There’s a version of AI deployment that exists in slide decks.
It’s clean. It’s linear. It shows a pilot, a go-live, and a steady upward curve.
Real healthcare operations don’t look like that.
Over the past couple of years, we’ve deployed AI inside a number of healthcare provider organizations and their operational teams. Some rollouts were smoother than expected. Others were humbling. A few taught us lessons the hard way.
Here’s what we’ve learned – not from theory, but from being inside real workflows.
1. The workflow you’re shown isn’t the workflow that runs
Early on, we would map processes based on documented SOPs and leadership walkthroughs. Then we’d go live – and discover the real workflow was different.
Shortcuts existed. Tribal knowledge mattered. Edge cases were handled informally. Teams had built invisible scaffolding to make the system function.
AI doesn’t struggle with clean diagrams. It struggles with reality.
We learned that shadowing isn’t optional. It’s foundational. Until you understand how work actually gets done – including the messy parts – you’re not ready to automate anything.
2. Trust erodes faster than it builds
You can earn confidence over weeks – and lose it in one silent failure.
One missed escalation.
One incorrectly handled edge case.
One patient stuck without visibility.
Healthcare operators have long memories, and for good reason. Their systems are fragile in ways outsiders don’t immediately appreciate.
We learned to design escalation paths before expanding autonomy. We learned to surface uncertainty instead of hiding it. We learned that visible imperfection builds more trust than invisible mistakes.
3. Humans don’t want to be replaced – they want to be protected
There’s a narrative that staff resist AI predominantly because they’re afraid of being replaced.
That hasn’t been our experience.
What people actually fear is being left holding the bag when something breaks.
When teams see that there’s governance, that exceptions are clearly routed, and that someone is accountable for system performance, resistance drops dramatically.
Automation feels threatening when responsibility is unclear.
It feels helpful when guardrails are visible.
4. Volume changes everything
A workflow that works at 200 transactions per day can behave very differently at 5,000.
We’ve seen systems that looked stable in pilot mode behave unpredictably under scale. Monitoring, instrumentation, and release discipline aren’t “enterprise add-ons.” They’re survival mechanisms.
Scaling AI in healthcare isn’t about flipping a bigger switch. It’s about building the plumbing to handle stress.
5. Ownership matters more than intelligence
We’ve worked with very advanced models that underperformed because ownership was unclear.
Who watches the metrics?
Who responds to anomalies?
Who decides when to expand scope?
AI is not self-governing. It requires operational ownership. When that’s defined, even imperfect systems improve over time. When it’s not, even strong systems drift.
6. The hardest part isn’t the model – it’s the change
Technically, many things are possible.
Operationally, not everything is digestible.
We’ve learned that pacing matters. Phased rollout matters. Allowing teams to move from shadow → assist → autonomy at their own speed matters.
If you move too fast, you create backlash.
If you move too slow, you lose momentum.
Finding that balance isn’t a science. It’s a partnership.
The biggest lesson
If there’s one thing I’d say to any healthcare leader considering AI, it’s this: The technology is rarely the limiting factor. The operating model is.
AI doesn’t fix broken workflows. It amplifies them. It doesn’t eliminate responsibility. It redistributes it. And it doesn’t create trust automatically. It has to earn it.
We’ve had deployments that made us proud. We’ve also had moments that forced us to redesign parts of our approach. Both were necessary.
Healthcare AI doesn’t need more confidence.
It needs more humility.
And in our experience, humility is what turns pilots into production.
