Scoping review of governance, privacy, and crisis safety in AI-supported conversational agents for perinatal mental health

Abstract Purpose Perinatal mental health conditions affect nearly 20% of mothers and 10% of fathers, yet access to timely care remains limited. AI-supported conversational agents (CAs) offer scalable support, but their governance remains poorly characterised. This scoping review examined how perinatal CAs are built, governed and evaluated, whom they serve, and what safeguards they incorporate. Methods Following JBI guidance, PRISMA-ScR, and a registered protocol, we searched nine interdisciplinary databases (January 2015–September 2025) and tracked citations. Two reviewers independently screened records, and findings were charted and summarised descriptively. Results Of 6,118 records, 35 studies describing 30 unique CAs across 14 countries were included. Rule-based architectures predominated; large language model and hybrid systems emerged rapidly after 2022. Twenty-two studies described screening, treatment, or crisis-related functions, although therapeutic foundations were inconsistently reported. None reported advanced privacy-preserving computation. Fourteen studies sent conversational content to commercial application programming interfaces (APIs), none reporting data jurisdiction or retention. Crisis safeguards narrowed progressively: 18 studies reported crisis provision, ten automated risk detection, and three linked detection to human follow-up. Only one specified a response timeframe; none reported crisis-detection sensitivity. Governance reporting was limited (median exploratory Governance Maturity Score, 7/24), with monitoring and post-deployment updating least documented. Fathers, couples, non-birthing partners, linguistically diverse populations, and low-connectivity users were poorly represented. No system was developed in a low-income country. Conclusions We propose eight minimum standards to guide developers, clinicians, and policymakers towards governance-first development. These prioritise validated crisis-response pathways, transparent data handling, inclusive design, human oversight, and continuous post-deployment monitoring.