Should Health Systems Buy AI Agents One Use Case at a Time?
The best day my old company ever had was also the day I should've been most worried.
We were a DTC brand, and I watched the Stripe dashboard tick past $1M in a single day, and I felt like a genius for roughly six hours, until the returns and the support tickets started telling me what the dashboard couldn't. The product was bad. Scale doesn't fix problems. It hides them.
Then it multiplies them.
I keep thinking about that dashboard when I read how health systems are buying AI agents right now.
Should health systems buy AI agents per use case or as a platform?
As a platform, or at least with a platform mindset. Buying one agent per use case feels safe and fast, but you end up with a dozen tools that don't share context, governance, or accountability.
Peter McCaffrey, chief digital and AI officer at UTMB, told Healthcare Innovation that his system now runs 22 agents across triage, lung nodule navigation, prior authorization and more, all on one platform. His framing stuck with me. It's not a tool for this and a tool for that. It's describing the workflow properly and placing intelligence inside it, alongside people.
Why does buying agents one at a time break down?
Because every new agent brings its own governance review, its own integration, and its own blind spots. At small numbers that's fine. At scale nobody can see the whole picture.
Mayo Clinic has roughly 450 AI solutions in its pipeline and 128 in clinical practice, according to Micky Tripathi's interview on Mayo's governance overhaul. Their committees couldn't keep up with the volume, so they rebuilt the structure around one accountable executive reporting to the CEO.
That's Mayo. With Mayo's resources.
Now picture a regional system with a fraction of the staff and the same vendor pitches landing in the inbox every single week.
Are AI agents tools or part of the workforce?
Treat them like workforce. They take instructions, run workflows, and make mistakes that compound, which is exactly why they need managers, reviews, and clear owners.
McCaffrey said agents become "a segment of the workforce" at some point. I agree, and I'd push it further. My daughter is ten, and one of the things I tell her (usually while she's bossing me around, fair enough) is that I work for her, not the other way around. Good managers work for their teams. Good AI agents work for the patient and for the nurse who picks up the escalation.
If you can't name who an agent works for, it shouldn't be live.
Who should own an AI agent inside a health system?
The service line owner, not IT alone. The person who runs lung nodule navigation or discharge follow up knows what good looks like, so they're the one who should shape, champion, and answer for the agent.
McCaffrey was blunt: if a use case has to be evangelized from the top down, that's not a great sign. Intermountain shows the other half. They're aiming for AI to handle 20 to 30% of roughly 260,000 monthly contact center calls, and their consumer experience lead credited board level KPIs for getting every function to buy in. The top sets the goal. The service line shapes the work.
What should a patient engagement platform actually give you?
One patient memory, one escalation path, and one place to see what every agent did. If your follow up agent doesn't know what your scheduling agent just told the patient, the patient notices before you do.
That's why we built HANA as open source and self hosted infrastructure, with no OpenAI dependency. Health systems should own the layer their patients talk to. Across more than a million patient interactions with zero critical adverse events, what's kept it safe isn't a clever model. It's one consistent layer that knows every patient and knows when to hand off to a human. We publish how we measure that on our research page.
How do you measure whether an agent platform is working?
Measure outcomes you already count, then add the ones patients feel. Scheduling and billing numbers are easy to verify. Engagement over months is harder, and it's where most programs quietly die.
The industry baseline for ongoing patient engagement sits around 15 to 20%. We run at 85% weekly, and we see a 31:1 clinic ROI across deployments. Honestly, the first number matters more to me than the second. A platform patients stop answering is just an expensive Stripe dashboard. Looks great. Tells you nothing.
Key Takeaways
Buying AI agents one use case at a time feels fast, and it is, right up until you have twenty of them and nobody can tell you what they're all doing. Treat agents like a segment of your workforce: give each one an owner in the service line, someone who reviews its work, and a clear answer to who it works for. Put them on shared infrastructure so patient context and escalation live in one place. And measure engagement over months, not just the admin wins from week one. Scale will hide your problems if you let it. If you want to know why we care so much about this, our story and team are here.
FAQ
What is an AI agent platform in healthcare?
It's shared infrastructure where multiple AI agents run with common patient context, governance, and monitoring. Instead of separate tools per use case, each agent plugs into the same workflow layer and escalation paths.
How many AI agents do large health systems run today?
It varies a lot. UTMB runs 22 agents on one platform, and Mayo Clinic reports around 450 AI solutions in its pipeline with 128 in clinical use.
Who should govern AI agents in a hospital?
A central executive should set policy, and the service line owner should run each agent day to day. Clinical leadership, not IT alone, should decide whether an agent is actually delivering benefit.
If you're sitting on a stack of agent pitches and trying to figure out which ones deserve a place in your workforce, book a call with me. I'll tell you what I'd do, even if the answer isn't us.