In most industries, the AI hesitation sounds like "we're not sure where to start." In healthcare, it sounds like "legal hasn't signed off," "we can't touch patient data," and "what if it hallucinates a dosage?"
Fair. All of it. But that caution has a cost — and it's measured in months of manual work that didn't need to be manual.
We recently ran the healthcare cohort of Opal University — five days of live training and hands-on agent building with marketing and digital leaders from health systems, pharma, medtech, and payer organizations across the US and Europe. One hour of live training followed by one hour of building, every day, for a week. A global cohort working through the same challenges in the same room.
The goal wasn't to talk about AI in the abstract. It was to get people building things that actually work inside the constraints they operate in — which in healthcare, are real, specific, and non-negotiable.
What follows isn't a recap of the curriculum. It's a field report on what happened when healthcare marketers stopped theorizing about AI and started building with it.
We put 50 healthcare marketers in an AI lab for a week. Here's what surfaced.
Pattern 1: Healthcare isn't behind — it's constrained. Those are different things.
Before we taught a single prompt or built a single agent, we spent time on something that almost never appears in AI training: the emotional reality of working in a regulated industry while being told, constantly, that you're not moving fast enough.
One participant put it simply: "We're still just starting out. Our company has a lot of red tape in terms of using AI. That's acted somewhat of a buffer."
That buffer isn't reluctance. It's institutional responsibility. The organizations in this cohort weren't behind on AI because they didn't care. They were behind because AI governance conversations in healthcare are genuinely complex — who owns the data, where it lives, what models can touch it, what the legal exposure is if something goes wrong.
One participant described their internal AI governance meeting in a way that stuck with everyone who heard it:
"They were literally building the plane in the meeting — talking about what model this is, whether data is shared, what is within our risk tolerance. Everybody was throwing out new questions and trying to figure out what's the bare minimum, what is within our risk tolerance."
What changed the energy in the room wasn't pretending those constraints don't exist. It was reframing what's possible within them. You cannot paste clinical call transcripts into a shared training environment. You can build the agent architecture using publicly available content and deploy it on a secure instance later. You cannot eliminate legal review. You can dramatically reduce how many substandard drafts reach legal in the first place.
The most important thing we said on Day 1 was this: AI is surprisingly dependent on your expertise. After you start using it heavily, you realize it desperately needs you — the person who knows what a compliant first draft looks like, who understands which claims need clinical sign-off, who has spent years knowing the difference between approved messaging and promotional overreach. That expertise doesn't become less valuable. It becomes the variable that determines whether AI output is useful or just generic.
Pattern 2: The biggest use cases in healthcare aren't content creation — they're quality assurance
Every other industry cohort we've run has seen participants gravitate toward content generation as their primary use case. Healthcare went somewhere different almost immediately.
The agents that generated the most excitement — and the most genuine organizational interest — were the ones that catch problems before they become problems.
A web operations manager at a healthcare system built a page summary agent. You give it a URL or a section of a site, and it generates a plain-language summary of what that page says, the key takeaways a reader would walk away with, and a set of FAQs. The summary function, in her words, acts as a "mirror" — it shows you how AI systems are interpreting your content. Not how you intended it, but how it actually lands.
For healthcare organizations where the gap between "what we meant to say" and "what AI thinks we said" can affect how patients understand their treatment options, that mirror function is worth a lot. It reveals messaging gaps before patients or search engines encounter them.
Another participant built a brand voice and clinical compliance checker. Feed it a URL, a customer persona, and a target keyword, and it evaluates the page against your brand standards and flags anything that reads as clinical enough to warrant human review. The planned next step: a conditional routing layer that automatically sends flagged content to clinical reviewers while letting non-clinical content move through standard approval. Marketing gets speed. Compliance gets control. Neither team has to manually sort which content is which.
A third participant — an IT director managing over 150 websites across neurology, orthopedics, dentistry, skincare, and gastrointestinal products — was drafting a multi-agent pipeline to solve a problem that was eroding trust with regional business teams:
"When we get the websites, there will be special characters, obviously. Sometimes when the team builds it, they do not know the exact meaning of a specific apostrophe or comma in another language. They miss it. And they don't know what to check. So that's a big time where we spend a lot of effort to correct it and go back and forth."
The solution isn't hiring native speakers in every market. The solution is an agent that validates language-specific characters, punctuation, and medical terminology systematically, across every site, every time.
None of these are flashy. None of them will go viral on LinkedIn. All of them will materially reduce risk and save hours every week for teams that have no margin for error.
Pattern 3: Skills are healthcare's fastest path to measurable value
If you take nothing else from this, take this: the single highest-ROI thing any healthcare organization can do with AI right now costs no money, requires no technical expertise, and takes less than an afternoon.
Turn your existing documentation into AI skills.
Skills — in the context of Opal and AI agents more broadly — are persistent instructions that you teach AI once and it applies everywhere. Your brand tone of voice document, your approved product messaging, your required legal disclaimers, your audience persona definitions. All of that institutional knowledge that currently lives in PDFs that nobody reads becomes an active set of rules the AI applies to every output it touches.
One participant said something that's stayed with me:
"Brand guidelines, tone guides, strategy docs — those things usually go somewhere in a folder to die. Nobody looks at them. Skills basically take those documents and are able to action them on everything you create."
For healthcare organizations specifically, this matters because compliance knowledge is some of the most expensive institutional knowledge there is to develop and some of the easiest to lose when people leave. When a new content creator joins the team, or when you brief an external agency, you currently rely on them absorbing complex regulatory and brand standards over weeks or months. Skills compress that onboarding. The AI has the rules from day one.
The healthcare participants who got the most out of the week were the ones who came in with well-documented standards. That turns out not to be a coincidence.
Pattern 4: AI found problems that humans had been looking at every day and missing
This one surprised even us.
Several participants discovered, in the course of building or testing agents, that those agents surfaced issues with existing content that no human had caught — despite those humans looking at the same content regularly.
A web operations manager found that a page summary agent generated an interpretation of a clinical page that didn't match the intended message at all. The content team thought the page said one thing. The AI thought it said something else. That gap wasn't obvious to anyone reading the page in context, but it would be obvious to a search engine or an AI assistant trying to answer a patient's question.
A participant building a brand voice checker discovered that several pages on their site, including some high-traffic ones, contained language that would trigger their planned clinical review routing — language that had been live for months without anyone flagging it.
This is what happens when you build an agent to audit content systematically rather than spot-check it. Humans are good at reading for meaning. We're less good at reading every page, every time, against a consistent set of criteria. Agents are very good at exactly that.
The agent that a home care provider was building — one that would eventually listen to live care assessment calls and surface relevant clinical questions from a knowledge base in real time — was grounded in the same insight.
"Being able to take that information and train people to better understand people. When a specialist call comes in for dementia, asking the questions that are absolutely relevant to that particular caller at that particular moment in time is something I would dearly love to get to."
That specific use case is blocked right now by data residency requirements. The platform's servers are currently US-based, with no confirmed UK hosting option, which means anything touching clinical call data from UK patients is a hard stop for now. But the thinking behind it — using AI to surface what humans need to say, not to replace what humans say — is the right direction.
What we're still figuring out
The data residency wall is real. The most ambitious use cases in healthcare — clinical call transcript analysis, real-time care guidance, compliance checking of genuine marketing materials — are all blocked or severely constrained by data sensitivity requirements. The workaround we taught: build and validate agent architecture using publicly available content, then deploy with real data on a dedicated secure instance. It works. But it means the learning-to-deployment gap is wider in healthcare than in almost any other industry.
AI agent scans corrupt your own analytics. One participant ran website quality scanning agents and discovered a three-fold jump in organic search traffic in Google Analytics. The agents' page requests were being counted as real user visits. There's currently no native way to prevent this. Before you deploy monitoring agents, coordinate with your analytics team to filter agent-generated traffic. Otherwise your performance reports will reflect activity that has nothing to do with actual patient or provider behavior.
CMS version determines what AI can actually do. This was a surprise to several participants. Older CMS versions (like 12, PaaS) limit agents to front-end crawling — visiting pages one at a time like a user, which times out at scale. Newer versions (CMS 13 and SaaS) enable native graph-based content retrieval, querying content directly from the database. The difference between an agent that times out on 250 pages and an agent that processes them in minutes often comes down to which version of your CMS you're running. If your organization is planning a CMS upgrade, add AI agent capability to the business case. It's a legitimate ROI argument.
Legal still reviews everything. AI can perform a first-pass compliance check. It can reduce how many drafts go back-and-forth before legal sees them. But in healthcare, legal sign-off remains mandatory and cannot be automated away. The bottleneck shifts — it doesn't disappear. That's still worth pursuing. Fewer revision cycles before legal review means faster time-to-publication. But it's worth being honest with stakeholders that this is a reduction in friction, not an elimination of it.
The most complex things require dedicated support. Simple workflows — two or three agents in sequence — are manageable for individual builders. The ambitions that emerged by Day 5 (multi-market site directories, automated pre-release auditing across hundreds of URLs, clinical review routing with conditional logic) are multi-person builds that require ongoing maintenance. The week gave participants the mental model and the starting point. Getting to production scale is a different conversation.
A note on who this is for
We ran the healthcare cohort because the generic version of this conversation — the one full of startup case studies and consumer brand examples — wasn't serving the people in this room.
Healthcare marketers operate in environments where the consequences of getting content wrong are measured differently than in most industries. Where the word "patient" carries a weight that "customer" doesn't. Where the ability to move fast and the obligation to move carefully are in constant tension.
The participants who showed up — some of them on a public holiday, some of them from organizations that hadn't yet officially approved any AI tools — showed up because they wanted to figure out how to be useful and responsible at the same time.
The most impactful outcome of the week wasn't any individual agent. It was a team walking away with a one-page list of agents and skills they wanted to build together and a meeting booked to prioritize them. That's the transition from AI adoption to AI culture. It takes longer in healthcare. But when it happens, it sticks.
If you work in healthcare marketing and want to join the next cohort, sign up here. We run small groups, industry-specific, and spend more time on your specific constraints than any generic AI course will.