How to Build a Healthcare AI Education Program

We have been designing a healthcare AI education program. The outline begins with a decision that comes before prompts, tokens, or risk definitions.

What should the organization say to its people?

We moved the definitions down and put the organization’s AI mission first: why it is pursuing AI and how the technology should reinforce the humanity of healthcare.

From there, the program can build toward real use cases, safeguards, governance, role-based learning, and measurement.

Employees bring curiosity, experience, fear, and skepticism into the room. Some use AI every day. Others hear that it will change healthcare and wonder what that means for their jobs.

Leadership owes them an honest answer.

In the American Medical Association’s 2026 survey of 1,692 physicians, 40% said they were equally excited and concerned about AI. Eighty-eight percent expressed at least some concern about losing skills, 85% wanted to be involved in adoption decisions, and 92% wanted more education and training.

The numbers match what we hear: people want context, a voice in decisions, and practical help.

Begin with the reason for AI

Most AI education opens with definitions. Machine learning is this. Generative AI is that. A large language model predicts the next token. Those lessons can wait a few minutes.

Start with what the organization is trying to accomplish. Connect AI to the mission, vision, and goals employees recognize, then explain how it can support patient care, the workforce, access, quality, and operations.

Healthcare depends on listening, judgment, empathy, trust, and the relationship between patients and the people caring for them. Explain how AI can create more room for those things, which decisions stay with qualified people, how tasks may change, and how workforce impacts will be handled.

Carry that message through the examples, safeguards, governance process, and measurement.

The evidence remains insufficient to connect a mission-centered opening with higher adoption. We recommend it as a program-design decision. If an organization wants people to understand its AI strategy, its education has to sound like that organization.

What must be ready before the first module

The content should come from a small set of operating decisions. We would gather these before production begins:

Show people where AI is already appearing

“AI” is too broad to feel real.

Show the workforce the actual portfolio: production tools, pilots, uses under review, and areas where the organization has chosen to wait. A fall 2024 peer-reviewed survey examined 37 prespecified use cases in 10 categories at 43 health systems. The risks and responsibilities change with the work.

An organization might share examples across four groups:

• Patient care and clinical decisions: imaging support, clinical decision support, deterioration, readmission and sepsis prediction, remote monitoring, virtual nursing, and computer vision.

• Documentation and communication: ambient listening and voice, AI scribes, chart summaries, patient messages, translation, and coding support.

• Access and service: patient chatbots, employee chatbots, contact-center support, scheduling, navigation, and enterprise search.

• Operations and business functions: revenue cycle, prior authorization and denials, staffing, capacity, patient flow, supply chain, quality, cybersecurity, and agentic workflows.

For selected use cases, explain the problem, users, owner, approval status, and lessons so far. In our experience, that tour makes the strategy concrete and gives the rest of the education somewhere to land.

Teach enough about how AI works to understand how it fails

We recommend a useful mental model for everyone. Current guidance more directly supports deeper training for people who use, supervise, procure, govern, or monitor AI.

The NIST Generative AI Profile explains that a language model predicts the next token or word from patterns in its training data. Its answer can be fluent and wrong. A predictive model estimates the likelihood of an outcome from the data and population used to build it. Performance can change across populations, settings, and time, requiring local evaluation and monitoring.

Generated content can hallucinate, use stale information, invent a citation, or omit a critical detail. Predictive systems can create false alarms, miss cases, perform unevenly, or drift. People add automation bias, overreliance, skill loss, and unclear accountability.

Privacy and security risks include indirect prompt injection through retrieved content. Employees need to know which data can enter an approved tool, which sources it can reach, and whether that access fits the task.

Teach a response with each risk:

• For fabrication, omission, and stale information, trace material claims to the chart, policy, record, or approved source.

• For bias, uneven performance, false alarms, missed signals, and drift, know the intended use and population. Compare the result with the person and situation in front of you, then report patterns.

• For automation bias and overreliance, form an independent judgment when the decision allows it, check the details with the greatest consequence, and keep practicing the underlying skill.

• For privacy and security, use approved systems, share the minimum necessary data, confirm permissions, and treat retrieved content as untrusted.

• For responsibility gaps, name the person who reviews the output, the person who owns the system, and the path for raising a concern.

Training belongs inside a wider safety system

A 2026 randomized vignette study shows the limits of education alone. Forty-four physicians who had completed 20 hours of AI-literacy instruction evaluated six simulated cases. Those offered AI recommendations with deliberate errors in half their cases averaged 73.3% on diagnostic reasoning, compared with 84.9% for those offered error-free suggestions. The adjusted difference was 14 percentage points.

The study used one model, one country, six simulated cases, and no untrained comparison group. It leaves the protective effect of training and the results in routine care unanswered.

Prior instruction did not eliminate vulnerability in this setting. Guidance from NIST and the Joint Commission with the Coalition for Health AI also calls for monitoring, accountable review, incident reporting, and clear responsibilities. Pair education with point-of-use checks and an easy way to stop or escalate a concerning use.

That is why we teach The SIGN Method as a practical routine:

• Scope the task, tool, data, and decisions that stay with a person.
• Instruct the system with a clear goal, context, boundaries, and expected result.
• Ground material claims in the chart, policy, record, or other verified source.
• Name who checked the work, who is responsible, and where concerns go.

SIGN is a teaching framework. Organizations should evaluate whether it changes decisions and behavior in their own workflows.

Make governance something everyone can use

As we wrote in Education for AI Governance, governance often reaches employees as a policy link and a warning. Education should turn it into routes for action by showing the real policy, intake form, reporting channel, responsible group, and escalation process.

By the end, people should know how to:

• check whether a tool and use case are approved;
• request a tool, feature, or use case, or propose an idea;
• report an error, near miss, harmful pattern, privacy concern, or security issue;
• escalate when an AI result conflicts with clinical or professional judgment.

This is governance for everyone.

It gives people defined responsibilities and power. A nurse can report a recurring alert problem, a scheduler can propose a capacity use case, and any employee can verify a chatbot before entering sensitive information.

Joint Commission and Coalition for Health AI guidance supports layered education, voluntary safety-event reporting, and feedback channels for people monitoring a tool.

Build a common foundation, then teach the work

We recommend assessed e-learning as a starting point for an organization-wide program. It creates a common vocabulary, introduces the strategy, and acknowledges that AI is already here. That shared foundation is the larger opportunity we described in Healthcare AI Literacy is the Key to Unlocking AI.

Use scenarios: an ambient note adds an old medication, someone pastes chart content into a public tool, a risk score conflicts with clinical judgment, or an employee proposes a chatbot. Ask what happens next, explain why, and point to the organization’s process.

From that shared foundation, add three layers:

1. Role-based and tool-specific learning when a scribe, predictive model, chatbot, imaging tool, or agent enters a workflow.

2. Practical reinforcement through job aids, office hours, sandbox exercises, team discussions, and lessons from monitoring.

3. Optional deeper learning in prompting, tokens and context windows, model evaluation, architecture, security testing, and agentic systems.

Everybody should leave with something they can use that day, perhaps a safer review step, the route for proposing an idea, or the reason a confident chatbot answer can still be fabricated. And yes, someone will be delighted to learn what a token is.

Measure what the education changes

Completion shows who finished; scenario scores show performance in those scenarios. Follow-up checks, workflow observation, approved-tool data, quality reviews, near misses, and escalation patterns can indicate whether behavior is changing.

Connect those measures to adoption, safety, experience, and patient-care goals. Attribute operational or patient outcomes to education only through a deliberate local evaluation.

AI education gives an organization a chance to tell its people what it is building and how they can participate safely.

If your workforce finished AI education tomorrow, would they understand the mission, the risks, their responsibilities, and what to do next?

If you are working through this design for your organization, let’s compare notes.

Frequently Asked Questions

What should the common foundation cover?

We recommend covering the AI mission, major types of AI, common risks, approved-use boundaries, employee responsibilities, and request, reporting, and escalation routes. Examples and depth should reflect each role.

Who should own and update the curriculum?

The executive sponsor owns the mission, AI governance owns policy and risk content, and the learning team owns delivery and assessment. One curriculum owner should coordinate reviews and update material when tools, policies, or monitoring findings change.

What should be operational before organization-wide education begins?

The organization needs a current AI portfolio, approved-use boundaries, and working request, reporting, incident, and escalation routes. Each route needs a named owner and response expectation.

Is e-learning enough for healthcare AI education?

E-learning can establish the foundation and test scenario decisions. Tool launches, workflow changes, and role responsibilities also need job aids, practice, discussion, and point-of-work reinforcement.

How should a health system measure AI education?

Measure reach, completion, scenario performance, retention, observed behavior, approved-tool use, quality findings, and reporting patterns. Use a defined evaluation design before attributing operational or patient outcomes to education.

References