April 12, 2026
Harvard Business School dropped a number last week: $109 billion. That's what enterprises spent on AI this year. You know what they got for it? Almost nothing.
I run a 2-day Healthcare AI Workflow Governance Audit. We find 2-3 workflow changes that save real money. Every single time, we uncover the same three questions that nobody asked before spending millions. These aren't complex questions. They take 30 minutes to answer. But somehow everyone skips them.
## "What specific workflow does this replace, and what happens when it's wrong?"
You've sat in those meetings. Someone says "let's add AI to our discharge planning." Everyone nods. Six months later you've got a model that suggests discharge dates. But nobody mapped the actual workflow first.
Who reviews those suggestions? What happens when the AI says discharge Tuesday but the patient needs dialysis Wednesday? Where does the override live? If you can't answer these questions, you built a very expensive random number generator.
The workflow comes first. Not the AI. You need to know every step, every handoff, every place a human makes a judgment call. Then you need exception paths. Because AI will find new ways to be wrong that you never imagined.
## "Who owns this when it breaks at 2 AM?"
Your AI will break at 2 AM. Not if. When.
And when it does, someone needs to explain why it recommended IV fluids for a dehydrated patient who's already in heart failure. That person can't just restart the service. They need to understand what the model saw, why it made that call, how to prevent it next time.
AI doesn't eliminate operational responsibility. It redistributes it. Now your on-call staff needs to debug neural networks instead of SQL queries. If they can't trace a decision back through the model, you've created a black box. Regulators hate black boxes. So do malpractice attorneys.
## "Can you explain this to your board in one sentence?"
Try this exercise. Explain your AI system to someone who doesn't work in tech. One sentence. No jargon.
"We use machine learning to optimize patient flow" doesn't count. That's word salad.
"Our system reads admission notes and suggests which patients need case management review within 24 hours" - that works. Specific input, specific output, specific timeline.
If you can't do this, your governance isn't ready. The gap between "we deployed AI" and "we deployed AI we can explain and defend" - that's where careers end. Where lawsuits start. Where you discover that nobody actually knows what the system does, they just know it has 98% accuracy on the test set.
## Why These Questions Matter
These aren't strategy questions. They're operator questions. They're the difference between a system that runs in production for years and one that gets quietly turned off after burning through the implementation budget.
I know because I built and ran healthcare systems for 12 years. Real systems with real patients and real liability. The technology was never the hard part. The hard part was making sure humans could operate it, fix it, explain it when something went sideways.
You can implement AI without asking these questions. Lots of places do. They're part of that $109 billion with nothing to show for it. Their AI works great in demos. Falls apart in production. Because they treated it like a technology problem instead of an operations problem.
## What Actually Works
Start with one workflow. One specific problem. Map every step, every decision point, every failure mode. Assign an owner who understands both the clinical workflow and enough about the model to debug issues. Write that one-sentence explanation.
Then implement. Then expand.
The Healthcare AI Workflow Audit I run focuses on exactly this. We spend two days going through your workflows, finding the 2-3 places where AI could save real money - if implemented right. Not a strategy deck. Not hand-waving about transformation. A working blueprint with names attached to responsibilities.
Most organizations need help with the governance piece, not the technology. The models work. The deployment process doesn't. These three questions fix that. Ask them before you spend another dollar on AI. Your CFO will thank you. Your compliance team will thank you.
Your 2 AM on-call staff will thank you most of all.