AI Exposure
By Jeri Garner | Principal Advisor, LEV Advisory Group
There is a lot to be excited about with AI in healthcare. I use it, I see what it can do, and some of it is pretty amazing: it can summarize information, help with documentation, automate patient communication, prioritize messages, and take on administrative work that has frustrated patients and clinicians for years. I absolutely believe AI has the potential to make healthcare better.
But after more than 20 years implementing healthcare technology and working with clinical and operational teams, there is one question I keep coming back to: did we actually remove the work, or did we just move it? Because healthcare has done this before.
I Learned This Lesson Long Before AI
Early in my career, I was involved in a large EHR implementation. We created a series of tasks and alerts that would route automatically to the nurse, and on paper it looked great. Every task had an owner. Everything was accounted for.
For one patient, the workflow worked beautifully. The problem was, the nurse didn't have one patient. She had six. Multiply those tasks and alerts across six patients, throughout an entire shift, on top of everything else she was responsible for, and suddenly our beautifully designed workflow wasn't so beautiful.
That experience taught me something I've carried through every technology implementation since: you can't evaluate technology one transaction at a time. You have to evaluate it at the scale of the person who has to live with it. That matters even more with AI.
AI Can Create Demand Faster Than People Can Absorb It
Think about patient engagement. AI can make it easier for patients to ask questions, report symptoms, request appointments, or follow up on results, and that's a good thing. But what happens on the other end? If AI enables hundreds of additional patient interactions over a weekend, what happens Monday morning? Do they become messages, tasks, reviews, another queue someone has to manage?
If so, we may have made the patient's experience easier while making the care team's job harder. We didn't eliminate the bottleneck. We moved it. AI doesn't get tired. The care team does.
That's why I don't want to see clinicians turned into AI babysitters—reviewing everything AI produces, correcting bad information, managing false positives, and still doing everything they were doing before. The best AI should take work off people's plates, not quietly put more work on someone else's.
The Exceptions Are Where It Gets Interesting
Imagine an AI agent helping manage test-result follow-up. An abnormal result comes back. AI recognizes it, messages the patient, creates a follow-up task, and routes it to the appropriate team. Perfect. Then the patient responds:
"I'm scared. I've been having chest pain since yesterday. What should I do?"
Now what? Does AI recognize this is no longer routine? Who does it escalate to? What if that person isn't working? How quickly does someone respond? Most importantly, how do we know someone actually acted on it?
The AI could have done everything it was designed to do, and the patient could still be sitting at home waiting. Technically, the workflow worked. Operationally, it failed. That distinction becomes even more important as we give AI greater autonomy.
"Human in the Loop" Isn't Enough
We hear this phrase constantly. Of course we need humans in the loop, but anyone who has worked in healthcare knows there's another risk: the danger isn't always that there won't be a human in the loop—it's that five people will think somebody else is. A task gets routed to a shared queue. Several people have access. Someone assumes someone else saw it. Nobody actually owns it. AI doesn't eliminate that problem. It can make it happen faster and at a much larger scale.
This is the distinction I keep coming back to: human in the loop is a design choice. Human on the hook is an operational commitment. A workflow can have a human in the loop at every step and still have no one on the hook when it matters. The more autonomy we give AI, the less ambiguity we can tolerate in who's accountable. We can delegate a task to AI. We cannot delegate accountability to it.
Five Questions to Ask Before Scaling AI
Before expanding an AI-enabled workflow, I think every healthcare organization should be able to answer five questions:
Can AI see what it needs to see? Is the information complete, current, accurate, and trustworthy?
Does AI know what it's allowed to do? Where can it act independently, and where does human judgment remain?
Who owns what happens next? Not just who receives the task—who owns the outcome?
Did we remove work, or just move it? Follow the impact all the way through the workflow.
Did the loop actually close? Don't stop measuring when AI completes its part. Did the patient actually get what they needed?
Those aren't really technology questions. They're operational questions.
Before You Scale AI, Understand What It Will Scale
AI is incredibly good at scale. Give it a great process, and it can scale a great process. Give it a broken one, and it can scale that too.
That's why healthcare AI readiness is about more than the technology. It's about the operating environment around it: workflow, data, accountability, and people. At LEV Advisory Group, that's where we work. Our Signal Scan™ Healthcare AI Readiness & Operational Assessment helps uncover the friction, ownership gaps, downstream burden, data risks, and operational weaknesses that can keep great technology from delivering great results.
Because before we ask: "What else can AI do?" I think we need to ask: "What happens to the rest of the system when it does?"
AI isn't the disruption. It's the exposure. What it exposes is where the work begins.
Comments