. That has been “really successful”, he considered, in helping pinpoint when AI outputs should be overridden, what fallback processes are, and so on.
Ensuring success with pilots
Neill and Pip moved on to offer some insight into their own experience with pilots and best practice in making those successful. “Make it fun!” Pip advised. “Pilotitis is absolutely a thing, and we’re crazy if we think people have time on their hands – I’m an ED nurse, and the idea that I use another piece of technology when I’ve got a huge waiting list of patients is ridiculous.” With this in mind, her team decided to make pilots small and purposeful, running for a maximum of 12 weeks, and deliberately targeted toward digitally-minded individuals. “We wanted people already engaged in this space, because we truly wanted to test the technology and whether it delivered the right results, so we could go out to tender. I think that’s what made the pilots so useful.” With AVT, the 12 weeks offered just over 1,000 completed scribes, granting a good amount of data, she added.
Thinking of something as a clinical capability that you want to introduce, rather than a technology experiment, is key, Neill advised. “We’ve made sure we have strong clinical ownership, with clear governance frameworks in place, and that we’re focusing on placing things into the actual workflow of clinicians rather than having a separate process outside.” Scaling is much easier if done in collaboration across the system, he went on, “and that way you can make sure you’re sharing learnings and also approaching procurements at scale”.
Developing AI literacy
Neill recommended taking a practical approach to AI literacy training, building that alongside scaling. “We’ve also got the Multiverse approach, where we’re giving people the opportunity to go on a data academy as either a one-year or three-year course, which is a great opportunity to develop champions across the organisation.” Using a skills framework is good for building confidence and capability, as well as getting people to realise the kind of capabilities they are building over time just through some of the work they’re doing implementing in their day-to-day roles. “We use Skills for the Information Age, where we recognise what all the different skills are that are used in the design and deployment, all the way through to post-evaluation. We’re also taking that a step further and mapping those skills to roles to provide career development opportunities for people, showing AI capability is actually part of professional practice.”
“Digital literacy isn’t a term I particularly like,” Pip admitted, “because people don’t know if they’re necessarily digitally literate; they know whether they’re confident using the tools that they have; and to some extent, it’s down to us as an organisation to make the tools easy to use.” Failing to make that happen can have an impact on digital confidence, she continued, “as if people have an issue they will quite often be put off using things again – it’s almost like one strike and you’re out”. UHL and UHN are also using Multiverse, and have set up a number of leadership academies and AIs to build confidence in anything like rostering or automation, where they need to be confident. A change champion model has been set up specifically for AI, and there is also a tracker covering what capabilities are out there, what tools are in use, and how often they are being accessed.
Pip credited the change champion network for helping lead the charge, noting the importance of having people on the ground to offer realistic feedback. “Quite often, it isn’t the AI itself; it’s how it’s being implemented that is the frustration. For us, it was getting feedback from the grassroots level, then coming back to our development teams to make what we have in place better, and information accessible when we think about training.”
Looking ahead
Having reliable, accessible, and structured data is the starting point for AI, Neill told us, to help generate insights, that in turn demonstrate value and build clinician trust. Once that trust is in place, tools can start to be adopted into everyday workflows, which releases the benefits in improvements to care and productivity. “As those systems are used, they then generate better structured data to help toward the next cycle of innovation – it’s getting that continuous loop in place,” he continued, “and if we think in that way, it will actually accelerate the way we’re doing things.”
UHL and UHN have a bold strategy set out by their CEO for AI and automation, seeking to replace manual tasks by 2027, according to Pip. “The reality of that is breaking down legacy systems and looking at things like NHS App use and EPR optimisation – the problem is, we don’t have a lot of money, and procurement takes a long time. As well as having a clear vision of what we want to do and a clear strategy to achieve it, we need to be pragmatic about the time it takes in reality.”
In five years’ time, AI will largely be invisible, Neill predicted, as it will have become part of the everyday infrastructure. Data will be being collected from wearables and home monitoring to inform a continuous view of health, rathe than looking at individual episodes of health. “We’ve spent a lot of time over the last three years creating a data platform, putting in a data lakehouse, the structuring, the semantic models, and so on, to give us the foundation layer and infrastructure that will allow AI to work on top and give us predictive models, early warning systems, decision support, and things like that.”
“I’d just like to think we could have managed to get rid of pen and paper, reducing the amount of printing and the number of manual tasks,” said Pip. “I would also love for AI to become less of a dirty word – people are scared of it, and it’s really new technology, but if we look at other industries they are using it ten times better than we are.” Patients are already accepting of the fact things like AVT are in use, she shared, “and when we talk to them to let them know, they will ask us why we are asking them, and tell us that they already thought we were using it”.
“We had some clinicians in critical care that were really interested in machine learning,” Neill told us, “and they’d created a dataset over many years – what they found working with us is that by going through that process, they could redesign and understand how they could better deliver care. They’re now asking whether certain outcomes could have been different if they could have had that information at the point of care – I think that’s working toward predictive care.” One of the barriers is the timeline that you have to actually developing a medical device, he suggested, “but it’ll be interesting to see how we can do some of that in the future”.
We’d like to thank Neill and Pip for sharing these insights with us.
