Primary Care Doctors' Perspectives and Experiences With a Chest X-Ray AI Triage Program: Qualitative Study.

BACKGROUND: AI has the potential to support chest X-ray (CXR) triage in primary care, but adoption depends on whether clinicians perceive its outputs as credible, useful, and workable within routine clinical workflows. Evidence on how primary care doctors experience AI-supported CXR triage in real-world practice remains limited. OBJECTIVE: This study explored primary care doctors' perspectives on a pilot CXR-AI program and identified barriers and enablers influencing adoption during early implementation. METHODS: We conducted a qualitative descriptive study in a Singapore public primary care center where an AI system was embedded into the CXR workflow as a triage tool. Doctors who had used the program in clinical practice were purposively sampled across age, gender, and clinical seniority. Data were collected through semistructured in-depth interviews and focus group discussions, audio-recorded, transcribed verbatim, and analyzed using thematic analysis. RESULTS: Twenty primary care doctors participated in 10 in-depth interviews and 2 focus group discussions. Adoption was variable and shaped by three interconnected themes: (1) AI validity and workflow integration, (2) clinician beliefs and confidence, and (3) organizational culture. Initial engagement appeared to be shaped by whether doctors understood the program's purpose, perceived a need to change existing practice, and were open to workflow change. Continued use was shaped by the perceived accuracy of the AI tool and its usefulness in clinical practice. Doctors perceived the AI tool as more valuable when they were confident in CXR interpretation. Institutional endorsement, phased implementation, positive peer experiences, and the safety net provided by continued radiologist reporting helped build trust. However, concerns about AI overcalling, lack of clinical context and interaction, and medicolegal responsibility limited clinicians' willingness to rely on AI alone. CONCLUSIONS: Adoption of AI-supported CXR triage in primary care depended not only on the technology itself, but also on how it was introduced, understood, and experienced in practice. These findings support the need for implementation strategies that are responsive to end user perspectives and contextualized within local workflows and clinical settings. Further research should examine later-stage implementation outcomes and objective operational and clinical outcomes of the program.