Join the conversation
.jpg)
Get ahead of the curve and learn from leading care providers and companies such as Cityblock Health, One Medical, Oak Street Health and more
A conversation with Ed Lee (Nabla), Rik Renard (Sword), and Thomas Vande Casteele (Awell) on what actually works - and what doesn't - in scaling Healthcare AI. Moderated by Andrew Schutzbank. Watch the full recording here.
In the 1950s, people hired "knocker-uppers" who came to their houses in the morning with long sticks to tap on windows to wake them up. When alarm clocks were invented, we didn't automate the window-knocking. We redesigned the solution.
This conversation brought together three companies working across different AI patterns: Nabla (ambient scribes), Sword (voice AI), and Awell (care orchestration). Despite different offerings, they shared many similar lessons about what works and what doesn't in scaling AI.
About halfway through the conversation, Rik points at the emperor wearing no clothes:
"Are we building a window knocking machine, or are we building an alarm clock? I talk to an executive, I talk to people on the floor, and we shadow their workflow, and then they just want to automate the current workflow in itself, but in that case, we are building those window knockers... whenever you shadow a healthcare workflow it is extremely broken, and if you think about it from first principles it shouldn't look like that."
His challenge:
"Building the window knocker is so much easier than building the alarm clock, because it requires so much thinking and so much cognitive load that I don't have, most people don't have, and most people don't have the gut to convince the executives... Do you just want to have a shiny object, or do you want to go from first principles and totally rethink your processes?"
The translation: doing things properly (= building alarm clocks) is an important barrier to scaling AI, potentially leading to incrementalism and losing the transformational power of the technology.
Thomas from Awell on the decision framework:
“There are organizations out there that understand that AI is the future […] where top-down leadership is saying, “I want to start solving these problems with AI, even if AI is not there yet, I just want it to be AI first.”
This mindset and cultural shift is important is crucial. If you want to be transformational, a top-down buy-in or mandate is paramount - even if it’s not always with the right framing.
On a more tactical level, there’s plenty of stuff AI is not ready for today. According to Thomas:
"When it is really important that something happens deterministically from a medical-legal perspective, you likely want to have it not AI do it for now, but you want a deterministic if-this-then-that logic flow. On the other hand if it is very complex to model in a flowchart and it is not that important from a medical-legal perspective what the exact sequence of events is ... then it is a very good fit for AI."
The killer use cases for AI today: patient outreach, eligibility checking, post-discharge coordination, appointment scheduling - anywhere complexity is high but medical-legal risk is manageable with a simple human-in-the-loop.
Pilots run in silos. Scale cannot happen unless key integration points are tackled. Ed from Nabla learned three important lessons the hard way:
1. Integration Points Matter
In the early days, Nabla expected clinicians to copy/paste notes from their product into EHRs. Adoption significantly increased when they added direct EHR integration (Epic, Cerner, Athena, NextGen). This is obvious as a principle but not obvious to execute on.
2. Timing Is Critical
Again, early on, Nabla’s product took some time to spit out the (full) note at the end of the conversation. Ed recalls:
"If you have a note that takes minutes to create, the patient's gone already by the time you're done... you gotta have a very rapidly created note within seconds."
Andrew reinforces the point:
"How powerful that post-visit synthetic moment is... the amount of things you remember right when you have just 2 seconds to think after the person is done is insanely valuable."
Timing is not an obvious principle as a driver of implementation; this can only be observed by shadowing actual users or relentlessly collecting feedback. Nabla allows all its users to provide feedback right from the product.
3. Don't Break What Works
Nabla built clinical nudges as push notifications during patient visits. But their core value was reducing distractions.
"When the premise of the product is to allow clinicians to be removed from those distractions, having push notifications on their phone was kind of the opposite of what we wanted to do."
The fix: nudges now appear after the visit, before the patient leaves; not during the conversation.
Ed's principle:
"When you try to add something on top of or outside of a workflow that is sort of natural, that's when you have challenges getting true adoption."
Rik's counterintuitive advice:
"Don't expect AI to ever be as good as a human being at all, at this point, and you will need to accept the downsides of that."
Andrew's response:
"People expect perfection out of AI, which is a mistake, because we don't expect it out of people... don't even expect human-level performance, just expect faster, more available lower-level performance. It's a good trade-off."
The trade: 80% accuracy at 10x speed and 24/7 availability beats 95% accuracy during business hours only - for most coordination tasks (and non-medical-legal critical aspects cfr. comments above)
Orchestrating AI across many use cases and organizations, Thomas sees the pattern of what breaks AI: it’s human error. He gave two examples:
As long as we rely on humans to do tedious and repetitive tasks, they will go wrong some part of the time. Shift humans to oversight, not execution.
Nabla: Over 25% reduction in clinician burnout in studies at academic medical centers like University of Iowa. Clinical documentation in seconds, not minutes.
Sword: Rik wasn’t explicit about specific results but Sword uses outcome-based pricing: only charge when the end-to-end workflow executes successfully. "We're going to save you $1,000 per successfully executed workflow, and we're going to capture $100 for each one."
Awell: one customer was able to save an equivalent of 20 FTE through an AI-enabled workflow. Although productivity metrics like utilization rate, throughput, patient to care team ratio are still the most common metrics, organizations are also hitting clinical (readmission) and business (e.g. outreach rates) outcomes.
Watch the full conversation here for deeper insights on governance, clinician adoption, and lessons from implementation failures.
Further reading:
