Responsible AI Off-Boarding in Radiology: Staff Perspectives on Decommissioning and a Proposed Withdrawal Framework.
Authors
Affiliations (5)
Affiliations (5)
- Epsom & St Helier University Hospitals NHS Trust, London, United Kingdom, SM5 1AA; CRRAG Research group, City St George's University of London, Northampton Square, London EC1V 0HB. Electronic address: [email protected].
- Epsom & St Helier University Hospitals NHS Trust, London, United Kingdom, SM5 1AA. Electronic address: [email protected].
- St George's University Hospital, Blackshaw Road, London, SW17 0QT. Electronic address: [email protected].
- CRRAG Research group, City St George's University of London, Northampton Square, London EC1V 0HB. Electronic address: [email protected].
- CRRAG Research group, City St George's University of London, Northampton Square, London EC1V 0HB; Department of Clinical Radiology, Great Ormond Street Hospital for Children, London, WC1H 3JH, UK; UCL Great Ormond Street Institute of Child Health, Great Ormond Street Hospital for Children, London, WC1N 1EH, UK; NIHR Great Ormond Street Hospital Biomedical Research Centre, 30 Guilford Street, Bloomsbury, London, WC1N 1EH, UK. Electronic address: [email protected].
Abstract
AI governance commonly emphasises procurement, validation, deployment and performance monitoring but lack guidance on how embedded AI tools should be withdrawn when funding, contracts or strategy change abruptly. We examined staff experience after withdrawal of a chest radiograph AI triage tool and developed a practical framework for responsible AI off-boarding. An anonymous staff survey was circulated two months after decommissioning of a chest radiograph AI triage tool at a multi-site NHS trust. The survey repeated selected items from three earlier implementation-phase surveys, with added decommissioning-specific questions on workflow, perceived patient benefit, emotional burden and future AI engagement. Quantitative responses were summarised descriptively. Free text responses were analysed deductively and interpreted using the Job Demands-Resources (JD-R) model. The response rate was 21.4% (40/187), comparable to earlier survey rounds. Perceived patient benefit from AI remained stable, with 70% (28/40) agreement post-decommissioning versus 71.1% (32/45) pre-implementation, 65.5% (19/29) early implementation, and 67.9% (36/53) late implementation. Perceived logistical burden post-decommissioning (35%, 14/40) was higher than at late implementation (26.4%, 14/53) but lower than at early implementation (51.7%, 15/29). Response rate from reporting staff was low (5/40), with two voicing disappointment in AI tool withdrawal, and one stating they had grown reliant on the tool. Across 22 free-text entries, frequent themes included loss of clinical value or pathway efficiency, operational relief after withdrawal, and patient-facing emotional labour during AI-enabled escalation. Decommissioning produces both operational relief and perceived clinical loss, with effects differing by staff role. We propose a three phase AI off-boarding protocol - pre-withdrawal assessment, graduated transition and post-withdrawal support. Decommissioning management should be treated as a central part of responsible governance across the AI lifecycle.