The Committee Nobody Staffed

The 2025 Health Affairs article examined American Hospital Association survey data to explore predictive AI incorporation and evaluation by US hospitals. Predictive modeling was widely adopted by hospitals to enhance operations and patient care. However, very few reported reviewing those models for accuracy against their patient population, and less than half those surveyed evaluated for bias. The distance from the model adoption to model evaluation is the primary focus of this edition, as it is the gap the AI committee announcement aspired to close. The functioning version of that committee follows the announcement in both time and money, as the announcement is the part that is required.

It is an easily accessible announcement. For the past two years, hospitals have been establishing AI oversight committees, and the trend continues. A charter has been created, members have been respectively named, and the first meeting has been scheduled. Then the operational question follows. Who reads the model documentation on a Tuesday afternoon? A committee that meets, and has no one working in between meetings, has the ability to approve and monitor very little. Its charter will define a purpose. The actual purpose of the committee will be shown in the scheduling.

What Review Actually Demands

The actual demands of a review body are considerable. Requests must be prioritized and triaged. Each part of a step must be assigned to a process owner. Drafts must be read, and local validations must be scheduled with the data teams. The behavior of models must be validated post deployment, and an issue reporting framework must be created that defines and enables clinicians to know when and how to report issues. Each of the demands must be addressed continuously. Each of these activities steals hours away from the team, and in order to meet some of the demands the team must have the capacity to address the demands.

Think about what is required to perform the intake step. In distributed systems, service lines are triggered to activate newly shipped vendor features. An analyst has built a tool to support a manager and that tool is being used. In systems where no intake step exists, the requests are processed without review.

The healthcare sector has working knowledge of a well-functioning review process since they have operationalized review processes. Institutional Review Boards are fully staffed offices with paid coordinators who manage the review process. No one expects a fully staffed IRB to operate in volunteer mode. The AI committee is expected to run like an IRB. This is the fundamental flaw of the design.

Staffed functions do not completely analyze all aspects of a system, similar to a small team managing a large portfolio. A scheduling system that autonomously books follow up meetings should not be held to the same standard as a system that makes clinical decisions. The principle of risk based tiering means intensive review focuses on areas of the highest potential risk. By contrast, areas of light review should process requests quickly enough that the overall system is not undermined. Those present in the review process are as critical as the tiers themselves. The presence of clinical informatics, nursing, data science, and a privacy owner, is critical as they will each identify failure modes the others do not.

The Expense That Wasn't Addressed

What would it take to make this functional? An actual owner with actual guarantees—not a title that gets layered on top of an already full schedule. Money that we can see, unlike one contract for enterprise AI. Budget cycle after budget cycle, this will exist because it costs nothing. What would the committee be protecting from the announced systems? One example of a model that reaches clinicians, if it does not have an executive level review, requires more money than paying for the model and the reviewers for a year. It will be the worst time for the executives.

What you see is the staffing problem. It will provide a response to an inquiry given by a board member. It will describe what systems at peer organizations do on the same day the announcement is made. Money for staffing directs attention nowhere. It shows up as a visible and repeated expense, with no formal celebration for its existence. It would be recognized when something goes wrong and would have been preventable without it. The structures of organizations favor the first expense, which is why the deviation from the norm has to be deliberate.

External gap expectations are widening. In 2025, The Joint Commission and the Coalition for Health AI began jointly publishing documents regarding the responsible use of AI in hospitals. It is evident that what the review body is responsible for is becoming what the systems are going to be assessed against. A committee will certainly be assessed very differently when it can produce a record of what it has reviewed.

The First Deliverable Is a List

The first step before an organization makes a decision and creates a new review function is to list the models that are in use and deployed outside of formal processes, and that are left outside of any policies. In the last edition, The Tools You Did Not Approve, to say that it is a contradiction to oversee a portfolio of things that you do not know about is an understatement. In the inventory that is created, the committee can see the systems that they have, and it creates a backlog for the review sessions. This system creates a backlog for the review sessions. Simply put, this process shows that it can be accomplished easily, but it requires a budget. The announcement does not indicate whether they have the budget to support the work.

This record is straightforward and lists, for each model, its owner, and the date and findings of the review. A board that receives this record quarterly will have more oversight in total than has been accomplished by most AI committees after two years of meetings. This record should be kept as a paid, named responsibility.

Half of that inventory includes unofficial integrations — and it matters just as much as the deployed half. Clinicians will often adopt tools that have not been sanctioned — and bringing those tools to light without fear of retribution is work that the review function will do by default. The tools that go unmessaged are the tools most likely to bring about the kind of disruption that the committee is trying to avoid.

During my recent interview with Pranava Adduri, on The Signal Room, we talked about how AI agents are being placed in healthcare organizations at a speed that exceeds the capability to build oversight structures. The systems that create oversight functions before the healthcare AI agents are deployed will most likely be the systems that function the best. The same is true even in the case of the review function. The committee that is built before a healthcare AI agent breaks the system will most likely be the most effective at that point in time.

What to Ask This Quarter

Adequate oversight of AI systems can be evaluated by two questions posed to executive and board members: What has our AI review body analyzed in the last 90 days? Where can we see the output of that review? If the response is a charter and not a report, the committee is a theoretical construct. The decision to appoint members to that functional committee is negligible. It is a decision in favor of committing resources to a function that the organization claims to have in its public disclosures.

A third question may follow after the first two have been answered. Who is compensated to complete this work between committees? If the answer is a list of people for whom this is the fourteenth duty, then that committee has members and no staff, and work between meetings is nonexistent.

An honest ninety-day record shows a lack of vanity, and the record's credibility improves with a lack of vanity. It may show a few reviews that have been completed, and one tool being restricted and the rationale for this. A record that shows perfect coverage and no findings has been created for the reader instead of being produced from the real work.

Context and Sources

This edition draws on the 2025 Health Affairs analysis of American Hospital Association survey data on hospital evaluation of predictive AI, and on joint 2025 material from the Joint Commission and the Coalition for Health AI on responsible use of AI in hospitals. It continues themes from issue 29, 2026 Is the Year of Oversight, issue 38, Why Oversight Without Ownership Fails, and issue 46, Who Answers for the Model?

The AI Health Pulse is a weekly briefing on healthcare AI strategy and oversight that provides independent insights for busy executives. It is written by Christopher Hutchins, a former health system data executive and the founder of Hutchins Data Strategy Consultants. The publication is free from sponsorship and paid placements. Each edition is based on named sources, and the operational aspects of healthcare AI strategy and oversight are covered, such as data readiness, model oversight, and the impact of AI on an organization. You can subscribe at https://aihealthpulse.beehiiv.com/subscribe. The complete archive is also free to access.

Companion reading: Best Healthcare AI Newsletters for Executives, An Honest List: https://aihealthpulse.beehiiv.com/p/best-healthcare-ai-newsletters-for-executives-an-honest-list

Christopher Hutchins Founder & CEO, Hutchins Data Strategy Consultants