Physicians’ perceptions and attitudes on artificial intelligence in hospital management: a qualitative study Authors: Xizhen Zhou, Nazean Jomhari, Nik Daliana Nik Farid, Yiming Xing Published: 2025 ABSTRACT BackgroundArtificial intelligence has begun to find its way into hospital management, where it is used to ease tasks, such as allocating resources, coordinating workflows and handling data. For physicians, who directly experience effects of management systems, their perspectives are critical to whether AI tools can be meaningfully adopted and integrated. Yet research in this area has tended to remain on clinical side, with far less attention given to how physicians perceive and respond to AI in hospital management.ObjectiveThis study aims to examine how physicians in China perceive, experience, evaluate use of artificial intelligence in hospital management. It seeks identify potential application scenarios, highlight factors that facilitate or hinder integration, explore differences across physician groups.MethodsThe study used qualitative approach, with semi-structured interviews involving 25 practicing physicians in China. Data were analysed in NVivo with reflexive thematic analysis, carried out in six steps.ResultsThe study identified four main themes about physicians’ views on artificial intelligence in hospital management: potential of AI applications, factors that drive use, barriers to use, and integration strategies, along with 12 related subthemes.ConclusionThis study points out both opportunities and challenges in bringing artificial intelligence into hospital management. AI shows clear potential to improve efficiency and support institutional and digital transformation, but its use is limited by weak infrastructure, financial constraints and cultural resistance. Physicians’ attitudes also differ by role, hospital level, age, experience and views on risk. Because this study uses purposive sampling, a cross-sectional design and includes only physicians, the generalizability of the findings is limited.Supplementary InformationThe online version contains supplementary material available at 10.1186/s12913-025-13815-1. Introduction Computer systems and digital tools play an important role in improving healthcare quality, and also ensuring patient safety [2]. With expansion of information and communication technologies across different domains, digitalisation has become a key strategy to address challenges in health systems [16]. In this trend, artificial intelligence (AI) has gained increasing attention, as a fast-developing technology in healthcare, with applications across clinical practice, hospital management, and health policy. The opportunities brought by digital transformation in healthcare organisations cover a wide scope in hospital management [31]. AI is reshaping operating models and the daily work environment, of physicians through its roles in resource allocation and process optimisation [2]. This reshaping involves more than changes in workload and workflow, it embeds parts of resource allocation into algorithms and platforms, which alters physicians’ room for discretion and their influence in decision-making, shifts how responsibility for efficiency and fairness outcomes is understood. Although AI shows increasing potential in hospital management, existing studies still focus mainly on algorithm development, AI-based clinical decision systems, and views of healthcare staff [17, 52]. Work that examines attitudes towards AI for hospital management remains limited, and research on physicians—who are the direct recipients of management systems—is especially scarce. This gap makes the study of physicians’ attitudes towards AI in hospital management a critical missing piece. Without addressing how physicians accept and set boundaries for management-oriented AI, the promise of efficiency may turn, at the frontline, into new burdens of training, reconciliation, and system switching, and may also raise issues around responsibility and fairness. Physicians’ views and willingness to adopt AI are crucial for the implementation of AI-based hospital management. Greenhalgh et al. [19] found that the single most important factor in the success or failure of a new technology-enabled service at the local level may be whether the staff who use it accept it. If an AI management system does not fit with physicians’ values, priorities, and workflows, it easily leads to resistance or non-use. Physicians’ acceptance is therefore a basic condition for turning the potential of AI management into actual impact in practice [14]. For this reason, understanding physicians’ perceptions, concerns, and expectations is essential for identifying the factors that support or hinder AI adoption in hospital settings, and for offering practical guidance for user-centred design and implementation. This study addresses this gap by exploring physicians’ views on AI for hospital management. The main aim is to build on earlier work in other healthcare settings and examine Chinese physicians’ perceptions, experiences, and expectations regarding the use of AI in hospital management. On this basis, the study seeks to identify the key drivers, barriers, and governance considerations that shape their acceptance and use, so as to advance technology adoption research and hospital practice, and to offer a human-centred reference for the design and implementation of management tools. To achieve this aim, three research questions were proposed: RQ1: What expectations and concerns do physicians have about the use of AI in hospital management? RQ2: What governance and implementation arrangements can support physicians’ use of AI? RQ3: How do physicians’ perceptions of AI for hospital management differ from those in other medical contexts? In this study, “AI in Hospital Management” is defined as the use of technology to support daily hospital operations and management processes, including staff scheduling, resource allocation, supply management, document handling, compliance reminders, and prediction of patient flow and care pathways. Clinical applications such as diagnostic support, image interpretation, and treatment assistance are excluded to avoid overlap with existing literature on clinical AI adoption. Review Scholars described AI’s role in healthcare in different ways. Topol [49] outlines AI can affect three levels-clinicians, health systems, patients. Nasef et al. [30] suggest AI may improve treatment outcomes, simplify hospital operations. AI has become important force in healthcare, exerting far-reaching influence across three key domains: clinical practice, hospital management, health policy [28, 30]. In these areas, clinical decision support systems (AI-CDSS) can assist personalised diagnosis and treatment choices [6, 46]. It entered medical specialties such as radiology [42], dermatology [18], surgery [20]. These systems help improve clinical workflows, reduce medical errors [33, 42]. In health policy, AI starting to play a role in policy design and regulation. Predictive models can estimate policy outcomes, and to support the development of effective health policies [30]. In hospital management, AI shows value for efficient operations in many ways. Within this background, some empirical studies examined specific uses and outcomes of AI in hospital management. These include applications in resource allocation, workflow optimisation and data management [3, 39, 45]. In resource allocation, studies show that AI-driven predictive analytics can use historical and real-time data to forecast patient flow, and improve the use of staff, materials, facilities, and in the process identify high-risk groups, to support personalised patient education, simplify visits, which strengthens the ability of health systems respond to changing needs [8, 53]. Some studies also note that AI tools can support patient flow prediction, bed capacity management, improve resource use through data integration, reduce inefficiency [10, 29]. Surveys show that some hospitals already used predictive models and smart tools in operations and management, such as systems for forecasting inpatient trajectories, optimising appointments, these applications show potential to improve efficiency and resource allocation in practice [32]. Other studies focus on staff scheduling and workflow optimisation, the results show that AI improves scheduling, daily task management through process automation and intelligent scheduling, reduces administrative inefficiency, increases overall operational performance [30]. Related studies also show that intelligent scheduling can improve staff allocation, ease administrative workload, allow health professionals to spend more time on clinical work [21, 22, 24]. In addition, at the level of data integration and management, AI algorithms can process electronic health records (EHR) efficiently and combine them with information on bed use and human resources, to reduce information overload, improve the accuracy and timeliness of administrative decisions [32]. Some operational use cases such as intelligent documentation systems are also expanding. These include automatic generation of medical record summaries, clinical notes, discharge documents based on extracted diagnoses and test information, they show potential to improve efficiency and optimise resources [12, 26, 44]. At the same time, these systems still face challenges in data quality, system integration and evaluation, but the overall trend shows that AI is being introduced into hospital data management, also starting to play a role [3]. At the same time, some studies begun focus on an issue that runs in parallel with “hospital operational efficiency,” namely psychological burden and well-being of healthcare staff in medical settings [40]. This line of work offers evidence for understanding how digital technologies can support the well-being of healthcare workers, also reminds us that when discussing AI for hospital management, it is necessary to consider its impact on staff experience, stress and well-being [41]. Methods Study design This study is descriptive and uses semi-structured in-depth interviews, which is a qualitative research method [13]. The study follows the COREQ (Consolidated Criteria for Reporting Qualitative Research) guidelines to ensure completeness of reporting [48]. These guidelines are well recognised in the literature and are recommended for the design and reporting of qualitative research. Participant recruitment This study used purposive sampling with attention to diversity. The inclusion criteria were: (1) practising physicians aged 18 or above; (2) able to communicate in Chinese and give informed consent; (3) experience in hospital- or department-level management activities; (4) use of or exposure to hospital information systems such as HIS, EMR, LIS, or other digital management tools; (5) some understanding or views on the use of AI in hospital management. The exclusion criteria were physicians without a valid licence and those who had never taken part in any management activity. Recruitment took place through three routes: recommendations from hospital managers, circulation of the study invitation in department WeChat groups by department heads, and referrals by participants. Invitations were sent to 46 physicians, and 25 valid interviews were obtained (response rate 54.3%). Those who did not respond or declined often did so because of high workloads or scheduling difficulties. To ensure diversity, this study included physicians with different educational work experience and job titles. 25 physicians came from a broad range of specialties. They worked in tertiary, secondary, and primary hospitals in eastern China. This region was chosen because it has a full hospital hierarchy, relatively strong digital infrastructure, and good representativeness. All participants joined based on their willingness to share their experiences and views with the researchers. Participant characteristics are shown in Table 1. Table 1Descriptive characteristics of the physicianN = 25Categoryn (%)SexMale14(56%)Female11(44%)Age (years)Mean ± SD (min: 30; max: 59)38.52 ± 7.44Age< 39 years14(56%)≥ 39 years11(44%)Median age36EducationBachelor’s degree3(12%)Master’s degree19(76%)Doctoral degree3(12%)Hospital levelTertiary hospital17(68%)Secondary hospital5(20%)Primary hospital3(12%)DepartmentInternal medicine13(52%)Surgery7(28%)Other departments5(20%)Work experience (years)Mean ± SD (min:4;max:35)13.48 ± 7.73Work experience< 10 years10(40%)≥ 10 years15(60%)Median years of work experience14Professional titleAttending physician17(68%)Associate chief physician7(28%)Chief physician1(4%)Notes: In this study, the term “physician” refers to licensed medical doctors holding the Chinese Physician Practice Certificate and engaged in clinical practice. The term “physician” in this study is defined by licensure rather than educational degree Table 1Descriptive characteristics of the physicianN = 25Categoryn (%)SexMale14(56%)Female11(44%)Age (years)Mean ± SD (min: 30; max: 59)38.52 ± 7.44Age< 39 years14(56%)≥ 39 years11(44%)Median age36EducationBachelor’s degree3(12%)Master’s degree19(76%)Doctoral degree3(12%)Hospital levelTertiary hospital17(68%)Secondary hospital5(20%)Primary hospital3(12%)DepartmentInternal medicine13(52%)Surgery7(28%)Other departments5(20%)Work experience (years)Mean ± SD (min:4;max:35)13.48 ± 7.73Work experience< 10 years10(40%)≥ 10 years15(60%)Median years of work experience14Professional titleAttending physician17(68%)Associate chief physician7(28%)Chief physician1(4%)Notes: In this study, the term “physician” refers to licensed medical doctors holding the Chinese Physician Practice Certificate and engaged in clinical practice. The term “physician” in this study is defined by licensure rather than educational degree Descriptive characteristics of the physician Notes: In this study, the term “physician” refers to licensed medical doctors holding the Chinese Physician Practice Certificate and engaged in clinical practice. The term “physician” in this study is defined by licensure rather than educational degree Data collection This study used a semi-structured interview guide to support the exploratory research approach. The guide contained ten core questions. The interview questions were shaped step by step based on the study aims and insights from the literature. We first carried out an extensive review of studies that examined healthcare staff’s views, attitudes, and adoption of digital health tools or AI, and summarised the areas that physicians often care about when they take part in digital management changes. As the focus of this study was physicians’ experiences and views of AI in hospital management settings, the interview questions were then designed in line with this context. They centred on management-related use, such as views of AI, actual experiences, barriers and concerns, and preferred ways of implementation. These steps led to a set of ten questions designed for this study. To ensure the guide was valid, the development process included two rounds of refinement. The first step involved an expert review by two specialists: one scholar with experience in hospital work settings and one researcher in artificial intelligence, who examined the outline for clarity, coherence and contextual relevance. A pilot with three participants then helped test whether the questions could draw out views grounded in real management experience. Minor adjustments were made based on their feedback, and the introduction, follow-up prompts and closing section were added before the guide reached its final form (see Supplementary File 1). The interviews were conducted after ethics approval, between 4 April 2025 and 4 September 2025. Physicians who gave informed consent were contacted and scheduled at times convenient for them. 20 interviews were held online, and 5 interviews were held face to face in hospitals. Before each interview, participants were told about the study and informed that the interview would be recorded with their consent. Each interview lasted about 30–50 min. Data collection continued until data saturation. The recordings were transcribed verbatim. Online interviews gave busy doctors greater flexibility in time and location and raised the likelihood of participation. Their limitation was the reduced access to non-verbal cues, which may affect the interpretation of emotions and attitudes. In contrast, in-person interviews helped build trust and encouraged deeper conversation, yet they were more likely to be interrupted by clinical duties and may introduce a social expectation bias that prompts more positive views of AI. To reduce differences between the two modes, the same semi-structured guide was used in both, and probing, paraphrasing checks and post-interview reflexive notes were applied to maintain completeness and comparability. These steps helped strengthen the consistency and trustworthiness of the data. Data analysis Thematic analysis was used for qualitative data analysis. To protect participant privacy, physicians were coded as P1, P2, P3, and so on. Researchers recorded and transcribed the interviews. The six-step thematic analysis of Braun and Clarke [4] was applied, with themes taken directly from the data rather than set in advance. NVivo 15 was used to code the data, with a semantic and data-driven inductive orientation. Initial codes were assigned to relevant text segments in a systematic way, and participants’ quotes were placed under the related codes. The codes were then organised and compared with the support of NVivo functions, and related codes were gradually grouped into candidate themes and sub-themes. The research team revised and refined the codes during this process and merged them into broader sub-themes and overall themes. For any differences in coding labels or categorisation, the two coders went back to the surrounding context in the original interview quotes and reviewed the participant’s intended meaning. We reached agreement through joint discussion, and no disagreement remained unresolved. We used a “dual saturation” standard: code saturation and meaning saturation. The first fifteen participants produced most of the primary codes and shaped the four-theme structure. Participants sixteen to twenty-two added a small number of more detailed sub-codes, all of which fitted into existing themes and did not form new ones. Participants twenty-three to twenty-five did not bring new conceptual codes and mainly offered repeated or supporting material. Based on this, we judged that saturation was reached. Rigor and trustworthiness This study followed the four criteria of Lincoln [25]: credibility, dependability, confirmability, and transferability, and it reported the research process with reference to the COREQ checklist [48]. To keep consistency, one researcher conducted all interviews. This researcher had no hierarchical or collegial relationship with the physicians, which reduced power differences and role bias. Because the first author has a background in human–computer interaction and digital health, this may shape the way the data is understood and the direction of follow-up questions. In addition, the interviewer is familiar with the potential value of AI in hospital management, which brings the risk of unintentionally highlighting the advantages of the technology. To reduce the influence of the interviewer’s background on participants’ answers, the interviewer explained before each session that the study aimed to understand different views and real experiences, and that there were no “correct answers”. During the interviews, the interviewer avoided leading wording and encouraged participants to share doubts, worries or opposing opinions. During the data analysis stage, this study used Reflexive Thematic Analysis (RTA) [4, 5]. Two researchers carried out the initial coding independently, then compared differences and exchanged interpretations across several rounds of discussion until reached agreement. Themes grew through repeated interaction between the researchers and the data, rather than through simple categorisation. To improve transparency and trustworthiness, we kept reflexive memos throughout the process to record prior assumptions, key decisions and potential biases (for example, expectations about the efficiency of AI), and revisited these notes during the analysis to adjust our position. We also carried out negative case analysis and used it to refine the boundaries of the themes, so that each theme reflected the participants’ accounts rather than the researchers’ assumptions. This process continued across the whole analysis and helped strengthen the transparency, confirmability and credibility of the study. Ethical considerations This study followed the Declaration of Helsinki (World Med, 2013). It was approved by the University of Malaya Research Ethics Committee (Approval No.: UUM. TNC/UMREC_4442) and by the Ethics Committee of Liaocheng Second People’s Hospital in China (Approval No.: [2025] Medical Ethics Review (16)). To reduce privacy risks, recordings were stored and transcribed only in China. All identifiable information was removed before analysis, and only de-identified text was used for academic collaboration and research. All de-identified data were stored securely on password-protected devices with tightly controlled access. The research data will kept for five years, after that all digital files will be permanently and securely deleted. Before obtaining written and verbal consent, researchers introduced themselves, explained the study purpose, described the data collection methods. Participants who agreed join and have the interviews recorded signed written informed consent. All participants took part voluntarily and did not receive any form of financial compensation. Time and place of interviews were set according to participants’ preferences. Participants were told they had the right to withdraw before or during participation at any time. Data assurance was given to participants at each stage of the study. Participants were encouraged to contact the researchers if they had questions or concerns. Results This qualitative study included 25 participants, mean age is 38.52 ± 7.44 years. Participants were grouped into two age categories: younger than 39 years formed the majority (n = 14, 56%), 44% were 39 years or older. Most of the sample were male (n = 14, 56%). In terms of education, most participants held a master’s degree (n = 19, 76%). Regarding workplace distribution, 68% worked in tertiary hospitals (n = 17). For professional titles, attending physicians made up the largest group (n = 17, 68%). Mean work experience was 13.48 ± 7.73 years. To ensure diversity of results and reduce bias, the physicians came from a wide range of departments, including cardiology (n = 2), endocrinology (n = 1), respiratory medicine (n = 4), respiratory and critical care (n = 1), neurology (n = 1), general practice (n = 1), oncology (n = 2), paediatrics (n = 1), general surgery (n = 4), orthopaedics (n = 2), stomatology (n = 1), obstetrics and gynaecology (n = 2), radiology (n = 2), and health examination (n = 1) (Table 1). Thematic analysis of interviews identified four overarching themes and corresponding subthemes, which are summarised in Table 2. Figure 1 visually represents the interconnectedness of these themes. Table 2Themes and sub-themesThemesSub-themePotential and understanding of AI in hospital managementInformation and documentationRoutine operations and resource allocationForward-looking decision supportFactors driving AI adoptionWorkload reductionImproving operational efficiencyOptimizing resource utilizationDigital transformationBarriers to AI adoptionOrganizational and cultural constraintsTechnical and system limitationsAwareness and attitudesGovernance and implementation of AI in hospital managementEthics and accountabilityIntegration and implementation strategies Table 2Themes and sub-themesThemesSub-themePotential and understanding of AI in hospital managementInformation and documentationRoutine operations and resource allocationForward-looking decision supportFactors driving AI adoptionWorkload reductionImproving operational efficiencyOptimizing resource utilizationDigital transformationBarriers to AI adoptionOrganizational and cultural constraintsTechnical and system limitationsAwareness and attitudesGovernance and implementation of AI in hospital managementEthics and accountabilityIntegration and implementation strategies Themes and sub-themes Information and documentation Routine operations and resource allocation Forward-looking decision support Workload reduction Improving operational efficiency Optimizing resource utilization Digital transformation Organizational and cultural constraints Technical and system limitations Awareness and attitudes Ethics and accountability Integration and implementation strategies Fig. 1Four key themes of AI in hospital management Four key themes of AI in hospital management Theme 1: Potential and Understanding of AI in hospital management Subtheme 1a: Information and documentation Participants generally saw documentation as one of the areas most suited for the early use of AI, especially tasks such as medical-record summaries, clinical notes and discharge summaries, which have fixed formats and involve highly repetitive content. Physicians expected AI to generate a first draft based on clinical information, and then let them review and adjust it. They believed this could cut down mechanical typing, reduce writing errors and shorten the related workflow. One participant said:Writing up cases takes up a big part of our clinical work. If AI could turn a short consultation into a case record through speech recognition, it would cut a lot of the repeated work and take a real load off (P24, orthopaedics). Writing up cases takes up a big part of our clinical work. If AI could turn a short consultation into a case record through speech recognition, it would cut a lot of the repeated work and take a real load off (P24, orthopaedics). Most participants in this study tended to see AI as a “back-office tool” that handles standardised and structured information, while keeping professional authority in activities that carry higher levels of responsibility, such as clinical decision-making and communication with patients. Subtheme 1b: Daily operations and resource allocation AI was seen as useful for day-to-day resource scheduling and allocation. The interview analysis showed that physicians thought AI was better suited for management tasks that involve high complexity and frequent changes. One aspect mentioned by all scheduling administrators (N = 5) is potential use of AI in staff rostering. They expect AI cut high time cost from complex rules and last-minute changes, reduce personal preference factors, ease workload. They also see AI as a chance to make scheduling more transparent and fairer. Another commonly mentioned theme is the use of AI in triage and waiting-area flow. Physicians think AI may help optimize patient movement, cut waiting time and unnecessary back-and-forth so improve system efficiency. In high-consumption departments (such as surgery), many physicians said they feel the pressure from supply shortages and unclear information more directly. They believe AI through prediction and auto-replenishment can ensure timely drugs and materials, which helps efficiency and keeps continuity of care and patient safety:If AI helped planning, it might notice respiratory intervention group nearby needs laryngeal masks. When we don’t have enough, maybe it could arrange allocation, get someone to deliver masks to our place (P13, respiratory medicine). If AI helped planning, it might notice respiratory intervention group nearby needs laryngeal masks. When we don’t have enough, maybe it could arrange allocation, get someone to deliver masks to our place (P13, respiratory medicine). Subtheme 1c: Forward-looking decision support Interview data shows about half of participants did not only see AI as a tool, for routine tasks, also imagined it as an “early warning system”, that could support forward-looking management decisions. A few participants said if AI could give reminders for unfinished tests, or missed key steps, it might help lower the chance of errors:If AI could remind what tests haven’t been done, or point out diseases which need to think about, it would help us avoid missing things (P25, orthopaedics). If AI could remind what tests haven’t been done, or point out diseases which need to think about, it would help us avoid missing things (P25, orthopaedics). At the system levels, a small group of participants spoke about possible value of AI’s predictive ability. Include predictions of outpatient, inpatient volumes based on historical and real-time data, could help plan staff and beds in advance, to early signals for readmission or follow-up loss after discharge. One physician said if the system could match patients with physicians, by combining disease patterns and clinical expertise, it could improve operational efficiency, and the use of resources:For example, if we see more vascular cases or COPD patients on a day, the system may arrange doctors who are good at those conditions to take the clinic. That match patients and doctors more accurately (P15 respiratory and critical care). For example, if we see more vascular cases or COPD patients on a day, the system may arrange doctors who are good at those conditions to take the clinic. That match patients and doctors more accurately (P15 respiratory and critical care). Overall, compared with the earlier focus on less workload or fixing one process, this sub-theme points to another expectation: physicians care more about whether AI can bring better predictability. They hope AI helps on one small step, but can move those scattered decisions, once based on experience and moment-to-moment judgment into a more data-driven, overall planning. Theme 2: Factors driving AI adoption Subtheme 2a: Reduced personal workload Most physicians said that reducing non-clinical workload is a main reason they would accept AI management tools. Many noted that if AI can generate drafts of medical summaries or discharge notes from existing clinical information, it could cut a lot of time spent on documentation and let them focus more on clinical judgement and patient communication. Some also mentioned that AI could create follow-up plans and reminders, lowering the chance of missing tasks. A few physicians said they use AI to help with case writing or literature searches when the workload is heavy:For many patients, we need writing discharge diagnosis, going through the discharge plan takes a lot of time, we need to talk with them and make that judgement. If AI could generate first draft, I think it could save at least half time for each patient (P4, endocrinology). For many patients, we need writing discharge diagnosis, going through the discharge plan takes a lot of time, we need to talk with them and make that judgement. If AI could generate first draft, I think it could save at least half time for each patient (P4, endocrinology). Taken together, this subtheme shows that participants hoped to redistribute their professional time. They expected AI to take over standardised and low-value tasks that can be automated, so their workload could ease without reducing their clinical autonomy. Subtheme 2b: Operational efficiency and resource optimisation Many Physicians also viewed AI as a way to improve overall hospital efficiency, not only personal workload. Their comments often concerned complex scheduling, adjusting patient flow, responding to peak periods, and reducing repeated communication between departments. All physicians who manage schedules said that rule-based automated scheduling saves time, reduces human bias, and makes task allocation more transparent.What we really need is an AI that can handle sudden changes. Once something shifts, we set the conditions and it can sort out a new rota on its own, instead of me having to redo everything every time (P7, oncology). What we really need is an AI that can handle sudden changes. Once something shifts, we set the conditions and it can sort out a new rota on its own, instead of me having to redo everything every time (P7, oncology). A few senior physicians added that AI, if it predicts visit volumes or case types, may help hospitals prepare earlier for busy times and plan beds and staffing in advance. In triage and waiting areas, AI might reorder patients based on needs and planned tests, reducing unnecessary movement and long waits. Subtheme 2c: Improvements in supplies and processes In resource management, many participants saw AI as a main reason to try these tools. Physicians from high-consumption departments said that continuous monitoring and prediction of supply use could give early stock warnings and trigger automatic restocking, so they do not find shortages right before clinical use. Some also noted that better data integration could make storage, ward consumption, and real demand line up more closely, and in this way reduce overstock and waste.If AI can spot supply shortages ahead of time, automatically pull things from the warehouse, that would be great. If it could track how much we use, send alerts, or deliver items straight to the department, it save us a lot of time and effort (P14, respiratory medicine). If AI can spot supply shortages ahead of time, automatically pull things from the warehouse, that would be great. If it could track how much we use, send alerts, or deliver items straight to the department, it save us a lot of time and effort (P14, respiratory medicine). From these views, it is clear that some doctors are not satisfied with the older way that relies on manual stock checks, personal experience, and ad hoc coordination. They prefer AI to bring more predictable and traceable decisions. Their expectations of AI are no longer limited to a single step in the workflow but extend to more fine-grained control of the supply chain and process management. This search for more accuracy and more control is another reason why they support AI in hospital management. Subtheme 2d: Digital transformation Most participants also felt that introducing AI is a concrete way for hospitals to respond to “smart hospital” and digital transformation agendas. In their eyes, AI deployment is not only a technical upgrade but also a signal of organisational culture and future direction, and a response to national policy and broader development trends. Many physicians further linked AI with more standardised management, and thought its possible uses fit closely with the hospital’s long-term development plans.Honestly, we’re all for it. AI is a big direction for society now anyway (P24, stomatology). Honestly, we’re all for it. AI is a big direction for society now anyway (P24, stomatology). This sub-theme suggests that part of physicians’ positive attitude towards AI comes from its place in the story of organisational change. AI is seen as a tool that helps hospitals move from management based mainly on experience to management that relies more on data and structured rules. This positioning, at a more macro level, strengthens their favourable view of management-oriented AI tools. Theme 3: Barriers to AI adoption Subtheme 3a: Organisational and cultural barriers Interview findings show that the main barriers to AI come from the organisation itself, especially resources and culture. At the organisational level, all participants said hospitals invest limited money and staff into these systems. High costs for building and maintaining AI systems make it hard for many hospitals to really implement them, and even harder to keep updating and improving them. This not only means “lack of money” but also shows how much (or how little) hospital leaders truly commit to digital transformation over the long term.AI is just too expensive, it often costs hundreds of thousands, even millions. Our country already puts very little money into healthcare, our hospital doesn’t give any budget for this at all (P1, cardiology). AI is just too expensive, it often costs hundreds of thousands, even millions. Our country already puts very little money into healthcare, our hospital doesn’t give any budget for this at all (P1, cardiology). About one quarter of physicians stressed that the Chinese medical environment values “renqing” (personal warmth) and flexibility. If scheduling, bed allocation, or supply distribution all follow fixed rules and algorithms, these management tasks can become rigid. Then it is hard to take special personal situations into account, and the human side of care and trust may weaken. For some doctors, this is not just a style issue but a threat to professional autonomy. They worry that if AI goes too deep into management tasks, it will narrow their room to adjust and negotiate, and this may create resistance.AI isn’t very humane. Things like scheduling, arranging beds, sometimes you need to make exception, someone has family issues, someone isn’t feeling well, AI just can’t show that. Our hospital still values human warmth, but once switch to AI, everything may become rigid and mechanical, just following rules, no flexibility (P14, respiratory medicine). AI isn’t very humane. Things like scheduling, arranging beds, sometimes you need to make exception, someone has family issues, someone isn’t feeling well, AI just can’t show that. Our hospital still values human warmth, but once switch to AI, everything may become rigid and mechanical, just following rules, no flexibility (P14, respiratory medicine). Subtheme 3b: Technical and system constraints The interviews showed that limits at the technical and system level were among the most direct and widely felt obstacles. A few physicians questioned the reliability of AI systems, said there were still gaps and errors in real use, which made it hard for them to trust tools fully. Because physicians generally believed they still carry the final responsibility for decisions made with AI, so system error increased their sense of risk and led to the feeling “it is safer to do it myself.” One participant said even when all instructions were entered correctly, the system still produce results with bias or inaccuracy, which reduced trust, and limited the usefulness of AI in management:We tried using AI for scheduling. We put all rules in but schedule it gave us still had lots of problems. It just couldn’t be used as it was (P16, respiratory medicine). We tried using AI for scheduling. We put all rules in but schedule it gave us still had lots of problems. It just couldn’t be used as it was (P16, respiratory medicine). A smaller group of physicians focused on compatibility with existing hospital information systems (such as HIS and LIS). They pointed out that many systems do not connect well, so AI tools cannot easily fit into current workflows and may create information islands. This shows that integration across departments and systems is often slow and difficult. At the same time, hardware limits and difficulty in system upgrades, especially in secondary and primary hospitals are often seen as major technical obstacles for AI to take root in hospital management.Maybe we need better hardware and smarter devices. current setup with the system probably can’t run well. The whole hospital need upgrades, that’s not a small change (P23, stomatology). Maybe we need better hardware and smarter devices. current setup with the system probably can’t run well. The whole hospital need upgrades, that’s not a small change (P23, stomatology). Subtheme 3c: Awareness and attitudes The interviews showed negative views and attitudes formed internal barrier to spread of AI. Physicians differed on how they understood AI, some of them showed limited knowledge when talked it, they unsure about its value. Who used AI more often were generally more interested in AI-based management, but their acceptance still varied by role and department. Resistance to change, especially among older physicians, was seen as a important cultural and structural barrier. This resistance may come partly from their clinical experience, it could not be replaced by AI, and partly from concern about the effort needed to learn a new system.Senior physicians may harder to accept this stuff. Younger doctors usually pick it up more easily…maybe because they don’t have much experience yet. For us who’ve been around longer and seen more, we don’t feel we need it as much (P19, paediatrics). Senior physicians may harder to accept this stuff. Younger doctors usually pick it up more easily…maybe because they don’t have much experience yet. For us who’ve been around longer and seen more, we don’t feel we need it as much (P19, paediatrics). In addition, most physicians said they had not yet experienced clear, direct benefits from AI in daily management work or clinical practice. This lack of noticeable advantages reduced their motivation to use AI on their own and, led to a generally cautious attitude toward its role in hospital management. This finding suggests when there is no mandatory requirement, delayed or unclear perceived benefits may become a key factor that shapes physicians’ intention to adopt AI.Probably haven’t really tasted the “sweet side” of AI yet. It does help in some ways, but in day-to-day work I haven’t really felt its benefits very directly. Maybe after we use it more, get familiar with it, we’ll start to get used to it (P16, respiratory medicine). Probably haven’t really tasted the “sweet side” of AI yet. It does help in some ways, but in day-to-day work I haven’t really felt its benefits very directly. Maybe after we use it more, get familiar with it, we’ll start to get used to it (P16, respiratory medicine). Theme 4: Governance and implementation of AI in hospital management Subtheme 4a: Ethics and responsibility In the hospital management setting, physicians generally felt privacy protection and responsibility allocation were key governance issues, that would decide whether AI could become part of routine management. Some physicians said they removed names and other sensitive details when using tools, to avoid possible data-leak risks, a few believed hospital’s internal network is secure enough to support AI use. Which shows current explanations about data risks, compliance boundaries, protective measures were still not clear. As a result, frontline staff lacked a stable basis of trust in AI-based management systems. Unclear responsibility was seen as a deeper institutional barrier. Most physicians stressed when there were no clear laws or hospital rules, AI could only serve as reference tool, they would still carry the final responsibility. This formed a typical defensive pattern of use. Some participants also said boundary of responsibility between hospital managers and system vendors remained vague when errors occurred. This unclear chain of accountability stood out it as the highest-level ethical and responsibility challenge in AI-based management, far beyond a simple technical issue.If AI make mistakes, I guess company and doctors all have to take some responsibility. But it’s hard to say, because there’s no clear standard of how responsibility should be divided (P15, respiratory and critical care).This AI is just for reference, it’s not something with fixed rules. We still make our own decisions. So, if something goes wrong, I feel physician would take more responsibility. And that’s actually a pretty big issue (P19, paediatrics). If AI make mistakes, I guess company and doctors all have to take some responsibility. But it’s hard to say, because there’s no clear standard of how responsibility should be divided (P15, respiratory and critical care). This AI is just for reference, it’s not something with fixed rules. We still make our own decisions. So, if something goes wrong, I feel physician would take more responsibility. And that’s actually a pretty big issue (P19, paediatrics). Subtheme 4b: Integration and implementation strategies Interview results showed bring AI into hospital management in a meaningful way, hospitals need to make a strategic commitment at the organisational level. All participants said funding was starting point. They felt high cost of medical record systems, related equipment had to be financially workable, before system run in a stable way. This view highlighted key condition for introducing AI: hospital leaders must treat digitalisation as a long-term strategic capital investment, rather than a short-term IT expense.It’s quite rare to see two parts come together. But to sum up, it still comes down to cost. Hospital should put in money and training. Without these basic things, you can’t really make AI and hospital systems work well together (P17, general surgery). It’s quite rare to see two parts come together. But to sum up, it still comes down to cost. Hospital should put in money and training. Without these basic things, you can’t really make AI and hospital systems work well together (P17, general surgery). Attitude of leaders was widely seen as a key factor (mentioned by more than 80% of physicians). Participants generally felt support from hospital leaders would directly influence both introduction of AI, and physicians’ willingness to use it. This supported the top-down nature of AI implementation. Acceptance of senior management was viewed as main driver for overcoming internal resistance and creating a positive climate for use.You have to let the management; the leaders see AI is actually useful. If they feel that using AI makes work more structured and logical, cuts down the workload, saves time, then they’ll be more willing to bring AI into the hospital (P21, general surgery). You have to let the management; the leaders see AI is actually useful. If they feel that using AI makes work more structured and logical, cuts down the workload, saves time, then they’ll be more willing to bring AI into the hospital (P21, general surgery). In addition, some participants said that without a well-built digital platform and clear institutional procedures, even advanced technology would have limited impact. This showed that organisational infrastructure and policy support were basic conditions for the successful use of AI. The effect of any AI tool would, in the end, depend on how well the organisation prepared its systems and structures. Discussion Theoretical mapping and integrative model This study uses reflexive thematic analysis to build the coding structure in an inductive way. After completing open coding and theme development, we placed these findings in a side-by-side mapping with UTAUT and CFIR [9, 50]. It is important to note that these two theories serve as analytic tools for interpreting and organizing the results rather than preset templates for coding. This approach keeps the openness that allows themes to grow from the interview data and gives us a way to examine how well existing theories of technology adoption and implementation fit the context of AI in hospital management and where their limits may lie. Theme 1 and Theme 2 show that physicians tend to see AI as a tool for reducing workload, improving workflows, and managing resources. These views align with Performance Expectancy in UTAUT, as well as relative advantage and readiness for implementation in CFIR. Theme 3 reveals the other side of this “conditional acceptance.” Problems such as weak system integration, unstable outputs, learning or switching costs, cultural tension, concerns over professional boundaries, and doubts about reliability correspond to Effort Expectancy, perceived cost and risk in UTAUT, and to the Inner Setting and Characteristics of Individuals in CFIR. Theme 4 shifts to the institutional level. External digital strategies and industry benchmarks give momentum (CFIR: Outer Setting—external policies and incentives). At the same time, gaps in privacy protection, unclear responsibility, and missing standards and oversight (CFIR: Outer Setting & Process) create major constraints. These gaps weaken the positive role of Facilitating Conditions in UTAUT and lead frontline physicians to treat AI as a “for reference only” tool in a defensive way. Fig. 2Integrated multi-level model of AI adoption in hospital management Integrated multi-level model of AI adoption in hospital management Across the four themes, physicians generally recognize the potential of AI for management tasks, but their willingness to adopt it shows a clear threshold. When reliability is weak, when organizational conditions do not fit, or when governance safeguards are missing, a positive attitude does not turn into routine use. Based on this, the codes were reorganized into six categories (individual/organizational/governance and drivers/barriers). The four themes were then grouped into a three-level structure—individual, organizational, and macro/governance—to show that AI adoption does not move in a linear way. It is a threshold-based and conditional process: performance expectations and risk judgments at the individual level need support from organizational resources, system integration, and cultural adjustment. These, in turn, depend on clear rules and accountable data and responsibility governance at the macro level. Only when all three levels cross the adoption threshold together can perceived benefits shift from attitude-level agreement to stable embedded use in practice (Fig. 2). Section ”Conditional support: Drivers, barriers, and Tensions in Attitudes” builds on this model and explains, in layers, the internal mechanisms and tensions within this form of “conditional support.” Conditional support: drivers, barriers, and tensions in attitudes In line with research from other healthcare settings, this study shows that the first step in developing and implementing AI for hospital management is to identify the actual needs of stakeholders [43, 51]. The findings indicate physicians’ attitudes toward AI in hospital management are neither purely optimistic nor simply resistant. Instead, it shows a careful form, of support rests on a set of conditions. To describe the mixed attitudes more clearly, discussion below examines this form of “conditional support” across three levels: individual, organizational and governance. Individual level: efficiency gains and risk trade-offs The findings show physicians hold the most positive views toward the perceived value of AI; they see it as the strongest driver of adoption. Interview data (T1 and T2) indicate most participants view AI as a tool to lightens their workload, hope it can free them from repetitive tasks. This expectation reflects their pursuit of efficiency and their wish, to redirect time toward high-value work such as clinical judgment and research. This aligns closely with the “performance expectancy” dimension in the UTAUT framework: when physicians see AI as backend support that improves efficiency, work quality, their willingness to adopt it becomes stronger. However, this efficiency-driven positive attitude limited by physicians’ views of AI’s actual constraints and their own understanding of it. This is not a form of technology fear, but rational trade-off shaped by real use. Most physicians regard AI to gain efficiency, but the lack of steady, tangible benefits and the high cost of adapting workflows raise effort expectancy and lower the intention to adopt. At the same time, they remain cautious about errors in complex or non-standard situations, which reflects concerns about reliability at the individual level. Because of this, many physicians place AI within a model of human–AI collaboration so they can keep a controlled balance between gains and risks. In addition, physicians’ attitudes toward AI management tools do not unfold uniformly at the individual level; they shift in systematic ways with role identity and personal characteristics [2, 7, 23, 37]. This study finds that physicians who take part in scheduling, bed coordination, or supply management can feel the potential benefits of AI in reducing workload, so they are more open to adopting it. Those who do not take part in these tasks more often see AI as an optional add-on. This difference fits the view in UTAUT that performance expectancy and facilitating conditions shift with experience and the relevance of the use context. When a technology matches the tasks a person needs to do, the positive effect becomes easier to see; when it does not, the tool can feel like an extra burden. Table 3Distribution of codes across themes by years of clinical experienceTheme/Work Experience< 8 years(n = 9)8–17 years(n = 10)≥ 18 years(n = 6)TotalT11523947T232371887T321403293T411182251Total7911881278 Table 3Distribution of codes across themes by years of clinical experienceTheme/Work Experience< 8 years(n = 9)8–17 years(n = 10)≥ 18 years(n = 6)TotalT11523947T232371887T321403293T411182251Total7911881278 Distribution of codes across themes by years of clinical experience Age and career stage also form important lines of difference. Younger physicians who are familiar with digital tools show more tolerance toward AI, while senior physicians take a more cautious position. The grouped codes support this pattern. As shown in Table 3, senior physicians have a higher share of codes in T3 (barriers) and T4 (ethics/governance), which means that more experience often comes with more concerns and a higher adoption threshold. This fits well with UTAUT’s view that age and experience shape effort expectancy and risk perception. From a CFIR perspective, these differences show the interaction between “Characteristics of Individuals” and the “Inner Setting”: physicians do not make technology choices alone but judge whether an AI tool is trustworthy and worth bringing in under the culture and performance pressures of their workplace. Based on these findings, this study suggests that the design and rollout of AI management tools should not assume a single, uniform physician user group. Instead, design, communication, and responsibility arrangements need to shift with role and career stage to balance performance expectations and perceived risks, and in turn raise acceptance. Organisational level: structural constraints and cultural tension At the organisational level, physicians especially those with managerial duties tend to hold a more welcoming view of AI. They see AI as a tool that may help address long-standing structural problems in their daily work. Its expected value lies in using trackable rules and data to move resource allocation towards arrangements that are more standardised and fairer. This expectation aligns with “facilitating conditions” in UTAUT and with “organisational readiness” and “implementation climate” in CFIR. When AI is imagined as infrastructure that supports standardised management, the willingness to adopt it at the organisational level becomes noticeably stronger. However, the actual organisational conditions form major barriers to the use of AI. First, system development and long-term maintenance cost a lot, the existing systems lack compatibility and integration points, and some outputs are not stable, so staff need to check them repeatedly. Second, frontline physicians take part only to a limited degree in the design and evaluation stages, and this may cause misalignment with real work processes. Within the CFIR framework, these patterns show structural gaps in the inner setting, including available resources, technical fit, and the implementation process. When resource support, workflow integration and user involvement do not reach the required level, AI tools cannot form stable institutional use, and the expected efficiency gains cannot be reached. A further source of tension lies in the organisational culture itself. Some physicians hope that AI can reduce “relationship-based” and “personal favour” decisions and make rules clearer and easier to trace. Others feel that hospital work depends a lot on situational judgement and on giving flexible care in special cases. If an AI management tool is built as a system that only follows strict rules, it may narrow the space for professional judgement and reduce room for informal adjustment, which can trigger resistance among some doctors. Based on this, we argue that the key is not to choose between “relationships” and “rules”, but to write both into the formal arrangements. On the one hand, there should be unified and stable basic rules. On the other hand, it should also state in which situations exceptions are allowed, who can approve them, and that every exception needs to be logged and traceable. It can also give departments some decision boundaries so frontline staff can adjust within the rule framework according to real conditions. Taken together, the findings show that the organisational drivers and barriers in physicians’ views do not stand apart but come from the same underlying structure. AI is expected to address long-standing issues of efficiency and fairness, yet limited resources, weak system integration, and a culture that values flexibility introduce new uncertainties. Rather than seeing this simply as the presence or absence of “facilitating conditions” in UTAUT or support from the “inner setting” in CFIR, the more central question is whether the technical and institutional arrangements offered by the organisation can align with frontline practice and local norms. If not, AI may shift from a proposed solution to a new source of institutional friction. Macro and governance level: gaps in trust and unclear responsibility At the macro and governance level, most participants viewed AI as part of national policy directions and the wider digital transformation of the health sector. Introducing AI was understood as something hospitals ought to do to keep pace with broader trends. The driving force at this level came mainly from external policy signals rather than concrete institutional safeguards. Participants generally felt expectation to move forward, however, about how to apply AI in hospital management in a way of accountable and sustainable, they lacked clear guidance. In contrast, almost all physicians saw privacy and data security as basic conditions. Some felt the hospital intranet is relatively safe, but most did not really know how data move through the system or who is in charge of oversight. Some participants said they actively de-identify case information when using these tools, for example by removing names or other identifiers, to make up for their low trust in current data governance. A similar blur appears in how responsibility is divided. Most physicians stressed that AI is “for reference only” and that they, as clinicians, carry the final responsibility. At the same time, they also felt that hospital management and system vendors should share responsibility when there are technical errors. At present, there are no specific rules or clear responsibility-sharing frameworks for AI management tools. As a result, risks do not have a clear place in the formal system and are often pushed down to individual physicians, creating a kind of “responsibility vacuum” and concentrated pressure on frontline staff. From a CFIR perspective, this trust gap and responsibility vacuum show that external policies and the regulatory environment still do not give enough structural support for the deep integration of AI in hospital management. Under the UTAUT framework, even when performance expectancy and social influence are high, the lack of clear legal and institutional safeguards weakens the positive effect of “facilitating conditions” on adoption intentions. These findings suggest that the key task at the macro and governance level is not to reinforce technical visions, but to set up privacy and data rules, transparent processes for reviewing algorithms, and clear arrangements for defining responsibility among vendors, hospitals and physicians. Only when rules at this level are in place can the positive intentions at the individual and organisational levels turn into the sustainable use of AI management tools. Critical reflection: the paradox of optimism and the risks that remain under-discussed In this study, most physicians expressed a positive attitude. Although some physicians showed concern about parts of AI management, no one voiced complete opposition. This positive attitude should be understood as a form of conditional optimism and needs critical reflection. First, most hospitals have not yet deployed AI management tools in a formal way, so the views of participants rely more on expectations than on real and validated experience. Second, digital transformation and performance-oriented systems create institutional signals that may raise an expressive optimism toward technology. Last, the idea that automation is equal to neutrality and fairness hides the value choices behind indicator selection and resource redistribution. Because of this, the current optimism seems more like a forward-looking judgment shaped by the broader environment than a stable conclusion about real outcomes. At the same time, although participants mentioned responsibility boundaries and cultural tensions, they touched less on deeper negative consequences, which are key challenges for AI management tools in practice. For example, algorithmic bias and fairness risks may appear when AI uses historical data with built-in imbalances for training and evaluation. AI may then turn uneven patterns, such as night-shift load or promotion opportunities into seemingly objective rules, which conflicts with the goal of improving fairness [15]. When the line between advice and decisions is unclear, responsibility may shift away from its proper place, weaken professional autonomy, and increase regulatory pressure [47]. Privacy and the reuse of data also lack clear limits, and the balance among data minimization, purpose limitation, and traceable access has not yet formed [3]. These issues receive little attention in frontline discussions. In addition, efficiency alone does not turn into sustainable adoption, and the focus needs to shift to a human-centred view of integration. AI needs to fit into workflows in a way that allows adaptation, and shift the learning and coordination costs from physicians to the system (“AI-extended professional self”) [38]. When clear rules and safeguards are missing, the idea of efficiency often turns into training, switching, and reconciliation costs at the frontline [3, 30]. Dialogue with existing research: contextual differences and unique perspectives To examine the generalisability and limits of our findings and to clarify this study’s unique contribution, this section compares two adjacent contexts: nurses’ views of AI for management and physicians’ views of clinical AI-CDSS. We searched WOS and Scopus (2020–2025) for studies set in hospital environments, involving nurses or physicians, focusing on AI-enabled operational management or clinical decision-making, and reporting adoption drivers or barriers. We selected 10 representative studies from medical settings (5 from nursing; 5 from clinical AI-CDSS). We summarised the existing research into two clear lines of discussion, and the comparison matrix appears in Tables 4, 5 and 6. Table 4Nurse perspective: Management-AI (scheduling/operations)StudyN-S1 (Shift–WLB)N-S2 (NurseMgr-Impact)N-S3 (DE-Survey)N-S4 (Pref-Sched)N-S5 (RN-Fac/Bar)Role / participantsNursesNurse managersNursesNursesNursesContextScheduling / shift planningWorkforce and operationsMixed settingsScheduling / shift planningMixed settingsAI type / taskSchedule optimisation; transparent rulesResource allocation; workflow redesignPerceptions, intention, training needsPreference embedding; appeal optionsAdoption facilitators / barriersMethodQualitativeQualitativeQuantitativeQualitativeFocus groupsImplementation stagePrototype / pilotEarly adoptionReadiness assessmentCo-designPre-implementationPerformance expectancy●○△○●Organisational resources△●△△●Explainability / trust○△△●△Fairness●△–●△Key insightsBetter WLB and higher acceptance; participatory rule-setting is centralLeadership support and change management act as prerequisitesTraining and data literacy are thresholds; trust remains unstable“Preferences + coverage rules” improve fairness and uptakeRequires a combined strategy of “training + governance + phased rollout”Notes: ●= emphasised;○= present༛△= partial༛–= not evident Table 4Nurse perspective: Management-AI (scheduling/operations)StudyN-S1 (Shift–WLB)N-S2 (NurseMgr-Impact)N-S3 (DE-Survey)N-S4 (Pref-Sched)N-S5 (RN-Fac/Bar)Role / participantsNursesNurse managersNursesNursesNursesContextScheduling / shift planningWorkforce and operationsMixed settingsScheduling / shift planningMixed settingsAI type / taskSchedule optimisation; transparent rulesResource allocation; workflow redesignPerceptions, intention, training needsPreference embedding; appeal optionsAdoption facilitators / barriersMethodQualitativeQualitativeQuantitativeQualitativeFocus groupsImplementation stagePrototype / pilotEarly adoptionReadiness assessmentCo-designPre-implementationPerformance expectancy●○△○●Organisational resources△●△△●Explainability / trust○△△●△Fairness●△–●△Key insightsBetter WLB and higher acceptance; participatory rule-setting is centralLeadership support and change management act as prerequisitesTraining and data literacy are thresholds; trust remains unstable“Preferences + coverage rules” improve fairness and uptakeRequires a combined strategy of “training + governance + phased rollout”Notes: ●= emphasised;○= present༛△= partial༛–= not evident Nurse perspective: Management-AI (scheduling/operations) Notes: ●= emphasised;○= present༛△= partial༛–= not evident Table 5Physicians’ perspectives on clinical AI‑CDSSStudyD-C1 (HCP-Expect)D-C2 (UAE-Phys)D-C3 (UK-Docs)D-C4 (CN-fsQCA)D-C5 (AI-CDSS-Ethics)Role / participantsMixed HCPsPhysiciansPhysiciansPhysiciansMixed HCPsContextGeneral clinicalOutpatient/EmergencyUK NHSChina multi‑hospitalGeneralAI type / taskDiagnosis/prediction/recommendationClinical AI applicationsAI‑CDSS deployment experienceAI‑CDSS adoption intentionResource‑allocation decisionsMethodQualitativeQuantitative/modellingQualitativeQuantitative/fsQCAQualitativeImplementation stageEarly/conceptualPre‑deploymentPilot/early stagePre‑deploymentPre‑pilotPerformance expectancy●●○●△Organisational resources△△△△△Explainability / trust–––––Fairness△△●△●Key insightsNeed workflow alignment & clear responsibilityOrg support & training as key moderatorsTrust shaped by explainability & accountabilityPerformance × social influence × resourcesNeed transparent rules & retained human oversight Table 5Physicians’ perspectives on clinical AI‑CDSSStudyD-C1 (HCP-Expect)D-C2 (UAE-Phys)D-C3 (UK-Docs)D-C4 (CN-fsQCA)D-C5 (AI-CDSS-Ethics)Role / participantsMixed HCPsPhysiciansPhysiciansPhysiciansMixed HCPsContextGeneral clinicalOutpatient/EmergencyUK NHSChina multi‑hospitalGeneralAI type / taskDiagnosis/prediction/recommendationClinical AI applicationsAI‑CDSS deployment experienceAI‑CDSS adoption intentionResource‑allocation decisionsMethodQualitativeQuantitative/modellingQualitativeQuantitative/fsQCAQualitativeImplementation stageEarly/conceptualPre‑deploymentPilot/early stagePre‑deploymentPre‑pilotPerformance expectancy●●○●△Organisational resources△△△△△Explainability / trust–––––Fairness△△●△●Key insightsNeed workflow alignment & clear responsibilityOrg support & training as key moderatorsTrust shaped by explainability & accountabilityPerformance × social influence × resourcesNeed transparent rules & retained human oversight Physicians’ perspectives on clinical AI‑CDSS Table 6Comparison with existing researchDimensionNurse–Management AI (Table A Findings)Physician–AI-CDSS (Table B Findings)This Study: Physicians–Management AIOur ContributionResearch gapMultiple scheduling/fairness cases already exist; pathways relatively clearFocus on diagnosis; rarely addresses management decisionsPhysicians’ needs and boundaries in non-clinical management AIFirst systematic depiction of the adoption logic of physicians using management-AIObjective functionFairness × work-life balance × retentionAccuracy × timeliness × consistencyFairness × efficiency × safety × professional boundariesIntegrates fairness and efficiency within the same decision space and appeal processDecision mechanismParticipatory rules, explainability, override optionsLimited explainability/accountability; override rareIntroduces override/appeal and dual-layer explanationsTransfers nurse-side design concepts into the physician-management settingOrganizational conditionsLeadership support, continuous trainingLimited resources/workflow integrationTime protection + workflow redesign + clear role assignmentProvides an actionable checklist and milestones for embedding AIEthics and justiceStrong focus on procedural/distributive justiceEthical allocation; patient autonomy/accountabilityAuditable opportunity allocation and bias monitoringProposes audit and bias-monitoring indicators for management-AI Table 6Comparison with existing researchDimensionNurse–Management AI (Table A Findings)Physician–AI-CDSS (Table B Findings)This Study: Physicians–Management AIOur ContributionResearch gapMultiple scheduling/fairness cases already exist; pathways relatively clearFocus on diagnosis; rarely addresses management decisionsPhysicians’ needs and boundaries in non-clinical management AIFirst systematic depiction of the adoption logic of physicians using management-AIObjective functionFairness × work-life balance × retentionAccuracy × timeliness × consistencyFairness × efficiency × safety × professional boundariesIntegrates fairness and efficiency within the same decision space and appeal processDecision mechanismParticipatory rules, explainability, override optionsLimited explainability/accountability; override rareIntroduces override/appeal and dual-layer explanationsTransfers nurse-side design concepts into the physician-management settingOrganizational conditionsLeadership support, continuous trainingLimited resources/workflow integrationTime protection + workflow redesign + clear role assignmentProvides an actionable checklist and milestones for embedding AIEthics and justiceStrong focus on procedural/distributive justiceEthical allocation; patient autonomy/accountabilityAuditable opportunity allocation and bias monitoringProposes audit and bias-monitoring indicators for management-AI Comparison with existing research The two trajectories: Core focuses and boundaries In the nurse-focused literature, studies center on how to place fairness, transparency, and work–life balance into the explicit rules of intelligent scheduling and unit operations, and how to let frontline staff and managers explain and discuss these rules together. Several qualitative studies focus on writing fairness/ transparency/WLB into scheduling algorithms (preferences and constraints, open weight settings, auditable reasons) to reduce the sense of a “black box,” and to balance personal preferences, organizational needs, and patient load within a single view [35]. On the governance side, the literature suggests putting rule-based safeguards and role division in place at the start [17, 34]. The main frontline barriers are limited training, time pressure, increased workload, and low trust. The related strategies include system training, leadership support, clear governance, and position planning [36]. In the physician-focused trajectory, studies shift the focus to how AI-CDSS supports diagnosis, triage, and treatment decisions, and how it enters actual use. The overall pattern shows multiple paths to adoption (technology-driven, individual-driven, or both), but common concerns about professional autonomy, privacy, data security, maturity of the technology. Physicians show low levels of awareness and engagement, they need training support, flexible time to use these systems [14, 52]. Work in critical care and emergency settings show unintended effects (extra EHR work, alert fatigue) [51], so teams need redesign roles and workflows, to reduce these impacts. On the ethical side, studies point out tension between efficiency and fairness, questions around explainability, shifts in professional roles and data ethics issues. They call for frontline-oriented ethical guidance, and literacy frameworks [11]. Contextual differences: normative demands vs. performance-driven demands The two trajectories show different priorities. Nurses place more weight on normative demands, and ideas such as procedural justice, preference expression and respect, and work–life balance/retention become built-in system goals and evaluation criteria. Physicians place more weight on performance-driven demands, such as accuracy–timeliness–consistency in diagnosis, triage, and treatment, and clear responsibility and reviewability [35]. In addition, the physician trajectory often raises institutional conditions such as “learning pathways, time protection, and professional flexibility,” while the nurse trajectory highlights organizational conditions such as “governance and compliance frameworks, and managerial support” [14, 17, 52]. Shared mechanisms: a four-part cycle of trust, explainability, workflow, and governance Our comparative analysis shows that whether it is management-AI or clinical-AI, the move from being seen as promising to being used in practice depends on a shared set of conditions. Trust and role boundaries: there need to be workable rules for when to follow and when to override, and a clear line between professional judgment and algorithmic suggestions.Explainability: users need to trace why the system gives a certain result, which includes auditability of scheduling reasons, weight settings, and clinical recommendation bases.Workflow integration: adopting AI must come with redesigned time allocation, resource coordination, and cross-role collaboration; otherwise the expected efficiency gains will be offset by learning and switching costs.Organizational governance and ethical safeguards: job definitions, accountability, and compliance need to form one coherent system that can support long-term, stable use. Trust and role boundaries: there need to be workable rules for when to follow and when to override, and a clear line between professional judgment and algorithmic suggestions. Trust and role boundaries: there need to be workable rules for when to follow and when to override, and a clear line between professional judgment and algorithmic suggestions. Explainability: users need to trace why the system gives a certain result, which includes auditability of scheduling reasons, weight settings, and clinical recommendation bases. Explainability: users need to trace why the system gives a certain result, which includes auditability of scheduling reasons, weight settings, and clinical recommendation bases. Workflow integration: adopting AI must come with redesigned time allocation, resource coordination, and cross-role collaboration; otherwise the expected efficiency gains will be offset by learning and switching costs. Workflow integration: adopting AI must come with redesigned time allocation, resource coordination, and cross-role collaboration; otherwise the expected efficiency gains will be offset by learning and switching costs. Organizational governance and ethical safeguards: job definitions, accountability, and compliance need to form one coherent system that can support long-term, stable use. Organizational governance and ethical safeguards: job definitions, accountability, and compliance need to form one coherent system that can support long-term, stable use. These mechanisms appear repeatedly in frontline nursing studies and in physician studies across different countries, which suggests they are not context-specific coincidences but cross-setting requirements [1, 14, 27, 34, 36, 52]. The four mechanisms above form the basic conditions that cut across different settings; on this basis, this study further sets out three additional, context-specific insights in the doctor–management-AI setting (see below). Unique perspective and contribution of this study Unlike the two mature lines of research above, this study focuses on the intersection between physicians and management-oriented AI (such as scheduling and resource allocation). The key question is: when AI does not give clinical advice but enters physicians’ daily work through “operational decisions,” how do physicians understand, accept, and negotiate with it? Existing work either stresses fairness and empowerment in nurse–management studies, or stresses performance and responsibility in physician–clinical studies. What is missing is attention to the situation in which physicians become the subjects affected by management-AI while having little presence in rule co-creation. Therefore, we move UTAUT and CFIR from “clinical adoption” to “operational touchpoints,” and offer three insights: Fairness threshold: in management-AI settings, the path from organizational justice to adoption intention is stronger than the path from performance or effort expectations.Explanation and participation: case-level explanation, rule-level explanation, and participatory co-design form a combined mechanism that supports stable use.Institutional grounding: role-specific institutional conditions and cross-department coordination are needed to turn pilot projects into routine practice. Fairness threshold: in management-AI settings, the path from organizational justice to adoption intention is stronger than the path from performance or effort expectations. Fairness threshold: in management-AI settings, the path from organizational justice to adoption intention is stronger than the path from performance or effort expectations. Explanation and participation: case-level explanation, rule-level explanation, and participatory co-design form a combined mechanism that supports stable use. Explanation and participation: case-level explanation, rule-level explanation, and participatory co-design form a combined mechanism that supports stable use. Institutional grounding: role-specific institutional conditions and cross-department coordination are needed to turn pilot projects into routine practice. Institutional grounding: role-specific institutional conditions and cross-department coordination are needed to turn pilot projects into routine practice. Implications for theory and practice We place management-oriented AI (mostly in nursing) and clinical AI (aimed at physicians’ diagnostic work) under the same set of conditions (trust, explainability, workflow embedding, and governance safeguards) and compare them, and we find that the real gap lies in the “physician–management-AI” space. In other words, we need to bring fairness and empowerment concerns from the nursing side into the physician group, and at the same time bring efficiency and responsibility concerns from the physician side into hospital operational and governance rules. In line with recent systematic reviews that state that most AI tools still stay at the development or pilot stage and lack evidence on adoption and routinisation, we place this study as adding cross-setting, cross-role evidence at the institutional and organisational level, and we argue that, for AI to move from promising to used in everyday work, all four conditions must be met at the same time rather than only talking about efficiency or technical feasibility. Limitations The findings of this study have the following limitations. Sample and context: Although we tried to achieve maximum variation, participants still mainly came from eastern China, tertiary hospitals, and attending physicians with a master’s degree. This structure partly reflects the reality of institutions with higher levels of digital investment, but it does not adequately represent primary or resource-constrained settings. At the same time, China’s specific arrangements for hospital management and data governance limit the transferability of our findings across countries and health systems. Nature of data: The study focuses on early adoption context. Most participants had not yet built long-term experience with AI, so their feedback came more from perceived potential and expectations than from verified, routine use. This expressive optimism may underestimate the actual frictions in workflow and governance that appear during real-world implementation. Data collection and representation: Topics about responsibility, performance, and fairness may make participants hold back. Online interviews improved access but reduced non-verbal cues; in-person interviews faced interruptions from clinical work and were shaped by the local setting. Chinese transcription and necessary English translation may also cause small shifts in meaning. We used a shared interview guide, probing and paraphrasing checks, and reflexive memos to reduce bias, but we could not remove it completely. Theory and model: The four themes come from a bottom-up RTA process; their links to UTAUT and CFIR come from an interpretive step after the analysis, not from a prior deductive test, so the study does not give strong evidence for either framework. The integrated conceptual model in this paper is a hypothetical schema, and future studies need to test whether its proposed paths show statistical significance. Directions for future research Given that “management-AI” in Chinese hospitals is still at an early stage, more systematic work is needed to explain how frontline physicians form their views, how they actually adopt these tools, and what kinds of support they need. Based on this need, the study proposes several practical directions. Conduct small-scale quantitative studies: researchers can distribute short questionnaires across hospital levels, roles, and seniority groups. Descriptive statistics can show how physicians rate efficiency gains, learning costs, risk and responsibility, and trust and controllability. Stratified comparisons, by age, years of practice, clinical vs. administrative roles and hospital level can help identify which groups value or worry about which aspects. These steps can turn qualitative findings to measurable group-level patterns, support more tailored design and resource allocation in different settings. Longitudinal adoption studies: this study mainly captures physicians’ perceptions and expectations at early stage of implementation. Future work can use longitudinal designs to track how physicians’ attitudes shift before, during and after the actual rollout of AI management tools. These designs can show how perceived benefits compare with experienced benefits over time. Contextual comparison and design research: future studies can compare different countries or health systems, to see how culture and governance shape barriers and needs. Researchers also can turn to intervention or co-design studies, explore governance mechanisms that help solve problems, such as unclear responsibility and limited frontline participation. Conclusion This study shows physicians in China give conditional support for AI in hospital management. They agree AI has value for reducing workload, improving workflows, supporting resource allocation. But adoption does not grow in a linear way, it depends on whether physicians feel the benefits are clear, and controllable enough to cross an acceptance threshold. Their concerns center on system stability, learning costs, professional boundaries and unclear responsibility. To cross this threshold, the technical layer needs strong interoperability with existing systems and clear explanations. The organizational layer needs leadership commitment, real investment, co-designed workflow changes with frontline staff, role-specific training and fallback options. The governance layer needs clear responsibility, data minimization, routine model review and use audits. Compared findings from clinical AI-CDSS and nursing management studies, our results show that physicians in management-AI settings pay closer attention to fairness, transparency and accountability. In theory, this study proposes and integrates the concept of conditional support. We map the empirical themes onto the three layers of UTAUT and CFIR, and show how age, seniority, and job role shape these pathways. The model explains how individual, organizational, and governance conditions together determine the adoption threshold. In practice, the study argues that AI tools in hospital management should use a co-pilot rather than rigid automation approach. We outline three sets of checkpoints across the technical, organizational, and governance layers. The study suggests that only when trustworthy technology, supportive organizations, and clear governance act together can physicians’ general approval of AI turn into stable and routine use in hospital management. Implications for hospital management and policy To turn conditional support into stable adoption, Table 7 uses a three main tracks and two-phase rollout framework. It lists key actions, required resources, responsible parties, and KPIs by organizational level or role. This framework outlines the main pathways for introducing AI into hospital management. Table 7AI initiatives by levelLevel / ActorShort-term (0–6 months)Mid-term (6–24 months)Cost & ResourcesResponsible UnitsSuccess IndicatorsMedical Education & TrainingDevelop AI/digital health curriculum; run faculty/resident AI literacy workshopsIntegrate courses into CME credits; launch online training platformCurriculum budget; instructor hours; case library and venueMedical Education Dept; Teaching Office; IT DeptCourse completion ≥ 90%; Faculty training ≥ 70%Run 1–2 ‘shadow decision-making’ drills with real dataMake A/B scenario drills routineData de-identification workflow; case maintenanceMedical Education Dept; IT DeptDrill completion ≥ 85%; Avg. task time ↓20%Launch modules on AI safety and bias (explainability, rollback, logs)Include safety/ethics tests in annual CME creditsExam systemMedical Affairs; Ethics Committee; Medical Education DeptSafety test pass rate ≥ 95%Organization & LeadershipSet up AI project group; publish 6‑month roadmapEstablish AI R&D and application office; build use‑case library and shadow‑evaluation workflowProject manager; FTE/month budgetAdministration Office; IT DeptProject‑to‑evaluation cycle ≤ 8 weeksLaunch 1 MVP each for scheduling/follow‑upExpand to ≥ 3 departments with cross‑unit coordinationData engineering supportIT Dept; Pilot DepartmentsScheduling conflict ↓30%; Follow‑up delay ↓25%Set model registry, versioning, drift monitoring and rollbackImplement MLOps + full audit logDev/Ops tools and maintenanceQuality Control; Audit OfficeDrift resolution ≤ 72 h; Rollback success 100%Policy & GovernanceIssue AI ethics & data governance guidelines; clarify responsibility frameworkForm AI Governance CommitteePolicy drafting budget; templates and formsMedical Affairs; Info Security; Ethics CommitteeCompliance audit 100%; Data breaches = 0Set data‑sharing agreements and interface standardsAlign cross‑hospital collaboration and industry standardsIntegration & interface testing costIT Dept; Legal; VendorsInterface first‑pass success ≥ 95%; Documentation 100%Provide adoption incentives (time/funding) for frontline optimization workRun annual effectiveness review; adjust incentivesPerformance budgetAdministration Office; HR DeptParticipation ≥ 70%; Time saved ≥ 15% Table 7AI initiatives by levelLevel / ActorShort-term (0–6 months)Mid-term (6–24 months)Cost & ResourcesResponsible UnitsSuccess IndicatorsMedical Education & TrainingDevelop AI/digital health curriculum; run faculty/resident AI literacy workshopsIntegrate courses into CME credits; launch online training platformCurriculum budget; instructor hours; case library and venueMedical Education Dept; Teaching Office; IT DeptCourse completion ≥ 90%; Faculty training ≥ 70%Run 1–2 ‘shadow decision-making’ drills with real dataMake A/B scenario drills routineData de-identification workflow; case maintenanceMedical Education Dept; IT DeptDrill completion ≥ 85%; Avg. task time ↓20%Launch modules on AI safety and bias (explainability, rollback, logs)Include safety/ethics tests in annual CME creditsExam systemMedical Affairs; Ethics Committee; Medical Education DeptSafety test pass rate ≥ 95%Organization & LeadershipSet up AI project group; publish 6‑month roadmapEstablish AI R&D and application office; build use‑case library and shadow‑evaluation workflowProject manager; FTE/month budgetAdministration Office; IT DeptProject‑to‑evaluation cycle ≤ 8 weeksLaunch 1 MVP each for scheduling/follow‑upExpand to ≥ 3 departments with cross‑unit coordinationData engineering supportIT Dept; Pilot DepartmentsScheduling conflict ↓30%; Follow‑up delay ↓25%Set model registry, versioning, drift monitoring and rollbackImplement MLOps + full audit logDev/Ops tools and maintenanceQuality Control; Audit OfficeDrift resolution ≤ 72 h; Rollback success 100%Policy & GovernanceIssue AI ethics & data governance guidelines; clarify responsibility frameworkForm AI Governance CommitteePolicy drafting budget; templates and formsMedical Affairs; Info Security; Ethics CommitteeCompliance audit 100%; Data breaches = 0Set data‑sharing agreements and interface standardsAlign cross‑hospital collaboration and industry standardsIntegration & interface testing costIT Dept; Legal; VendorsInterface first‑pass success ≥ 95%; Documentation 100%Provide adoption incentives (time/funding) for frontline optimization workRun annual effectiveness review; adjust incentivesPerformance budgetAdministration Office; HR DeptParticipation ≥ 70%; Time saved ≥ 15% AI initiatives by level The integration of AI in hospital management should move forward step by step, starting with small-scale pilots to test its value, then expanding gradually to the whole hospital, and finally building a model for cross-institution adoption. In this process, the governance framework, digital foundation, and user experience should be treated as three parallel lines under continuous review and improvement, so that AI tools can secure sustained trust and acceptance from physicians in terms of reliability and usability. Supplementary Information Below is the link to the electronic supplementary material. Supplementary Material 1 Supplementary Material 1