Start with an inventory, an owner and an off switch.
I gathered the free and open resources I trust most for hospitals putting AI into practice, organized around seven steps every hospital can follow. Each resource is credited, dated and labeled with what you may do with it. It is the companion to my AI Education Atlas.
What I found
Guidance now exists for nearly every step, from CHAI's playbooks to NIST's risk framework. The weak links are local evaluation and knowing when to stop a tool.
So what
A course certificate is not proof of safe use. A 2026 review of 27 training studies found none that measured organizational change or patient benefit.
Now what
Pick your step, open the matching resources, and produce the document listed for that step. Small hospitals can start with one accountable group and a simple registry.
Educational resource only. This is not legal, regulatory, compliance or medical advice. I am not affiliated with, and not endorsed by, the organizations listed. Listing a resource is not an endorsement or a certification. See the full disclaimer at the end.
The hospital pathway
Seven steps, in the order a hospital should take them
My proposed sequence, drawn from the resources below. It is not a quoted accreditation checklist. Tap a resource to jump to its card.
If you run a small or rural hospital
Start with one accountable multidisciplinary group and a simple spreadsheet registry, not a new bureaucracy. Put deeper review only on high-consequence tools, and write local validation into vendor contracts while you still have leverage.
Worked example: an ambient scribe
Fields I would ask a hospital to evaluate locally: critical omissions, medication errors, clinician editing time, privacy, how patients are told, and what happens when the tool is down. These are proposed fields, not a claim about any product.
What training proves
27 studies of AI training. None measured patient benefit.
Source: Woods L, et al. BMC Medical Education 2026;26:549 [11]. Studies can report more than one level, so bars do not add to 27. Literature searched August 20, 2024. CC BY 4.0.
Knowledge
The learner knows more. Most studies stop here.
Competence
The learner uses AI safely at work. Rarely measured.
Outcomes
Patients or the organization benefit. Not yet shown.
Treat these as three separate claims. Completing a course supports the first, not the other two.
Resource library
Find the resource for your role and your step
How I label access and reuse
Open-source softwareCode with an explicit license that allows reuse and modification.
Open-access articleFree to read. Reuse depends on its Creative Commons license.
Free public guidanceFree to read. Check permission before copying or adapting.
Membership or applicationSome content or the program itself needs an account, membership or fee.
"Free to read" and "allowed to reuse" are different questions. Each card answers both separately.
Documents to produce
Ten documents a well-governed hospital can show on request
AI inventory with owner, users, vendor, version and status
Governance charter and responsibility matrix
Vendor due-diligence questionnaire
Risk and impact assessment
Local validation plan and signed approval
Role-based competency checklist
Patient communication worksheet
Monitoring and incident plan
Change-control record
Tested rollback and retirement plan
Keep patient-level data out of any self-audit. Use policies, aggregate results and named owners.
Words to avoid without a documented basis
"CHAI approved"
"Joint Commission compliant" or "certified" before certification is granted
"FDA approved" for a tool that is not on an FDA authorization record
"Bias free" or "validated" without local results
"Hospital-ready" for an open-source tool that has not had security and clinical review
A consensus guideline is not a regulator. A certification covers the organization's practices, not a product.
Credits
Every resource belongs to its authors
Summaries are mine, in plain language. Read the original before quoting or relying on any detail. Last verified November 13, 2026.