Why Consolidated Governance Statistics Remain Scarce
Industry discussions about AI governance in medical device manufacturing frequently cite adoption percentages, incident counts, and regulatory compliance rates. In practice, no single consolidated and independently verified dataset currently tracks these figures specifically for medical device manufacturers over a defined period. Surveys referenced in trade publications often originate from software vendors, consulting firms, or industry associations, each applying different definitions of a "formal AI governance framework" or an "AI incident." A recall tied to a labeling error is not the same as a recall tied to a model drift event, yet both may appear in the same aggregate figure if the source is not examined closely. Governance leaders building an investment case should treat any cited statistic as a starting point for inquiry rather than a finished benchmark, and should request the underlying survey instrument, sample size, and definitions before using a number in an internal business case.
Pre-Market Validation Versus Runtime Governance
Medical device AI governance is often discussed as a single category, but the practical requirements differ depending on where in the product lifecycle the AI model operates. Pre-market validation addresses how a model was trained, tested, and documented before a device reaches market, and is the stage most associated with existing FDA guidance on AI/ML-enabled device software functions. Runtime governance addresses what happens after deployment: how the model or agent behaves in production, whether its outputs or actions stay within approved boundaries, and how deviations are detected and recorded. A manufacturer can have mature pre-market documentation and still lack visibility into how a deployed model or AI agent is actually behaving across connected systems. When evaluating any governance statistic, it is worth asking which lifecycle stage it describes, since a high adoption figure for pre-market controls says little about post-deployment oversight.
AI Agents Create a Distinct Oversight Problem
AI agents used in manufacturing, quality management, or supply chain operations raise a different oversight question than static predictive models embedded in device software. A static model typically produces a decision-support output that a human reviews before acting. An AI agent may instead be granted permission to query systems, retrieve data, or initiate actions with limited human review at each step. This shifts the governance question from whether a model's output is accurate to what a given agent is permitted to do, under what identity, and who is monitoring its activity in real time. Oversight gaps tend to appear where agent permissions are broad, tool access is not scoped to least privilege, or activity logs are incomplete, rather than in the underlying model's predictive performance. Any statistic describing AI agent adoption in a manufacturing context should specify whether it covers agents with action-taking permissions or agents limited to generating recommendations, since the governance implications differ substantially.
Reading Regulatory Statistics With Appropriate Caution
Regulatory statistics require similarly careful reading. FDA has published guidance addressing software lifecycle management and change control for AI/ML-enabled device functions, but guidance documents are not binding in the same way as codified regulation, and specific provisions are revised periodically. Statistics describing industry compliance with FDA AI governance expectations should specify whether they refer to a binding rule, a non-binding guidance document, or a draft still open for comment. Jurisdiction also matters: figures describing EU MDR or EU AI Act compliance should not be conflated with FDA-specific statistics, since the underlying obligations and timelines differ. Governance leaders should confirm the current status of any referenced guidance directly against the regulator's own published materials before citing it internally, particularly given how frequently AI-specific regulatory language is revised.
Building an Internal Benchmark While Industry Data Matures
- Inventory AI systems by lifecycle stage: Separate pre-market validated models from models or agents operating in production, since governance obligations differ at each stage.
- Define agent permissions explicitly: Document what each AI agent is permitted to access or act on, under what identity, rather than relying on implicit trust.
- Establish runtime monitoring before scaling agent use: Confirm that deployed models and agents can be observed and audited in real time, not only validated at build time.
- Log and retain agent activity: Maintain audit trails of tool calls and agent actions to support internal review and future regulatory inquiry.
- Treat vendor statistics as inputs, not conclusions: Use external figures to frame questions for your own risk assessment rather than as a substitute for internal measurement.