SCPMG Builds a Governance Framework for Clinical AI That Can Move Beyond Oversight
Health systems are adopting clinical artificial intelligence faster than they can govern it, and the gap has become operational rather than conceptual. Southern California Permanente Medical Group has responded with the Collaborative Technology Adoption for Care Transformation framework, or C-TACT, a structure built to manage clinical AI across the full lifecycle rather than only at the point of risk review.
The framework is organized around three pillars — People, Process, and Technology — and four lifecycle stages: Vendor Evaluation, Pilot, Scale, and Maintenance. Each stage is paired with a domain-weighted scorecard, giving the model a practical way to compare projects at different levels of maturity. In value-based care systems, where technology adoption must align with clinical outcomes, workflow fit, and operational execution, that kind of staged governance is more than administrative architecture. It is a response to the reality that AI tools do not create value simply because they are technically accurate.
Why AI Governance Needed a New Operating Model
Traditional governance frameworks tend to concentrate on risk oversight, compliance review, and post hoc monitoring. Those functions matter, but they do not fully address the operational questions that determine whether a clinical AI tool survives beyond a pilot. Who owns implementation? How is adoption measured? What happens when a model moves from a controlled test environment into a live workflow with staffing constraints, reimbursement pressures, and changing clinical demand?
C-TACT attempts to answer those questions by making governance stage-specific. The Vendor Evaluation scorecard addresses the earliest point in the adoption pipeline, while the Pilot scorecard captures whether a tool can function in real clinical conditions. Scale and Maintenance then shift the emphasis toward business and operational execution, recognizing that a tool’s long-term value depends on sustained delivery, not simply a promising first deployment.
A framework like this reflects the economics of modern healthcare as much as the technology itself. Artificial intelligence that cannot be implemented reliably at scale does not change care delivery, even if early performance metrics look strong.
Scorecards, Weighting, and the Logic of the Framework
The C-TACT model uses a domain-weighted scorecard approach, and the weighting is not arbitrary. For the Scale and Maintenance stages, business and operational execution are explicitly incorporated as weighted domains. The framework also applies Analytic Hierarchy Process methods, with a reported consistency ratio of 0.0037, indicating internal coherence in the weighting structure.
That level of formalization matters because governance often fails when it remains too qualitative. A health system can create committees, checklists, and approval layers without ever standardizing how competing priorities are balanced. Clinical utility, implementation burden, organizational readiness, vendor maturity, workflow integration, and sustainment all influence whether an AI tool should advance. C-TACT translates those judgments into a structure that can be applied repeatedly across projects.
The framework is not presented as a finished regulatory standard. It is a governance system designed for internal operational use, and its real significance lies in how it organizes decision-making across the entire adoption pathway.
Retrospective Scoring Separated Deployment Stages by About Thirty Points
In an illustrative retrospective application, the Pilot Score Card separated two published SCPMG AI deployments and four anonymized comparators by approximately thirty points across three weighting schemes. The reconstructed comparator scores ranged from 47.0 to 53.5. In other words, the scorecard did not collapse distinct projects into a narrow middle range; it created a sizable separation between the published deployments and the anonymized comparators.
The result is useful as a demonstration of how the framework behaves in comparative scoring, but it stops short of validation. The comparator scores were reconstructed retrospectively, and they were not scored under blinded conditions. Without prospective, blinded testing, it is not possible to claim discriminative performance or predictive accuracy in real-world deployment settings.
This distinction is important because governance tools can appear persuasive when they are applied after the fact. Retrospective reconstruction may show that a framework can sort known cases, yet that does not prove it will reliably guide decisions before outcomes are known. The clinical and operational stakes are different when a health system is deciding whether to move an AI tool into broader use.
Operational Learning May Matter as Much as Risk Oversight
C-TACT is presented as a framework for operational learning as much as risk control. That emphasis reflects a central problem in clinical AI governance: systems need mechanisms that help them learn which tools can be supported, expanded, or maintained, not just which tools should be blocked. A governance model that only flags hazards can slow adoption without improving implementation quality. A model that also evaluates execution, scale readiness, and maintenance capacity gives leadership a way to manage AI as a clinical service line rather than an isolated project.
The framework’s design also acknowledges the different demands of each deployment stage. A vendor review is not a pilot. A pilot is not scale. Scale is not maintenance. Each phase brings different questions about clinical fit, technical reliability, staffing, and organizational burden, and each phase deserves a separate scoring lens.
That staged structure is especially relevant for value-based healthcare systems, where a technology’s worth depends on whether it can support measurable care transformation over time. The framework is therefore less about celebrating AI than about making adoption governable.
A Framework Waiting for Prospective Validation
C-TACT enters the field as a governance framework and a methodological demonstration, not as a validated prediction tool. The authors call for prospective, blinded validation to establish whether the scorecards can truly discriminate between deployable and non-deployable clinical AI systems under real-world conditions.
That next step will determine whether the framework can move from institutional design to wider operational utility. For now, the evidence shows that a structured, stage-based governance model can be built, weighted, and applied retrospectively in a way that separates deployments meaningfully. What remains unresolved is whether it will perform with the same clarity before a system knows which projects will succeed, stall, or require redesign.
The article leaves that question open, and it does so explicitly. The framework is ready for testing, but not yet for inference about broad predictive performance.