Seven lessons from healthcare leaders working through AI adoption, education, governance, and scale.
AI is moving quickly inside healthcare organizations. But the harder work isn’t choosing the technology. It’s preparing people to use it well.
Health systems are rolling out ambient documentation, copilots, and employee-built automations, often faster than their literacy programs, governance models, and training can keep up.
Amplifire recently brought together AI, clinical, and workforce leaders from health systems nationwide for a candid, anonymized roundtable on what’s working, what isn’t, and what they’re still figuring out.
The real conversation wasn’t about tools. It was about people: how to build confidence without overconfidence, how much training is enough, and how to govern AI without smothering useful experimentation.
Seven ideas kept coming up.
1. The first AI wins are not really about AI. They are about work.
One of the clearest early successes across the group was ambient documentation.
Participants described physicians spending less time documenting and more time with patients. In some cases, the impact was significant enough that clinicians became strong advocates for the technology.
But the same tool did not work equally well for everyone. Adoption varied by clinician, specialty, setting, and workflow. Some physicians found ambient documentation transformative. Others found little value in it. Organizations were also learning that an approach designed around a physician workflow could not simply be transferred to nursing.
Nurses work differently. They interact with patients differently, and they document differently. One participant described the challenge as learning to “care out loud” so an ambient system can capture information that may previously have been documented another way.
That distinction matters because it is easy to start an AI strategy with the technology: Where can we use AI?
A more useful question may be: Where is work unnecessarily difficult today?
Where are people spending hours on repetitive tasks? Where is documentation getting in the way of patient care? Where are people manually gathering, summarizing, or moving information?
The strongest early AI use cases discussed in the roundtable had something in common: the value was obvious to the person doing the work. That may be one of the simplest tests of a good AI use case.
If you have to spend a lot of time convincing people that the technology is valuable, the problem may not be adoption. It may be the use case.
2. Adoption travels through people faster than it travels through training.
One organization initially struggled with adoption of an ambient documentation tool. Then an influential physician began telling colleagues they should be using it, and the response changed.
Other participants described similar experiences with physician champions, informaticists, super users, and colleagues sharing successful AI use cases with one another. That peer influence matters.
Someone can sit through a presentation about how AI might save time and remain skeptical. Then a colleague says, “I used this to do something that normally takes me three hours, and I finished it in 20 minutes.”
Suddenly, the technology is worth exploring.
Participants described prompt-a-thons, mini-hackathons, AI jams, town halls, office hours, communities of practice, and informal team sharing as important parts of their enablement strategies.
The lesson is not that formal learning is unnecessary. It is that formal learning and peer-to-peer adoption do different jobs.
Formal learning establishes the foundation. Peer evidence creates pull.
Organizations need both.
A short AI foundations course might teach employees what they need to know about privacy, risk, approved tools, and responsible use. But the motivation to actually incorporate AI into daily work may come from watching someone they trust use it successfully.
That changes the role of workforce enablement. The goal is not simply to create more AI training. It is to create an environment where employees can learn from the organization and from one another.
3. AI literacy needs to teach judgment, not just prompting.
Prompting gets a lot of attention in AI education. It is useful, but it is nowhere near enough.
When the group discussed what belongs in a foundational AI curriculum, the list became much broader: different kinds of AI, approved and unapproved tools, privacy, PHI and PII, ethical and responsible use, human accountability, hallucinations, probabilistic outputs, prompting, appropriate trust, and knowing when to verify, edit, or escalate an AI-generated response.
One of the most interesting parts of the conversation centered on something even harder to teach: AI can be wrong and sound very convincing.
Poor information packaged extremely well can look more credible than good information communicated poorly. Employees need to understand that fluency is not the same thing as accuracy.
But there is another side to that problem. Humans are not perfect judges either. People have knowledge gaps. They get tired. They make mistakes. And sometimes they are very confident about something they have completely wrong.
At Amplifire, we call that Confidently Held Misinformation™. It is something we have spent years studying because uncertainty and misinformation create very different risks. Someone who knows they are uncertain is more likely to stop, check, or ask for help. Someone who is confident and wrong is much more likely to act.
AI introduces another participant into that equation. Now the employee has to evaluate not only what they know, but what the AI appears to know.
That makes judgment one of the most important skills in AI workforce readiness.
AI literacy is not knowing how to get an answer from AI. It is knowing what to do with the answer.
4. “Human in the loop” is not enough of a workforce strategy.
“Keep a human in the loop” has become one of the most common answers to AI risk. It makes sense, but it also becomes harder as AI gets embedded into more workflows.
Participants raised a practical question: What happens when there are simply too many AI outputs for a person to meaningfully review every one?
Telling employees to “check everything” sounds responsible. At scale, it may not be realistic.
The conversation began moving toward a more useful model: oversight proportional to risk.
A low-risk AI use case that summarizes recurring internal reports probably should not have the same governance requirements as an AI system informing a clinical decision. The same should be true of human oversight.
Some applications may require extensive validation. Others may need periodic monitoring. Some may be safe enough for routine use within clearly defined guardrails.
Participants also raised another uncomfortable but important point: AI is frequently judged against perfection. Humans generally are not.
Healthcare already has established approaches for understanding human error, acceptable performance, risk, and accountability. If an AI system is being evaluated against a standard no human consistently meets, organizations may not be asking the right question.
The goal is not to lower the standard for AI. It is to define the standard clearly.
Instead of simply asking, Is there a human in the loop?, organizations may need to ask:
What human judgment is required, at what point in the process, and proportional to what level of risk?
That is a harder question, but a much more useful one.
5. AI can expose problems that AI cannot fix.
There is a natural temptation to look at a slow process and automate it. The roundtable surfaced an important warning:
Automating a broken workflow only scales the problem.
Participants working on citizen-development programs were already encountering this. Employees would identify a process they wanted to improve with AI, only to discover that the underlying process itself needed work.
AI can magnify messy workflows. The same is true of data and institutional knowledge.
One participant gave a simple example. Many longtime employees have years of documents sitting in drives and folders. Some are current. Some are outdated. An AI system with access to those documents can retrieve both.
An experienced employee may recognize immediately that an old policy or document is no longer valid. A new employee may not.
That makes knowledge management part of AI readiness in a way organizations may not have anticipated.
Before asking what AI can do with organizational knowledge, organizations may need to ask harder questions about the knowledge itself: Is it accurate? Is it current? Who owns it? What should be retained? What should no longer be available?
The same applies to organizational data and processes.
AI workforce enablement therefore cannot live entirely inside an AI training program. It increasingly touches process improvement, data stewardship, and knowledge management.
Those may not be the most exciting parts of an AI strategy, but they may be some of the most important.
6. Citizen development changes the governance problem.
Some of the most interesting AI applications may never originate in IT. They are already being built by employees.
Roundtable participants described relatively simple automations that gather information from multiple sources, summarize recurring reports, or turn data into something easier to use. In some cases, work that previously took hours could be dramatically reduced.
That creates an enormous opportunity. It also creates a new question:
When does something an employee built for themselves become something the enterprise owns?
Imagine someone in finance creates an AI workflow that saves several hours every week. Then their team starts using it. Then another department wants it.
At what point does a useful personal automation become an enterprise application?
Participants discussed the need for review involving architecture, security, subject-matter experts, and finance before locally developed solutions scale broadly.
A useful way to think about that progression is:
Experiment → Validate → Govern → Scale → Own
Organizations need to make experimentation possible without pretending every experiment is ready for enterprise use. They also need a path for the good ideas to graduate.
Otherwise, one of two things happens: governance becomes so restrictive that employees stop experimenting, or successful experiments quietly become business-critical tools without the organization realizing it.
Neither is a particularly good outcome.
7. The most durable AI curriculum may be the one that talks least about today’s tools.
There was an interesting tension throughout the conversation. People need help using today’s tools, but today’s tools keep changing.
Features appear. Models change. Vendors change pricing. Enterprise capabilities change. Employees may hear about a new capability publicly before their own organization has had time to evaluate it.
Traditional training development cycles are not designed for that pace. Several participants described the challenge of creating training that can become outdated before it is fully approved and published.
That suggests AI workforce enablement may need three different layers.
The durable layer teaches principles that should survive changes in technology: judgment, privacy, responsible use, appropriate trust, verification, accountability, and core AI concepts.
The role layer helps people understand how AI applies to their actual work. A nurse, physician, analyst, instructional designer, finance leader, and IT professional do not need exactly the same AI education.
The dynamic layer handles the things that change quickly: specific tools, features, organizational policies, workflows, and current use cases.
That final layer may increasingly happen at the point of need rather than inside a traditional course. The group discussed contextual education, short clinician-led videos, communities, self-service resources, and learning that appears when an employee actually needs it.
This also addresses another challenge raised repeatedly during the roundtable: the workforce is not starting from the same place.
Some employees have barely touched generative AI. Others are already building agents and automations. Giving both groups the same 45-minute AI course does not make much sense.
The future of AI workforce enablement will need to become more personal, not simply more comprehensive.
Take this back to your organization
No one in the roundtable suggested they had all of this figured out. In fact, one of the most reassuring parts of the conversation was how often leaders from very different organizations were wrestling with the same questions.
The technology is moving quickly, and organizations are learning while they deploy it.
So rather than leaving this conversation with another AI checklist, here are seven questions worth bringing back to your team:
- Where is AI creating measurable value for our workforce today, and what specifically made those use cases work?
- What does every employee need to understand about AI to make good decisions, regardless of which tool they use?
- Where is human review actually necessary, and how should that requirement change based on risk?
- Are our processes, data, and institutional knowledge ready to be amplified by AI?
- How will an employee-built AI solution move from an individual experiment to an enterprise capability?
- Which parts of our AI education should be durable, which should be role-specific, and which should happen at the point of need?
- How will we know that people can apply what we have taught them, not simply that they completed the training?
That last question may become increasingly important.
AI workforce readiness is easy to frame as a technology problem. It is tempting to measure licenses, adoption, prompts, training completions, or the number of use cases deployed. Those things tell us something, but they do not tell us whether someone knows when to trust an AI-generated answer.
They do not tell us whether employees recognize when something is wrong, understand the limits of the tool, or know how to use it responsibly in the context of their role. And they certainly do not tell us whether someone is confident in something they should be questioning.
Those are human questions.
And as AI becomes more capable, getting those questions right may matter even more.