JMIR Publications

ABSTRACT

Background:

Artificial intelligence (AI) is transforming patient care but also raises ethical questions such as bias and transparency. While a range of well-established frameworks exist to guide responsible AI practice, most were designed for academic or regulatory settings and can be hard to operationalize within fast-moving, resource-limited small and medium-sized enterprises (SMEs).

Objective:

We introduce the Scalable Agile Framework for Execution in AI (SAFE-AI). SAFE-AI embeds ethical safeguards such as fairness, transparency, and continuous monitoring within standard Agile product-development cycles, while remaining practical for organizations without dedicated ethics teams.

Methods:

We followed a design-science, practice-oriented approach over 20 weeks. After a needs-finding workshop, a cross-functional team from an SME, ethics researchers, and academic partners met weekly in Agile sprints, continuously reviewing relevant literature and regulations. Through three prototype-feedback cycles the group iteratively refined a four-phase SAFE-AI lifecycle, acceptance/fairness/transparency checklists, and scenario-based responsibility metrics, recording decisions until unanimous consensus.

Results:

The co-design process produced a four-phase SAFE-AI life-cycle: Discovery, Assessment, Development, Monitoring. SAFE-AI’s phase-specific checklists melds acceptance, fairness, and transparency metrics into each Agile sprint. A novel scenario-based probability-analogy mapping (SPAMM) method was added to translate model risk and uncertainty into plain-language narratives for non-technical stakeholders, forming the framework’s core “responsibility metrics” layer. To keep oversight lightweight, SAFE-AI defines clear triggers that automatically reopen ethical review whenever models are retrained, tuned, or fed new data, ensuring consistent re-evaluation without duplicating earlier work.

Conclusions:

SAFE-AI shows that meaningful ethical safeguards can be embedded within standard Agile workflows without slowing delivery or requiring a full-time ethics team. Its checklist-driven phases and automatic review triggers provide a lightweight yet defensible way to track fairness, transparency, and responsibility throughout the model lifecycle.