The primary objective of this systematic review is to understand the experiences of nurses using AI in clinical practice.

A systematic review methodology was selected because the research question required a comprehensive, transparent, and reproducible synthesis of empirical evidence from multiple studies. This approach is appropriate for identifying and integrating findings across a body of literature to address a focused question about a specific population’s experiences, and is consistent with established guidance for evidence synthesis in health professions research [18].

The study protocol was registered with PROSPERO (CRD42022308051) [19]. The manuscript title has been refined from the registered title for clarity; the review question, eligibility criteria, and methods are consistent with the registered protocol.

Ethics approval was not required because this study synthesizes data from previously published literature and did not involve collection of identifiable participant data.

Eligibility was determined using predefined inclusion and exclusion criteria (Textbox 1).

A comprehensive literature search was conducted on September 13, 2023 across 5 electronic databases—PubMed, Embase, MEDLINE, CINAHL, and PsycInfo.

A search strategy was developed and refined through repeated testing and adjustment, using both keywords and controlled vocabulary tailored to each database. Concept blocks, illustrated in Textbox 2, provided the structure for the search, with keywords and controlled vocabulary terms merged using Boolean logic (AND/OR) to ensure comprehensiveness. No date restrictions were applied, allowing for the inclusion of all eligible studies up to the date of the search. Searches were not restricted by language, but inclusion was limited to studies published in English. Reference lists of identified articles were also reviewed, but no additional studies met the inclusion criteria. Full search strategies are provided in Multimedia Appendix 1.

The review was conducted following the Joanna Briggs Institute (JBI) methodology for systematic reviews, using the “JBI Manual for Evidence Synthesis” [20] as the guiding framework. The reporting of this review adheres to PRISMA (Preferred Reporting Items for Systematic Reviews and Meta-Analyses) 2020 guidelines [21]. Study management and screening were managed with Covidence software (Veritas Health Innovation Ltd). Furthermore, 2 reviewers (AJSS and QZ) independently screened titles, abstracts, and full texts, resolving discrepancies through discussion and consensus. In cases where multiple reports of the same study were identified, the most comprehensive version was prioritized, and duplicates were merged accordingly. The screening and selection process are presented in the PRISMA flow diagram in the Results section. The completed PRISMA 2020 checklist is also provided in Checklist 1.

Data extraction followed the JBI mixed methods data extraction form following a convergent integrated approach [20]. Where studies included multidisciplinary samples, nurse-specific findings were extracted separately where the primary data permitted. In cases where profession-specific disaggregation was not possible, this was noted explicitly in the data extraction table (Multimedia Appendix 2).

Study quality was assessed using the Mixed Methods Appraisal Tool (MMAT) [22] to accommodate the diversity of included study designs. This aligned with the convergent integrated approach used in this review [20]. Additionally, 2 reviewers (AJSS and QZ) independently assessed study quality, and any discrepancies were resolved through discussion to reach consensus. To capture the complexity of interpretive and critical understandings of AI use in nursing, an inclusive approach to study selection was maintained. As noted by Hannes and Lockwood [23], qualitative evidence can simultaneously illuminate both broad patterns and context-specific detail, and excluding studies entirely based on quality grounds risks narrowing this breadth. Including all studies regardless of MMAT scores allowed for a representative synthesis across diverse clinical contexts. Quality appraisal results were therefore used to inform the interpretation of findings, rather than as a basis for exclusion. Where lower-quality studies contributed to a construct, this is noted within the relevant Results section.

Thematic analysis was conducted using a predefined deductive framework based on TAM2, supplemented by the single UTAUT construct of Facilitating Conditions. All included studies were read in full, and relevant text segments, including direct participant quotations, authors’ qualitative summaries, and quantitative findings, were coded line-by-line into the framework constructs. Coding was undertaken by the first author (AJSS), with regular review and discussion of coding decisions with 2 supervising coauthors (BCB and DD) throughout the analysis period. These supervisory discussions served as a calibration mechanism for the application of the codebook; ambiguous extracts and decisions about construct boundaries were brought to these meetings, discussed against operational definitions, and resolved by consensus before being applied across the dataset. The codebook (Table 1) was iteratively refined through these discussions to clarify operational boundaries; construct categories themselves remained fixed throughout. Formal interrater agreement was not calculated, as only a single author (AJSS) coded the data; this is acknowledged as a limitation. Quantitative findings relevant to TAM2 constructs, such as survey results, were summarized and integrated narratively with the qualitative data under relevant headings. Where frequency of reporting is described, this refers to the number of included studies in which relevant data were coded to a given construct. Frequency counts indicate how commonly a construct appeared across the included studies and should not be interpreted as a measure of the depth or strength of any individual theme.

The PRISMA flow diagram (Figure 1) is presented to illustrate the screening and selection process. The initial searches yielded 952 records. After 424 duplicate or merged records were removed in Covidence, 528 articles were screened by title and abstract. Of these, 60 reports were sought for retrieval and assessed for eligibility, resulting in the exclusion of 40 reports for reasons such as the lack of AI experience (n=15), lack of experiential data (n=11), a focus on non-AI technology (n=6), and other contextual or methodological reasons as detailed in Multimedia Appendix 3. Ultimately, 20 studies [24-43] met the inclusion criteria and were included in the final synthesis.

Of the 20 included studies, 13 were appraised as high quality [31-43], 5 as medium [24-28], and 2 as low [29,30] (Multimedia Appendix 4). The 5 medium-rated studies were predominantly quantitative surveys with limitations relating to sampling representativeness and nonresponse bias [24-28]. The 2 low-rated studies had insufficient methodological reporting to fully assess rigor [29,30]. Job Relevance and Output Quality were supported exclusively or predominantly by high-rated studies. Perceived Usefulness was well-represented across quality levels but anchored by a majority of high-rated studies. Facilitating Conditions drew contributions from the broadest quality range, including both low-rated studies; findings for this construct are therefore interpreted with appropriate caution below.

Of the 20 included studies, the proportion of nurse participants varied substantially. Of the total, 9 studies reported that nurses comprised the majority of participants [24,25,29,31-36]. In the remaining studies, nurses participated alongside other professional groups, and in some cases were a small minority. While there was no requirement for studies to have a majority of nurse participants, this variability affects the extent to which findings directly reflect nursing perspectives. Where possible, data specific to nurses was extracted for synthesis. However, in several studies (eg, [26,34,37]) nursing data were supplementary or aggregated with other professions. A full summary of participant demographics and professional backgrounds is presented in Table 2.

Using the MMAT [22], adherence to quality criteria ranged from strong methodological coherence [31,33] to studies with notable limitations such as unclear sampling strategies [27], lack of representativeness [26], and insufficient reporting of nonresponse bias [24,25]. Common issues included limited discussion of divergences between qualitative and quantitative findings in mixed methods studies [29,34], and a reliance on single data sources [37]. Detailed MMAT findings for each study are presented in Table S1 in Multimedia Appendix 4.

Findings from the included studies, grouped by TAM2 constructs, are summarized below and presented in detail in Multimedia Appendix 2. Figure 2 illustrates a conceptual map of the TAM2 and UTAUT constructs weighted by frequency of them being identified in the included studies.

This systematic review synthesized findings from 20 studies reporting RNs’ experiences of AI in clinical practice. Nurses described AI as having the potential to support decision-making, workflow efficiency, and clinical confidence, although these reported benefits were consistently contingent on implementation quality. Perceived usefulness and facilitating conditions emerged as the most frequently reported constructs, with ease of use, workflow alignment, and organizational support recurring as determinants of whether nurses experienced AI as a benefit or a burden. Findings most directly attributable to nurses come from studies with majority nurse samples (notably for Job Relevance and several studies contributing to Perceived Usefulness), whereas constructs such as Output Quality, Subjective Norms, and elements of Facilitating Conditions drew partially on multidisciplinary samples; these are interpreted with caution below. Confidence in each finding is therefore not uniform; Job Relevance and Output Quality are anchored by high-quality studies, Perceived Usefulness is supported by a mostly high-quality evidence base, Perceived Ease of Use and Facilitating Conditions draw on a wider range of methodological quality, and Subjective Norms rests on the smallest and most heterogeneous evidence base.

Across the 20 included studies, nurses generally viewed AI as having the potential to improve clinical workflows and to contribute positively to care delivery, although the breadth of these perceptions varied by AI type and clinical setting. However, challenges were also evident, such as training needs, workload concerns, interoperability issues, and ethical considerations. While constructs such as perceived usefulness and facilitating conditions were frequently discussed, subjective norms and job relevance were less often explored, suggesting underresearched areas with potential to influence adoption.

A key strength of this review is its comprehensive synthesis across diverse AI applications and settings, supported by a transparent search strategy and appraisal process using the MMAT [22]. Another strength is the structured use of the TAM2 framework [16], which provided a consistent lens for analysis and allowed for thematic comparison across studies, although, as noted below, this carries trade-offs for inductive emergence.

There are 6 notable limitations. First, the included studies encompassed a heterogeneous range of AI technologies and contexts, from sepsis prediction tools to natural language processing documentation aids, which limits the generalizability of specific implementation recommendations. In particular, conclusions regarding workflow alignment and facilitating conditions are likely to be highly context-dependent and should not be applied uniformly across all AI tool types or care settings.

Second, many studies involved multidisciplinary participant samples in which nurses were not always the majority, and in several cases, nursing data were aggregated with data from other professions. Where studies included multidisciplinary samples, nurse-specific data were extracted separately where the primary data permitted, as documented in Multimedia Appendix 2. In a small number of studies where profession-specific disaggregation was not possible, findings may reflect broader health care professional perspectives rather than distinctly nursing experiences, and these cases are flagged explicitly in Multimedia Appendix 2.

Third, the deductive analytical framework carries an inherent trade-off. TAM2 was selected a priori and registered with PROSPERO, which ensures transparency and reduces post hoc analytical drift; however, coding line-by-line into prespecified constructs may have constrained the emergence of nursing-specific themes that did not map cleanly onto these categories. We did not encounter substantial extracts that fell entirely outside the framework, although themes relating to professional identity, relational care, or moral distress may be more likely to be visible under inductive analysis. Dimensions such as professional identity, relational care, ethical responsibility, and moral distress in AI-assisted practice may therefore be underrepresented in the synthesis not because they were absent from the primary literature, but because the framework did not create space for them. Inductive or hybrid reanalysis of this literature could potentially surface a complementary picture and is recommended as a priority for future qualitative synthesis.

Fourth, coding of the included studies was undertaken by a single reviewer (AJSS), with regular consultation and consensus discussion of coding decisions with 2 supervising coauthors (BCB and DD) but without independent double-coding of the data. While the consultative process was designed to support consistent application of the codebook, the absence of independent dual coding means that formal interrater agreement could not be calculated, and we cannot fully exclude the possibility of single-coder interpretative bias. This limitation is partially mitigated by the use of a transparent, preregistered framework and by the iterative supervisory review process, but readers should weigh the synthesis findings with this constraint in mind.

Fifth, although the search syntax was not restricted by language, inclusion criteria were limited to studies published in English. This decision was made on practical grounds of interpretive validity, but we acknowledge that it may have excluded relevant empirical work from non-English speaking health care contexts, which could have captured different cultural and organizational dimensions of nurses’ AI experiences.

Finally, the rapid pace of AI development means that nurses’ perceptions are likely to evolve over time, potentially affecting the long-term applicability of these findings.

This systematic review synthesizes nurses’ experiences with AI in clinical practice, identifying both the opportunities and challenges associated with its integration. While nurses often view AI as a means to enhance workflows, support decision-making, and contribute to patient care, adoption is reported to be hindered by barriers such as usability issues, limited training, and concerns about trust and clinical relevance. It is important to note that this review captures nurses’ reported perceptions and experiences, rather than direct evidence that AI improves patient outcomes, workflow efficiency, or clinical safety; such claims require prospective evaluation. Addressing these challenges requires aligning AI development with user needs and embedding robust education and training into both preregistration curricula and continuing professional development. The consistent findings that facilitating conditions are the primary determinant of positive experience suggest potential value in involving nurses meaningfully in the co-design and iterative refinement of AI tools; this inference is drawn from the weight of evidence on implementation prerequisites rather than direct evidence on participatory design processes, which were only sparsely reported in the included literature. Prospective evaluation using nursing-relevant outcomes, including usability, workflow integration, and trust calibration, is needed to establish which specific implementation conditions reliably produce benefit, and to test whether co-design approaches improve adoption in practice.

By closing these gaps, nursing can play a central role in shaping AI tools that enhance, rather than erode, the human elements of care. As AI capabilities evolve, the profession has an opportunity, and a responsibility, to lead their ethical, effective, and patient-centered implementation.