HTN was joined for a deep dive into ambient voice technology (AVT) by a fantastic panel including Wahida Jabarzai, clinical AI and automation delivery lead at University Hospitals of Northamptonshire and University Hospitals of Leicester; and Ravinder Kaur Sahota, group CIO at The Dudley Group and Sandwell and West Birmingham.
Our panel shared their learnings, experiences, and insights from AVT projects, covering clinical impact, risks, coding, governance, regulation, assurance, through to implementation, value, benefits realisation, and sustainability.
Introductions and journey to date
Ravi talked us through her role as group CIO, with responsibility for digital, data, technology, and medical engineering across The Dudley Group and Sandwell and West Birmingham, covering community, acute, and primary care. After doing a lot of research and identifying specifications for AVT around areas such as direct EPR integration, the group went out to suppliers to select those that could meet those requirements. “We knew there wouldn’t be anything on the market that was going to fit those immediately, as at that point nobody had developed full EPR integration in the UK,” she commented, “so we knew we were going to have to work in development with somebody to deliver that.”
The group was also approached about being part of a national AVT pilot, Ravi continued. “We trialled two across our organisation, and we were also using Microsoft Dragon at that point, as well, so we decided to do a comparison against the findings from that.” The trial ran in 2025, and building the functionality needed to map out the templates and areas for it to be trialled in took a few months, she said. “We decided to trial in our same-day emergency care service, and did our baseline assessment there before the trial started. We also trialled in rheumatology outpatients, which had the biggest backlog of letters waiting to be typed for patients.”
Results were published and shared after the first three months, with learning shared across the region, Ravi told us. The pilot was then expanded into community and paediatrics, and is being trialled now in inpatient areas. “We wanted to make sure everything was done correctly, so we undertook detailed assessments, penetration testing, DCB compliance, clinical safety, IG, and so on,” she continued. “As a result of that, we were asked to lead a Midlands-wide procurement of AVT in conjunction with the regional team, which was awarded very late last year, and which has 1,149 individual organisations who can draw down on that procurement.”
At present, the roll out plan is waiting to be approved, Ravi reported, with further work being done in the meantime on technical items, and with full EPR integration achieved. “That’s gone through alpha and beta testing,” she explained, “and we’re working on other items such as clinical coding, where we’ve entered the formal testing phase. We’ve had our first two testing rounds, and we started looking at our deployment last week.”
Wahida shared some details on her role as clinical AI and automation delivery lead at UHL and UHN, a large group with multiple sites. “We have got a division for digital to trial emerging technologies like AVT, where we’re trying to be frontrunners,” she said. “We were very interested in trying out AVT to record consultations and use natural language processing to create things like clinical notes and letters to be sent directly to patients via an integration. We decided to go ahead with that back in 2024, and we had to look carefully at how to approach it as it was a very new technology.”
Ultimately, UHL and UHN decided to go ahead with a pilot across four different suppliers, Wahida noted, beginning in early 2025. The project went through governance and was conducted with a select group of clinicians, which allowed benefits and nuances to be observed with the tech, offering insights into what was required so a detailed specification could be created for procurement. “After going out to procurement in Summer 2025, we chose our current supplier, Accurx, which uses Tandem’s AI scribe. Earlier this year we tested that with clinicians in a testing and development phase, before we went out for the full integrated solution that we worked on in parallel.”
Doing that allowed for the team to hone in on the technology itself, according to Wahida, who highlighted how it supported preparation for wider roll out in terms of clinical, operational, and safety perspectives. Product improvements and feedback from clinicians was addressed during that phase, she went on, and currently the group is “very close” to completing its HL7 bi-directional integration via HL7 FHIR. “We are hoping in the coming weeks to go live initially at UHL, and then follow that with UHN as well.”
Challenges and what worked well
Wahida discussed some of the challenges that her team encountered when approaching AVT, areas which could maybe be approached differently in the future, and things that worked particularly well in comparison. One challenge was compliance, as that proved “a little bit tricky” in terms of clinical safety, she noted, as well as cyber security and DPIA from an IG perspective. “As we’re a group, we needed to be sure every angle was covered, getting lots of different people involved in different teams and coordinating across sites was a bit of a challenge,” she said. “AI is also very new, so we needed to be sure we met the right governance standards from an AI perspective – we do have an AI board, and what really helped was making sure we had escalation routes, had the right people in the room, and had the right expertise to help us get that across the line.” Starting the process a little bit earlier could have been a benefit, Wahida offered on reflection.
“We’ve had great feedback from clinicians that AVT is saving time in clinic and allowing a greater focus on patients; but from a change perspective, it’s still a new technology and something people aren’t used to,” Wahida considered. “It can be a challenge to make sure people are on board and their voices are heard, especially when dealing with thousands of clinicians – we’re trying to get around 7,000 licenses out.” Efforts to tackle this have included attending clinical senates, getting involvement from chief medical officers, holding webinars, and running dedicated feedback sessions. “One thing we have noted is that personalisation is required,” she said, “because not every specialty requires the same level of detail in consultations, for example.” A dedicated template for each specialty is a way the group is trying to overcome this.
A final observation Wahida offered was the difficulty with levels of noise in the environments where AVT was being rolled out, or other environmental issues like network connectivity. “That did cause issues for some of the suppliers we were working with regarding what the solution could pick up, so addressing that has been helpful in dealing with that early on. We did things like checking internet connection and picking up on any problems with our IT teams, and in areas like fracture clinic we have dedicated lapel microphones to ensure minimal disruption.” The group is also looking at MDT meetings and how to make sure every voice is captured, she highlighted.
At The Dudley Group and Sandwell and West Birmingham, AVT means clinicians can hold conversations with patients whilst the solution captures what is being said, ready to be checked once the patient leaves to be saved to the EPR, Ravi shared. “Our clinics were historically always overbooked, so clinicians were always running over time, but as a result of the pilot we have found clinics are actually running to time,” she reported. “Community nurses were able to see an extra two patients per day, and in community I think you see that more, because people often have to go back to a base or capture things later on.”
It’s important not to include clinics being overbooked in your baseline, as that work is being done by clinicians in their own time, meaning it isn’t bankable time, Ravi suggested. The other thing to take into account is time for annual leave, sickness, and so on. “All of our benefits have that 30 percent taken off straight away to account for that, as you know when you’re rolling it out that clinicians aren’t going to be working 24/7. Having said that, we’ve seen a real shift in terms of the cognitive loads on our clinicians, and we’ve seen benefits in retention across the organisation. Feedback tends to be that this is the best thing they have ever used, and you can just see the impact it’s having.”
Optimisation and measuring benefits
Ravi pointed to the need to be able to optimise the product in order to get the best out of it, noting that having already completed work on mapping templates for EPR made it easier. “Clinicians may want a template a certain way, which is fine,” she stated, “just don’t open yourself up to having 30 different variations of each one, because that will get very messy when it comes to implementation. AVT will take time to learn accents and other nuances; it will take time to learn abbreviations and how you would like outputs; and you can have multiple different outputs from one consultation, like a more formal letter for a GP and an easy-to-read letter for patients.” It takes two-to-three weeks to optimise per clinician, she noted, “but once that is done it’s very accurate”.
Benefits are also being seen in improved EPR data quality, according to Ravi. “Due to clinicians being busy, sometimes by the time notes are made on a patient, they’re a very shortened version of what actually happened,” she said, “so the use of AVT has meant an observable improvement in the kind of quality of data in our EPR. We’re also finding in our trial in clinical coding that the solution is picking up much more detail than our coders would, and they are all very impressed with it.” It has been known for AVT to make suggestions that aren’t entirely accurate, she noted, so it’s always important to check for accuracy, but in general that is much less of a burden then previously when having to type things up and enter things into the patient record.
UHL and UHN has captured over 7,000 “scribes” from clinicians trialling the solution in real-time since January, revealed Wahida. Learnings from that are being captured regularly through weekly or bi-weekly drop-in sessions, and clinicians can also offer feedback on Teams. “If we have any issues with hallucinations and so on, we’ve got our team on the ground from our supplier to address them, and we get a weekly report from them, and a dashboard they have created to see the latest figures.” On a weekly basis, a review can identify if there is an issue with the amount of edit times, she went on, which usually indicates either inaccuracies or inconsistencies with the output. “Going back to clinicians helps pinpoint what those problems are, and typically we find it can be due to the template not being optimised enough to capture what is needed, so it has to be edited a lot.”
Hallucinations are being taken seriously, Wahida stressed, with plans being put in place even for post go-live to remain on top of things and maintain engagement with the supplier. “In terms of reporting, we had a baseline survey that was sent to clinicians to understand where they were at a particular point in time, and then a regular survey we’ve sent out throughout the project to capture those learnings. The most important thing is to check in with end users, as they are ultimately the ones who will be able to tell you if there are any issues with the technology – it’s been really helpful to do that via various different channels.”
Moving from pilot to deployment
Clinical leads are key when it comes to looking at deployment, Ravi discussed, making the suggestion that having a clinical lead in each service to help drive roll out forward is “really helpful”. The deployment plan across The Dudley Group and Sandwell and West Birmingham is fully mapped out, and restrictions have been put in place to limit the number of templates that can be used within a service. “We’ve taken the decision to prevent anything new from being created in service areas until they’ve been onboarded, because you don’t want to be creating new documents whilst rolling out the solution into existing documents,” she told us. “We now have a two-month running period that will give us time to get the templates ready to go and engage with those services – we’re going to have drop-in sessions, and we’ve requested our supplier comes in and shadows the service during that time to fully understand the flow of patients and how our clinicians work.”
Ravi shared how EPR integration means clinicians across the group will be able to click a button to commence the use of AVT, which will start the recording, and then that will sit in the background, before reminding clinicians to review and save. If things like test results remain outstanding, it can be saved later on, once that has been attached, and if there is a letter to be sent, that will be automatically generated and sent to the intended recipients. “We’re going to be rolling out to outpatient areas first, as that’s likely to be safer and fits nicely with work we’re already doing with clinical coding. And then we have our inpatient rollout that’s also already been very well defined and allows us to roll out to areas with highest usage.”
In the UHL and UHN group’s pilot of different AVT solutions, it observed differences in how each picked up audio, with some reportedly able to do that “much better” than others, Wahida detailed. Other differences included in the content generated, and whether or not the solution was able to work offline. “One thing we noticed very early on when we went live with a non-integrated solution in the testing and development phase was that the character limit was very small – if you’re working with a speciality that tends to need to send longer letters, that can become a bottleneck very quickly.” This was something taken to the supplier to be improved, she told us, which had since been a great benefit for clinicians.
Moving on to talk about personalisation, Wahida noted how the group used champions to help the supplier develop a template builder function that allowed clinical champions to build them for their specialty. “We had clinical leads from each specialty to do that, and we’ve got templates that clinicians can pick up that are very relevant to their area and how they actually conduct their consultations. Other things are like personal memories, so things specific to the clinician, allowing them to add that element to their consultation.” User acceptance testing has been another essential part of the process, she explained, “and that’s not only been limited to testers; it’s also included clinician testing and having them involved in the whole process”.
Wider learnings
The solution chosen by UHL and UHN has been certified as a Class II medical device, Wahida discussed, with that being something that was included in the specification to be in place by the beginning of the contract or shortly thereafter. Another thing the team looked at was translation, and whether there was a solution that negated the need for a translator. “We haven’t ultimately gone with that – it does pick up different accents, and if a consultant were to speak in their own language and a patient does too, it can pick up those languages, but that’s not one of the main purposes we’ve used it for,” she told us. “If you do have an interpreter, it’s important that you indicate that to the AI, so it knows who person A is, who person B is, and it will pick that up for you.”
Ravi talked about consent, and the process in place around informing patients about the use of AVT. “We inform them it’s in use,” she said, “and we explain what that means and ask for their permission when they come in for their consultation.” The response overall has been “overwhelmingly positive”, however, from patients, with feedback covering greater interaction during the clinic, more eye contact, and having more of a conversation rather than the clinician being busy typing into a computer.
For UHL and UHN, the decision has been to use implied consent, with patients being informed about the use of AI to record conversations, rather than being asked specifically for their agreement, Wahida reported. If patients voice their preference not to have that used in their consultation, the clinician can then choose not to use it, she continued, “but they can then use it after the consultation to produce a summary of what happened, just so it allows them to still generate the letter and so on”. Work has been done to engage with patients through specifically designed engagement sessions to gain feedback on whether current efforts to communicate about the use of AVT are effective, and what language could be used to make it clearer, she added.
Future outlook
Considering what use of AVT at the UHL and UHN group might look like in a years’ time, Wahida voiced hopes to have rolled out at UHL. “We’ve taken a phased approach with outpatients first, with the specialties within that phased as well, to be sure we have the right level of support to do that as safely as possible. The secondary phase will include things like MDT meetings, and the third will be inpatients.” A similar approach will be taken to roll out at UHN in the coming months, she went on, “and ideally I’m hoping in 12 months to be in a place where we’ve got clinicians using the system as effectively as possible for them, where it’s saving them time, and it’s much more integrated and embedded into their day-to-day work”. Continuing to focus on emerging concerns and developments will help with optimisation, and there will then be the chance to look at what else is possible with AVT, she concluded.
Wahida also highlighted: “At the moment we are looking at how to use our integration engine to enable patients to receive a digital copy, and we’re hoping in the next couple of months to be able to start storing documentation in the EPR. We haven’t fully explored clinical coding, either, so I’d like to see us able to move forward with that, and to start to address the more nuanced areas, and the specialties that maybe work a little bit differently.” By making those changes, it will be possible to make a “massive difference” for patients, she stated, “so we’ll be able to give them as smooth of a service as possible by next year”.
For The Dudley Group and Sandwell and West Birmingham, Ravi talked about being “lucky” to have an integration engine and an integration specialist, allowing more possibilities around accessing data and so on. “As I said, we’ve run alpha and beta testing and found it much faster,” she continued, “and it’s picking up on things that perhaps our clinical coders weren’t able to. It’s now being tested over a range of specialties with hundreds of patients to make sure it’s actually working, and we’ve been doing coding for both inpatients and outpatients during those testing cycles.” Having full EPR integration means being able to look at more potential opportunities for how AVT could be used, she went on, like AI triaging, referral management, or task management, “as long as you get the integration right and map out the correct workflow”.
We’d like to thank our panellists for taking the time to share these insights with us.
