1,000+ AI Devices Cleared. Who Governs How Facilities Use Them?

By mid-2024, the FDA had authorized over 950 AI and machine-learning-enabled devices in radiology alone. In cardiovascular medicine, the number reached 92 by early 2025 and continues to grow. Across all medical specialties, the total has surpassed 1,000 cleared devices.

The FDA's regulatory framework for these devices is well-established. Manufacturers submit evidence of safety and efficacy through 510(k), De Novo, or premarket approval pathways. The agency evaluates the device, clears or approves it, and the manufacturer can market it.

Here is what the FDA does not regulate: how a specific facility deploys that device in clinical practice.

The governance gap

When a hospital purchases an AI-powered cardiac imaging analysis tool, several questions arise that fall outside the FDA's jurisdiction:

• Who at the facility is responsible for validating the device's performance on their patient population?
• What happens when the AI's output conflicts with the clinician's interpretation?
• How does the facility track the device's accuracy over time?
• What are the escalation protocols when the AI produces an unexpected result?
• How does the facility ensure that staff using the device understand its limitations?

These are not hypothetical concerns. They are operational realities that every facility deploying clinical AI must address. And currently, no accreditation framework systematically evaluates whether facilities are addressing them.

The three actors and their boundaries

Understanding the gap requires mapping who governs what:

The FDA regulates the device manufacturer. It evaluates whether the device performs as claimed based on the evidence submitted. Once cleared, the FDA's ongoing oversight focuses on post-market surveillance, adverse event reporting, recalls, and manufacturer compliance.

Accreditation bodies evaluate facility-level operations. They assess whether a facility has qualified personnel, appropriate equipment, adequate quality improvement programs, and compliant documentation. But most accreditation standards were written before clinical AI deployment was common. The standards address the human and equipment components of clinical care, not the algorithmic components.

Professional societies publish clinical practice guidelines that may reference AI tools, but guidelines are advisory. They do not create enforceable governance frameworks, and they address clinical best practices rather than operational deployment requirements.

The result is a governance vacuum at precisely the point where clinical AI meets clinical care.

What facility-level AI governance would look like

A meaningful accreditation framework for clinical AI deployment would need to address at least five domains:

Vendor identification and tracking. Facilities deploying multiple AI tools from different vendors need a systematic way to track which tools are active, what versions are running, what FDA clearance each holds, and what risk tier each represents. Today, this information often lives in procurement records that are disconnected from clinical operations.

Validation protocols. A device cleared by the FDA based on a training dataset from one population may perform differently on another. Facilities should validate AI tool performance on their own patient demographics, imaging equipment, and clinical workflows before full deployment, and periodically thereafter.

Override and escalation policies. When a clinician disagrees with an AI output, what happens? The answer should be documented, consistent, and auditable. Currently, many facilities have no formal policy for AI-clinician disagreement.

Performance monitoring. AI device performance can degrade over time due to data drift, software updates, or changes in clinical workflow. Facilities need ongoing monitoring mechanisms, not just at deployment, but continuously.

Staff competency. Using an AI tool effectively requires understanding its capabilities, limitations, and failure modes. Accreditation standards should verify that clinical staff have been trained on the specific AI tools they use, not just on the clinical procedures those tools support.

Why accreditation bodies should care

Over 1,000 devices are already cleared. Hospitals are deploying them now. The question is whether accreditation standards evolve to cover this deployment or whether AI governance at the facility level remains unstructured and unaudited.

The case for accreditation involvement is strong. Accreditation bodies already evaluate facility quality across personnel, equipment, and clinical processes. Adding AI governance is a natural extension of that mandate. More importantly, it is a responsibility that no other actor in the healthcare system is positioned to fulfill.

The FDA cannot practically evaluate every facility's AI deployment. Professional societies can publish guidance but cannot enforce it. Payers can set requirements but lack the operational assessment capability. Accreditation bodies have the assessment infrastructure, the facility relationships, and the quality mandate.

The first movers

A small number of standards organizations have begun developing AI governance standards. These early efforts tend to focus on documentation requirements, policies, training records, validation reports. This is a reasonable starting point, but documentation alone is insufficient.

The next generation of AI governance standards will need to be evaluable against structured data. Does the facility's AI monitoring system detect performance degradation? Can the vendor identification registry be queried programmatically? Are override events logged and reviewable?

These questions require the accreditation infrastructure to evolve, from evaluating paper documentation to evaluating data pipelines. That infrastructure evolution is the precondition for meaningful AI governance at scale.

---

Regain Accreditation provides continuous compliance monitoring infrastructure for accreditation bodies worldwide. Request a demo →