Procurement and early deployment of artificial intelligence tools for chest diagnostics in NHS services in England: a rapid, mixed method evaluation

May combine two claims.

A recent review found substantial gaps in real-world evaluation of implementation and use of AI for diagnostics in clinical settings, including no studies on procurement.

“A recent review notes substantial gaps in real-world evaluation of implementation and use of AI for diagnostics in clinical settings, for example, it found no studies on procurement.16”

• cites [16] Lawrence R. Dodsworth E. Massou E. Artificial intelligence for diagnostics in radiology practice: a rapid systematic scoping review eClinicalMedicine 83 2025 10.1016/j.eclinm.2025.103228 PMC12140059 40474995 sourceSupported

May combine two claims.

Little is known about real-world implementation of AI in radiology diagnostics, including procurement, preparation for deployment, experiences of staff, patients, and carers, and impact on effectiveness and costs.

“However, little is known about real-world implementation, including procurement, preparation for deployment, experiences of staff, patients, and carers, and impact on effectiveness and costs—due in part to very few studies having been conducted on real-world implementation of AI tools for radiology diagnostics.16”

• cites [16] Lawrence R. Dodsworth E. Massou E. Artificial intelligence for diagnostics in radiology practice: a rapid systematic scoping review eClinicalMedicine 83 2025 10.1016/j.eclinm.2025.103228 PMC12140059 40474995 sourceSupported

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

Improving chest diagnostics and lung cancer detection are national priorities due to high rates of late-stage lung cancer diagnosis.

“Improving chest diagnostics and lung cancer detection are national priorities due to high rates of late-stage lung cancer diagnosis,3 and efficient chest diagnostics may be central to managing any future respiratory disease pandemic.19”

• cites [3] National Institute for Health and Care Excellence Artificial intelligence-derived software to analyse chest X-rays for suspected lung cancer in primary care referrals: early value assessment Available from: https://www.nice.org.uk/guidance/hte12/resources/artificial-intelligencederived-software-to-analyse-chest-xrays-for-suspected-lung-cancer-in-primary-care-referrals-early-value-assessment-pdf-50261967918277 2023Not checked yet

  Not checked in the free run.

• cites [19] Inui S. Gonoi W. Kurokawa R. The role of chest imaging in the diagnosis, management, and monitoring of coronavirus disease 2019 (COVID-19) Insights Imaging 12 2021 155 10.1186/s13244-021-01096-1 34727257 PMC8561360 sourcePartly supported

  Citation 19 attaches to the sentence’s second clause, not the national-priorities claim cited to reference 3. The paper’s Key points and Controversy of chest CT versus RT-PCR sections describe roles for chest imaging in managing COVID-19, including assessing progression and triaging patients. They do not address whether efficient chest diagnostics would be central to managing a future respiratory disease pandemic. (LLM, conf=0.82)

  Experimental — accuracy not yet measured

Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing).

Chest diagnostics may be key to managing any future respiratory disease pandemic.

“and chest diagnostics may be key to managing any future respiratory disease pandemic.19”

• cites [19] Inui S. Gonoi W. Kurokawa R. The role of chest imaging in the diagnosis, management, and monitoring of coronavirus disease 2019 (COVID-19) Insights Imaging 12 2021 155 10.1186/s13244-021-01096-1 34727257 PMC8561360 sourcePartly supported

  The Abstract describes chest imaging as integral to the work-up and staging of COVID-19, particularly for assessing complications and disease progression. The Introduction limits the review to COVID-19; the paper does not establish that chest diagnostics will be key to managing any future respiratory disease pandemic. (LLM, conf=0.85)

  Experimental — accuracy not yet measured

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

Evidence suggests AI might benefit diagnostic services by supporting decision-making, improving detection accuracy, reducing errors, increasing efficiency, and easing workforce burdens.

“Evidence suggests AI might benefit diagnostic services by supporting decision-making, improving detection accuracy, reducing errors, increasing efficiency, and easing workforce burdens.13, 14, 15”

• cites [13] Liu M. Wu J. Wang N. The value of artificial intelligence in the diagnosis of lung cancer: a systematic review and meta-analysis PLoS One 18 2023 e0273445 10.1371/journal.pone.0273445 PMC10035910 36952523 sourcePartly supported

  The Introduction (§1.2) says AI can help physicians screen for early-stage lung cancer while reducing errors and increasing efficiency. The Discussion (§4) says AI-assisted diagnosis has high diagnostic value and anticipates improved accuracy and reduced physician workload. These passages support the proposed benefits for lung-cancer diagnosis, but do not establish them across diagnostic services generally; the workload benefit is presented as a future prospect. (LLM, conf=0.85)

  Experimental — accuracy not yet measured
• cites [14] Joy Mathew C. David A.M. Joy Mathew C.M. Artificial Intelligence and its future potential in lung cancer screening Excli J 19 2020 1552 1562 10.17179/excli2020-3095 33408594 PMC7783473 sourceSupported
• cites [15] Chiu H.Y. Chao H.S. Chen Y.M. Application of artificial intelligence in lung cancer Cancers (Basel) 14 2022 1370 10.3390/cancers14061370 35326521 PMC8946647 sourceSupported

May combine two claims.

The use of AI tools as decision support, rather than operating autonomously, aligned with existing UK regulation and public attitudes.

“The use of AI tools as decision support, rather than operating autonomously, aligned with existing UK regulation and public attitudes.16”

• cites [16] Lawrence R. Dodsworth E. Massou E. Artificial intelligence for diagnostics in radiology practice: a rapid systematic scoping review eClinicalMedicine 83 2025 10.1016/j.eclinm.2025.103228 PMC12140059 40474995 sourceSupported

Classified as probably not a substantial claim.

These findings underscore the value of engaging stakeholders throughout the process.

“These findings underscore the value of engaging stakeholders (e.g., national leadership, network teams, Trust clinical, managerial, and technical staff, and AI suppliers) throughout the process, reflecting previous published research on implementing AI in radiology services.25,26”

• cites [25] Wenderott K. Krups J. Luetkens J.A. Weigl M. Radiologists’ perspectives on the workflow integration of an artificial intelligence-based computer-aided detection system: a qualitative study Appl Ergon 117 2024 104243 10.1016/j.apergo.2024.104243 38306741 sourceNot checked — no full text of the source
• cites [26] Kim B. Romeijn S. van Buchem M. Mehrizi M.H.R. Grootjans W. A holistic approach to implementing artificial intelligence in radiology Insights Into Imaging 15 2024 22 10.1186/s13244-023-01586-4 38270790 PMC10811299 sourceSupported

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

Many factors identified in our analysis reflect established findings on implementation of other innovations and NASSS framework constructs.

“Many factors identified in our analysis—including context (at system, organization, and service levels), leadership and engagement approaches, technological considerations, use of data, and local capacity and resourcing (e.g., finance, people)–reflect established findings on implementation of other innovations, digital or otherwise,21,27, 28, 29, 30 and NASSS framework constructs.20”

• cites [21] Shin H.D. Hamovitch E. Gatov E. The NASSS (Non-Adoption, Abandonment, Scale-Up, Spread and Sustainability) framework use over time: a scoping review PLoS Digital Health 4 2025 e0000418 10.1371/journal.pdig.0000418 PMC11913280 40096260 sourceSupported
• cites [27] Parmar J. Sacrey L.A. Anderson S. Facilitators, barriers and considerations for the implementation of healthcare innovation: a qualitative rapid systematic review Health Soc Care Community 30 2022 856 868 10.1111/hsc.13578 34558143 sourceSupported
• cites [28] Santos W.J. Graham I.D. Lalonde M. Demery Varin M. Squires J.E. The effectiveness of champions in implementing innovations in health care: a systematic review Implement Sci Commun 3 2022 80 10.1186/s43058-022-00315-0 35869516 PMC9308185 sourcePartly supported

  The Results (Description and operationalization of champions; page unavailable) describe champions promoting innovations and taking leadership roles, including acquiring resources. The review also examines technological innovations and organizational uptake, but it does not establish the full set of factors listed in the citing sentence, particularly use of data and context across system, organization, and service levels. The NASSS framework is attributed to a separate citation, not this source. (LLM, conf=0.83)

  Experimental — accuracy not yet measured
• cites [29] Syeed M.S. Poudel N. Ngorsuraches S. Veettil S.K. Chaiyakunapruk N. Characterizing attributes of innovation of technologies for healthcare: a systematic review J Med Econ 25 2022 1158 1166 10.1080/13696998.2022.2140591 36301001 sourceCould not verify

  The Discussion and Conclusions (pages unavailable) identify attributes used to recognize innovative healthcare technologies, including novelty, benefits and acceptable cost. They do not address whether the cited factors—such as leadership, data use and local capacity—are established findings about implementing innovations. Searches for implementation factors, leadership and resources found no passage establishing that link. The separate NASSS-framework clause is attributed to citation 20, not this source. (LLM, conf=0.73)

  Experimental — accuracy not yet measured
• cites [30] Milella F. Minelli E.A. Strozzi F. Croce D. Change and innovation in healthcare: findings from literature Clin Outcomes Res 13 2021 395 408 10.2147/CEOR.S301169 PMC8141398 34040399 sourcePartly supported

  The Results (Clusters from Keywords Analysis; page unavailable) link implementation of change to leadership, clinician involvement, communication, participation and IT, supporting several of the factors attributed to this source. The cited paper does not establish the full list, particularly use of data and local financial or staffing capacity. The sentence cites a different source for NASSS constructs, so that portion is not attributed to this paper. (LLM, conf=0.86)

  Experimental — accuracy not yet measured
• cites [20] Greenhalgh T. Wherton J. Papoutsi C. Beyond adoption: a new framework for theorizing and evaluating nonadoption, abandonment, and challenges to the scale-up, spread, and sustainability of health and care technologies J Med Internet Res 19 2017 e8775 10.2196/jmir.8775 PMC5688245 29092808 sourceSupported

Probably not a substantial claim: may not pass S3 (load-bearing).

The decision was to maintain a formal training log for diagnostics staff.

“The decision was to maintain a formal training log for diagnostics staff; however, it was recognised that such a solution would be much harder to manage safely if AI were to be used in emergency departments, where staff numbers accessing the diagnostic system was larger and less consistent.”

May combine two claims.

Preparing for deployment occurred at Trust-level, supported by local networks, and included integrating AI, governance processes, staff engagement and training, planning patient engagement, and collating impact data.

“Preparing for deployment occurred at Trust-level, supported by local networks, and included integrating AI, governance processes, staff engagement and training, planning patient engagement, and collating impact data (Fig. 2, box b).”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

Several panel members reported that information provided did not facilitate informed decision-making, potentially increasing likelihood of key details being missed.

“However, several reported that information provided did not facilitate informed decision-making, e.g., due to large amounts of supporting documentation submitted, resulting in some members having to deal with too much information, some of which was seen as excessively technical; this potentially increased the likelihood of key details being missed.”

May combine two claims.

While equity was a required consideration in all procurement processes, interviewees indicated that this did not generally change the selection made, and some perceived this element as 'box-ticking'.

“While equity was a required consideration in all procurement processes, interviewees indicated that this did not generally change the selection made, and some perceived this element as ‘box-ticking’.”

May combine two claims.

Procurement of AI tools involved engaging selection panels, assessing tenders, and contracting AI suppliers.

“Procurement involved engaging selection panels, assessing tenders, and contracting AI suppliers.”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

Meetings highlighted the potential to frame AI implementation as a "good news story" for the NHS and patients, whilst also recognising the importance of transparency.

“Meetings highlighted the potential to frame AI implementation as a “good news story” for the NHS and patients, whilst also recognising the importance of transparency, particularly as patients could learn of AI use independently by accessing their records.”

• AssumptionRests on an assumption: Patients could learn of AI use independently by accessing their records.

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

Network-level procurement meant that networks could potentially choose different AI tools, but that all Trusts within a network would have to use the tool selected by their network.

“Network-level procurement meant that networks could potentially choose different AI tools, but that all Trusts within a network would have to use the tool selected by their network.”

Probably not a substantial claim: may not pass S3 (load-bearing).

The majority of activities observed were governance meetings.

“The majority of activities observed were governance meetings (including national meetings between network leads, local planning and oversight meetings); we also observed a small number of other activities, including a training session and a local review of a patient safety risk assessment.”

May combine two claims.

Where dedicated project managers were present, there was greater capacity to facilitate progress and local knowledge of relevant individuals.

“Where present, there was greater capacity to facilitate progress (e.g., arranging meetings, managing timelines, engaging governance contacts) and local knowledge of relevant individuals.”

May combine two claims.

Changes to workflow involved the prioritisation of scans for reporting and the AI read being visible for staff to review after completing their report.

“Changes to workflow involved the prioritisation of scans for reporting (e.g., one network categorised prioritisation using ‘critical’, ‘high’, and ‘standard’ labels, with critical to be reported first) and the AI read being visible for staff to review after completing their report.”

May combine two claims.

Preparation for deployment involved AI integration, governance processes, staff engagement and training, planning patient engagement, and collating impact data.

“Preparation for deployment involved AI integration; governance processes; staff engagement and training; planning patient engagement; and collating impact data; patient communication plans varied and were still developing.”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

The AI tool was presented on the local diagnostic system and key features were demonstrated and discussed, including highlighting areas of interest in scans, a bar indicating AI confidence in its interpretation, and how to report clinician agreement with AI interpretation.

“the AI tool was presented on the local diagnostic system and key features were demonstrated and discussed, e.g., highlighting areas of interest in scans, a bar indicating AI confidence in its interpretation, and how to report clinician agreement with AI interpretation.”

May combine two claims.

Trusts operated independent governance procedures, with differing meeting structures and documentation templates, making many unwilling or unable to process forms that did not comply with their local requirements.

“Trusts operated independent governance procedures, with differing meeting structures and documentation templates, making many unwilling or unable to process forms that did not comply with their local requirements.”

May combine two claims.

Project meetings were attended regularly by Trust and AI supplier staff, working to agreed timelines, and operating a risk register where integration tasks were a central focus.

“We observed project meetings attended regularly by Trust and AI supplier staff, working to agreed timelines, and operating a risk register where integration tasks were a central focus.”

Probably not a substantial claim: may not pass S3 (load-bearing).

We recruited 10 of the 12 networks participating in AIDF.

“We recruited 10 of the 12 networks participating in AIDF: of the two that did not participate, one had joined the programme late, while the other was facing extensive challenges in implementation; neither felt they had the capacity to participate in the work.”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

The sequence and nature of processes was emphasised, i.e., the requirement for human readers to complete their assessment before checking the AI report, and that the human reader's interpretation should take precedence over AI.

“The sequence and nature of processes was also emphasised, i.e., the requirement for human readers to complete their assessment before checking the AI report, and that the human reader’s interpretation should take precedence over AI.”

May combine two claims.

In some networks, staff reported mixed responses to the prospect of AI, with concerns raised by more senior clinicians about the potential impact of AI making decisions without clinician input, and where accountability lay in the event of a condition being missed.

“In some networks, staff reported mixed responses to the prospect of AI, with concerns raised by more senior clinicians about the potential impact of AI making decisions without clinician input, and where accountability lay in the event of a condition being missed.”

May combine two claims.

Members were responsible for specific tasks, including developing IT capabilities, testing the AI tool using retrospective imaging data, implementing a period of 'shadow mode' before the AI was live in clinical practice, and updating standard operating procedures.

“Members were responsible for specific tasks, e.g., developing IT capabilities, testing the AI tool using retrospective imaging data, implementing a period of ‘shadow mode’ before the AI was live in clinical practice, and updating standard operating procedures.”

Probably not a substantial claim: may not pass S3 (load-bearing).

Three of the 54 people approached for interview did not take part.

“Three of the 54 people approached for interview did not take part: this was as a result of individuals not responding to invitations rather than refusal to participate.”

May combine two claims.

AI tools were operational in clinical services in 24 of 66 Trusts by November 2024 and in 43 Trusts at time of writing (June 2025).

“AI tools were operational in clinical services in 24 of 66 Trusts by November 2024 and in 43 Trusts at time of writing (June 2025).”

Probably not a substantial claim: may not pass S3 (load-bearing).

A Trust Clinical Safety Officer joined an extended project meeting to assess hazard identification (HAZID) documents, enabling faster approval.

“For instance, a Trust Clinical Safety Officer joined an extended project meeting to assess hazard identification (HAZID) documents, enabling faster approval.”

Probably not a substantial claim: may not pass S3 (load-bearing).

One site's hazard assessment raised questions about how the organisation ensured new and established staff were trained before using the AI tool.

“One site’s hazard assessment raised questions about how the organisation ensured new and established staff were trained before using the AI tool.”

May combine two claims.

Challenges to procurement and deployment included engaging staff with high clinical workloads, staff's limited AI knowledge, time to participate, and concerns over appropriate tool usage, managing unsuccessful suppliers' responses, and varied local governance processes, IT systems, and data availability and quality.

“Challenges included: engaging staff with high clinical workloads; staff’s limited AI knowledge, time to participate, and concerns over appropriate tool usage; managing unsuccessful suppliers’ responses; and varied local governance processes, IT systems, and data availability and quality.”

Probably not a substantial claim: may not pass S3 (load-bearing).

Observed super user training sessions were led by the AI supplier team.

“Observed super user training sessions were led by the AI supplier team: the AI tool was presented on the local diagnostic system and key features were demonstrated and discussed, e.g., highlighting areas of interest in scans, a bar indicating AI confidence in its interpretation, and how to report clinician agreement with AI interpretation.”

May combine two claims.

Patient communication plans varied and were still developing.

“patient communication plans varied and were still developing.”

Probably not a substantial claim: may not pass S3 (load-bearing).

Another network would share information only upon patient request.

“while another would share information only upon patient request.”

May combine two claims.

All decisions still had to be made by humans: AI readouts were only to be examined once human readers had conducted their assessments.

“Importantly, all decisions still had to be made by humans: AI readouts were only to be examined once human readers had conducted their assessments.”

May combine two claims.

Another network used the AIDF programme to align their local IT infrastructure to a common PACS platform: although this would potentially facilitate future AI implementation, the complexity and scale of this process delayed deployment of AI tools.

“Another network used the AIDF programme to align their local IT infrastructure to a common PACS platform: although this would potentially facilitate future AI implementation, the complexity and scale of this process delayed deployment of AI tools.”

May combine two claims.

While procurement was coordinated at network level, contracting varied due to the autonomy of NHS organisations and a history of independent procurement and contracting across Trusts.

“While procurement was coordinated at network level, contracting varied due to the autonomy of NHS organisations and a history of independent procurement and contracting across Trusts.”

May combine two claims.

Enablers included programme leadership's support, networks sharing expertise and capacity, committed clinical, technical, and procurement specialists and AI suppliers, clinical champions, and dedicated project management.

“Enablers included: programme leadership’s support; networks sharing expertise and capacity; committed clinical, technical, and procurement specialists and AI suppliers; clinical champions; and dedicated project management.”

Probably not a substantial claim: may not pass S2 (checkable).

Procurement, deployment, and evaluation of the impact of AI are complex social and technical processes, requiring significant resources (e.g., time, capacity, expertise).

“Procurement, deployment, and evaluation of the impact of AI are complex social and technical processes, requiring significant resources (e.g., time, capacity, expertise).”

Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing), S4 (based).

Early and ongoing engagement may have helped address these concerns.

“early and ongoing engagement may have helped address these concerns.”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

It was important that staff knew how to use the tool and understood its limitations, i.e., AI should only act as decision support.

“It was important that staff knew how to use the tool and understood its limitations, i.e., AI should only act as decision support, reflecting the NASSS constructs of ‘work needed to implement change’, ‘staff adopters’, and ‘knowledge needed to use technology’ (Fig. 2, boxes 3, 5, and 6) (quotes: Supplementary file C, section 3c).”

Probably not a substantial claim: may not pass S3 (load-bearing).

It was recognised that such a solution would be much harder to manage safely if AI were to be used in emergency departments, where staff numbers accessing the diagnostic system was larger and less consistent.

“however, it was recognised that such a solution would be much harder to manage safely if AI were to be used in emergency departments, where staff numbers accessing the diagnostic system was larger and less consistent.”

May combine two claims.

Implementing AI involved complex social and technical processes, requiring significant resources.

“Implementing AI involved complex social and technical processes, requiring significant resources.”

• Also stated in: § Funding · boosted

  “Our analysis demonstrates that, in line with many healthcare innovations, procurement and preparation for deployment of AI tools are complex social and technical processes, requiring significant time and resources.”

• AssumptionRests on an assumption: The experiences of the participating networks and Trusts are sufficiently representative to draw lessons for future real-world implementation of AI in diagnostics.

Probably not a substantial claim: may not pass S3 (load-bearing).

There was no change in how data were used.

“as the tools were classified as a medical device there was no change in how data were used”

• cites [23] Information Commissioner's Office Artificial intelligence 2023 Available from: https://ico.org.uk/for-organisations/uk-gdpr-guidance-and-resources/artificial-intelligence/Not checked yet

  Not checked in the free run.

• cites [24] Medicines and Healthcare Products Regulatory Agency Medical devices: software applications 2023 Available from: https://www.gov.uk/government/publications/medical-devices-software-applications-appsNot checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing).

Tools should not influence clinical decisions.

“it was stipulated that tools should not influence clinical decisions.23,24”

• cites [23] Information Commissioner's Office Artificial intelligence 2023 Available from: https://ico.org.uk/for-organisations/uk-gdpr-guidance-and-resources/artificial-intelligence/Not checked yet

  Not checked in the free run.

• cites [24] Medicines and Healthcare Products Regulatory Agency Medical devices: software applications 2023 Available from: https://www.gov.uk/government/publications/medical-devices-software-applications-appsNot checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing).

AI is anticipated to be increasingly central to health services.

“as AI is anticipated to be increasingly central to health services,12”

• cites [12] UK Government Department of Health and Social Care Fit for the future: 10 year health plan for England 2025 Available from: https://assets.publishing.service.gov.uk/media/68760ad755c4bd0544dcae33/fit-for-the-future-10-year-health-plan-for-england.pdfNot checked yet

  Not checked in the free run.

May combine two claims.

Influential factors identified here mirror research on other healthcare innovations, suggesting AI may not address service challenges as straightforwardly as policymakers anticipate.

“Influential factors identified here mirror research on other healthcare innovations, suggesting AI may not address service challenges as straightforwardly as policymakers anticipate.”

• Also stated in: § Funding · hedged

  “While the factors we identify parallel other healthcare innovations, the relative novelty of AI—at policy, regulation, and service levels—meant there was limited guidance and knowledge on procurement and preparation for deployment: this increased uncertainty throughout these processes; further, the influential factors we identify suggest that AI may not address service challenges as straightforw”

• AssumptionRests on an assumption: The experiences of the participating networks and Trusts are sufficiently representative to draw lessons for future real-world implementation of AI in diagnostics.

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

AIDF provided £21 m to services and aimed to increase diagnostic efficiency, reduce backlogs and waiting times, and ease workforce pressure.

“AIDF, which was led by NHS England, provided £21 m to services and aimed to increase diagnostic efficiency, reduce backlogs and waiting times, and ease workforce pressure.”

• cites [18] UK Government Department of Health and Social Care £21 million to roll out artificial intelligence across the NHS 2023 Available from: https://www.gov.uk/government/news/21-million-to-roll-out-artificial-intelligence-across-the-nhsNot checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing).

Staff feedback at our webinar provided some validation of our findings.

“However, staff feedback at our webinar provided some validation of our findings.”

Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing). May combine two claims.

Our theory-driven approach let us consider these processes from diverse perspectives and reflect on how different factors interacted to shape implementation progress.

“Our theory-driven approach let us consider these processes from diverse perspectives and reflect on how different factors interacted to shape implementation progress.”

May combine two claims.

Due to implementation delays, we could not study several important issues, including day-to-day use of AI in clinical practice, impact on care, outcomes, and costs, and adaptation and sustainability of AI tools over time.

“First, due to implementation delays, we could not study several important issues, including day-to-day use of AI in clinical practice, impact on care, outcomes, and costs, and adaptation and sustainability of AI tools over time.”

May combine two claims.

Our lessons may apply less to other NHS organisations that did not take part in the AIDF programme, other healthcare systems with different regulatory arrangements, or to situations where services are commissioning AI tools without financial incentives to do so.

“therefore our lessons may apply less to other NHS organisations that did not take part in the AIDF programme, other healthcare systems with different regulatory arrangements, or to situations where services are commissioning AI tools without financial incentives to do so.”

Probably not a substantial claim: may not pass S3 (load-bearing).

Our rapid approach enabled timely formative feedback of learning to services and programme leadership.

“Our rapid approach enabled timely formative feedback of learning to services and programme leadership.”

Probably not a substantial claim: may not pass S2 (checkable).

The novelty of AI suggests a need for greater understanding of and guidance and education on AI and its implementation at system and organizational levels.

“suggesting a need for greater understanding of and guidance and education on AI and its implementation at system (policy and regulation) and organizational levels.”

Classified as probably not a substantial claim. May combine two claims.

Our hybrid sample let us understand how networks approached procurement and deployment, providing in-depth insights on how preparation for deployment progressed at Trust-level.

“Our hybrid sample let us understand how networks approached procurement and deployment, providing in-depth insights on how preparation for deployment progressed at Trust-level.”

Probably not a substantial claim: may not pass S3 (load-bearing).

Due to the tight timeline of this phase of our evaluation, we were unable to interview several important groups, including patients and carers who experience diagnostics supported by AI, and clinical and managerial staff who work in or oversee care pathways affected by chest diagnostics.

“Second, due to the tight timeline of this phase of our evaluation, we were unable to interview several important groups, including patients and carers who experience diagnostics supported by AI, and clinical and managerial staff who work in or oversee care pathways affected by chest diagnostics.”