Clinical AI governance framework guides AI tool evaluation

Health systems adopt clinical AI faster than governance structures allow. A new npj Digital Medicine article in press from Southern California Permanente Medical Group (SCPMG) proposes a clinical AI governance framework that scores tools across people, process and technology at each stage, from vendor evaluation through to maintenance. The authors, led by Dinh Nguyen, present it as a governance instrument and a methodological demonstration, not as proof that it works.

How the clinical AI governance framework works

The Collaborative Technology Adoption for Care Transformation (C-TACT) framework organises governance across three pillars: People, Process and Technology. Four lifecycle stages run across them: Vendor Evaluation, Pilot, Scale and Maintenance. Each stage has a scorecard with weighted domains, reviewed by a multidisciplinary council that includes clinical, operational, financial, IT, quality and legal expertise.

Widely used frameworks, including the NIST AI Risk Management Framework and WHO guidance, cover business and operational execution only indirectly. C-TACT weights those areas explicitly, alongside safety, ethics and equity. Under its rule, a pilot needs at least 70 out of 100 in aggregate and no domain below 5 out of 10 to receive a conditional go-ahead to scale. A pilot that clears the aggregate but has a weak domain, such as data readiness, goes to an extended pilot instead. A strong scalability score cannot make up for weak data readiness.

Ad

Weighting domains with an analytic hierarchy process

The Pilot scorecard has eight domains. The paper derives its weights with the analytic hierarchy process (AHP), a structured pairwise-comparison method. The exercise was built from the existing expert weights with small deliberate changes, not from independent comparisons by council members. The consistency ratio of 0.0037 therefore shows that this worked example is internally coherent, not that reviewers agreed with one another. The AHP weights differed from the expert weights by at most 3.3 percentage points, and patient safety ranked highest.

Two deployed tools and four comparators

The authors scored two SCPMG products reported in earlier studies: the Smart Messaging Tool, an NLP system for triaging secure messages, and the Kaiser Permanente Intelligent Navigator, a self-service booking and navigation platform. They scored them alongside four anonymised comparator tools built from the literature cited in those studies. The two deployed tools scored 79.0 to 80.5 across three weighting schemes. The four comparators scored 47.0 to 53.5. The gap of roughly 30 points held under every weighting approach, and the 70-point threshold separated the two groups.

The comparator scores were reconstructed after the fact from public evidence, and two of the four comparators had already been superseded by the SCPMG products they were compared against. Reviewers were not blinded to product identity. The gap should not be read as an estimate of how C-TACT would discriminate between competing products assessed prospectively. The 70 and 85 point thresholds are heuristic, not empirically calibrated.

What comes next

Prospective, blinded validation is the priority. Planned work includes inter-rater reliability studies with at least four trained reviewers per scorecard, retrospective calibration of thresholds against SCPMG governance outcomes, and multi-system comparisons that use deployment outcomes as the endpoint. Governance should support iterative learning, not only gatekeeping, since an unsuccessful pilot can still reveal workflow barriers worth fixing. Similar efforts are underway in Portugal and through regional responsible AI networks aimed at low-resource settings.

The study was deemed exempt from review by the Kaiser Permanente Institutional Review Board, used no patient-level data, and followed the SQUIRE reporting guideline.

Source: Nguyen D, et al. “Clinical AI governance for scalable innovation in value based healthcare systems.” npj Digital Medicine, article in press, doi:10.1038/s41746-026-03409-4.