AI Governance for Healthcare¶

Know which AI is running in your hospital, who approved it, and whether anyone is still watching.

Try the demonstration What adopting it takes

What is this?¶

Your hospital already uses AI, and someone approved each tool at some point. This is free, open source software that keeps one shared record of all of them: who approved each tool, on what evidence, and when it is due for another look. It does not test AI tools, and it does not promise safety. It gives your committee one place to keep the answers and your board one place to read them.

Why does this matter now?¶

AI is already here. In 2024, 71% of US non-federal acute care hospitals used predictive AI built into their electronic health record, up from 66% a year earlier (federal data brief).

Most hospitals check it. Showing the checks is the hard part. Of the hospitals using predictive AI, 82% said they evaluated it for accuracy and 74% for bias. Those answers are self-reported. They do not say whether a check was written down, repeated, or owned by a named person, and that is what a board, an insurer or a regulator asks to see.

Expectations are rising, mostly as guidance. As of October 2026:

• In September 2025 the Joint Commission, which accredits hospitals, and the Coalition for Health AI (CHAI) published guidance on the responsible use of AI. CHAI is a coalition of health systems, patient advocacy groups, technology companies and others. The guidance is not a legal requirement. In 2026 the Joint Commission also launched a voluntary certification for how a hospital governs AI.
• A federal rule, 45 CFR 92.210, requires organizations that receive federal financial assistance, which includes many hospitals, to make reasonable efforts to find the patient care decision tools that use race, color, national origin, sex, age or disability as inputs, and to reduce the risk of discrimination from them.
• States are adding their own. In October 2026 California approved two laws that reach hospitals using AI clinical decision tools. From January 1, 2027, AB 1979 requires health facilities to take reasonable steps so that a licensed provider uses independent judgment when AI informs care. SB 503 requires facilities that deploy such tools to make reasonable efforts to identify systems that could produce biased results, to monitor their effects, and to act on problems (California Health Care Foundation).

Laws change, and this is not legal advice. Ask your counsel which apply to you.

Why is one approval not enough?¶

Published studies show that an AI tool can disappoint once it is in your hospital, and can change after you approve it.

• A tool can work worse for you than the vendor said. Michigan Medicine tested a widely used sepsis alert on its own patients (Wong, 2021). It missed about two thirds of the patients who developed sepsis and raised an alert on 18% of all hospital stays. The authors called this substantially worse than the developer reported. It was one hospital, looking back at past data.
• A tool can change under you. A letter in the New England Journal of Medicine reports that Michigan later switched the same alert off because of false alarms after a shift in its data (Finlayson, 2021).
• A tool can look accurate and still be unfair. A widely used program for choosing patients for extra care favored white patients over sicker Black patients, because it predicted cost instead of illness (Obermeyer, 2019).

None of this means AI should stay out. It means a hospital owes its patients a process, and needs to be able to show that it ran one.

What happens when you cannot answer?¶

Sooner or later a board member, an insurer, a regulator or a lawyer asks:

1. Which AI tools are in use in our hospital today?
2. Who approved each one, when, and on what evidence?
3. Is anyone still checking that it works for our patients?
4. If something went wrong, what can we show?

Without one list, the answers sit in meeting minutes, email and one person's spreadsheet. They take time to gather, and they are hard to vouch for.

We found no published figure for what that costs, so we will not invent one. Ask your committee secretary how many hours the last board question took. That is your baseline.

What do you get?¶

• One list, with an honest status. Each tool shows as Out of compliance, Needs update, On track or Retired. Fixed rules set the status, so nobody types it in. The labels are this software's own, not a legal finding, and the rules check that a record is complete and current, not that it is true.
• Decisions with names on them. Four checkpoints (approve the design, approve a pilot, approve deployment, and a periodic review) each record who decided, when, and the reason the reviewer chose to write. In the demonstration copy a name is only typed in, and nobody checks it.
• Reviews that come back. Each deployed tool has a review date. In the 30 days before it, the tool shows as Needs update. Once the date has passed it shows as Out of compliance.
• Reports your board can read, and a spreadsheet of the whole list.

What it does, in detail covers all of this, including what the record protects and what it does not.

What does it cost, and what does it not do?¶

There is no license fee. You pay for hosting, your IT team's time and your committee's time. You also need a sponsor, a committee with the power to say stop, and a named owner for each tool. We give no dollar figure.

For scale, one academic medical center reported that a fairness and reliability review of two prediction models took 115 person-hours over 8 to 10 months (Stanford, 2022). That is one center's experience, not a price list, and this software does not do that work for you.

What it does not do:

• It has no support. There is one maintainer and no support contract, and no response times are promised. You hold the source code and the license to carry on without us.
• It does not find or test AI tools. People enter what they know. "Met" on a checklist is a reviewer's statement, not a verified fact. Nothing checks whether a model has drifted.
• It is not a certification. It does not say that any use of AI follows a law.
• It is not built to hold patient data. Nothing technical stops someone typing it into a free-text box, so your staff need to be told.
• It has not been shown to make patients safer. A 2025 review found only 16 studies of AI ethics frameworks being put into practice in health care, and none showed a link to patient outcomes (the evidence, honestly).

How can you start small?¶

Open the demonstration. It runs in your browser with sample data and needs no server. It does not keep a trustworthy history or check who signs, so it is for looking, not for real records.

Then ask your own team three questions. Who would sponsor this? Which committee could say "stop" to an AI tool? Which one tool would you start with? In a trial, measure how long the next board question takes to answer, and how many tools you did not know about. Adopting it has the full list.

See it in action¶

A 90-second walkthrough, using sample data only. It shows the portfolio dashboard and its statuses, opening a project and its access settings, the stage checklists with their evidence and owners, the go/no-go checkpoints, the applied model card preview, the OPTICA adoption review, and creating a project.

Where to go next¶

Sources¶

Checked October 12, 2026. Laws and guidance change.