Most healthcare AI conversations happen on stages or in pitch decks. The ones that actually teach me something happen over dinner.

We recently hosted a small gathering on the sidelines of the America’s Physician Groups (APG) annual conference in San Diego, bringing together physician group leaders to talk about operations.

No agenda and no sales pitch. Simply executives and senior leaders who are actually running these organizations, comparing notes on what’s working and what isn’t.

Several  themes kept surfacing. And they’re the same ones I hear everywhere.

The Technology Is Spreading. The Operating Model Hasn’t Caught Up.

One group at the table had recently deployed a surgical scheduling tool that lets physicians book cases directly from a mobile app. The AI layer generates productivity analytics visible across the physician group, creating a performance culture that didn’t exist before. Expected case load increase: 8 to 11%

Work that used to require a dedicated analyst now happens automatically.

That’s a real outcome.

But the same organization has the opposite adoption problem elsewhere: ungoverned use.

Staff running their own AI tools with no structure, no security guardrails, no shared methodology. The technology is spreading faster than anyone is governing it.

This pattern comes up constantly. Real results in one pocket. An operating model that hasn’t caught up everywhere else.

What closes that gap isn’t more technology. It’s building the operating model around it: clear ownership, defined workflows, and someone accountable for whether adoption actually holds.

We’ve seen what happens when groups treat this as an IT project rather than an operational one. The tools stick around. The results don’t.

The Right AI. Blocked by a Lack of Interoperability.

Another group at the table has an AI tool handling intake calls. It works. But it isn’t connected to their EMR. The AI sits on one side; the clinical system sits on the other. Any integration requires navigating a lengthy corporate approval chain, adding months of delay to work that could have been done in weeks.

What’s blocking them isn’t a product limitation. It’s structural. EMRs that hinder the free flow of data to third party AI tools aren’t just an inconvenience. They can be a real blocker for what AI can do inside these organizations. Until data moves freely, workflows can’t change meaningfully.

Every organization at the table named some version of this. It isn’t a niche problem. It’s the foundational constraint underneath most of the AI ambition in healthcare right now. Until EHR vendors embrace interoperability and internal IT teams are properly incentivized and resourced, this won’t change.

Trust Is Built Through Involvement, Not Reassurance.

One operations leader at the table, coming from a QA background, had built a structured process where humans actively review AI outputs before they become action. It’s working. Staff feel like participants rather than subjects. Accountability is visible and their comfort with the technology is growing.

The lesson isn’t just about QA process design. It’s about what actually builds trust in AI among frontline teams. It doesn’t come from telling people the system is accurate. It comes from giving them a role in verifying it.

This matters more than most vendors want to admit. The practices where AI adoption stalls aren’t usually dealing with a bad product. They’re dealing with staff who feel like something is being done to them rather than with them. Getting frontline teams involved early, giving them a real role in validating outputs, and making adoption a shared process rather than a top-down mandate, is what separates a pilot that works from a program that scales. We’ve seen this pattern repeatedly.

At Graybill Medical Group, where a two-year partnership produced a 50% reduction in front-office costs and a 24% reduction in no-shows, the operational shift wasn’t just technical. It required rebuilding workflows across nine locations with the people doing the work inside the process, not watching it from the outside.

No One Has Cracked Repeatable Upskilling. And Everyone Knows It.

One leader said it plainly: their organization needs to upskill its team, and no one at the table had found a methodology that actually scales. Not a complaint. Just an honest observation that landed with the whole group.

This means the tools are moving faster than the organizational capability to absorb them. Groups that are making headway tend to share one thing: they invest in their operators, not just their software. They identify the people who understand both the clinical context and the workflow logic, and they build around them. It’s not a scalable system yet. But it’s a starting point.

The Gap Is Never Where You Think It Is.

The organizations getting AI right aren’t the ones with the best tools. They’re the ones treating adoption as an operational discipline, not an implementation task. The difference between a pilot that works and a program that scales isn’t the product.

If your organization is somewhere in this gap, between the pilot that worked and the program that hasn’t, come grab a seat at the dinner table and  join the conversation

*Frederik Mueller is co-founder and CEO of Third Way Health. Frederik has dedicated his career to solving healthcare’s biggest operations and technology challenges.

Frequently Asked Questions

How can physician groups improve front office operations with AI?
The highest-impact starting point is usually the highest-volume, most standardized workflow in your front office, typically inbound scheduling or patient intake. AI handles the volume; the  team handles the exceptions. The key is not the tool itself but how it’s integrated into workflows and whether the staff are involved in the process from the start, not just handed a new system.

What does it actually cost to outsource front office operations for a medical practice?
The more useful question is what your current model is actually costing you. Most physician groups significantly undercount their true front-office spend because labor costs, overhead, turnover, and lost revenue from unanswered calls are tracked separately. When groups consolidate those figures, outsourcing to an AI-enabled partner typically delivers 30 to 50 percent cost reduction while improving access metrics.

Why do AI implementations fail in healthcare practices?
The most common failure mode isn’t a bad product. It’s deploying AI onto a workflow that was never redesigned to support it, and doing so without adequate staff involvement or ongoing operational support. AI surfaces process gaps that were previously invisible. Without someone accountable for navigating those gaps after go-live, adoption stalls.

What’s the difference between AI tools and AI-enabled operations in healthcare?
AI tools automate a task. AI-enabled operations redesign the workflow around the automation and ensure the human behavior required to sustain it actually changes. Most healthcare organizations have the former. The ones seeing durable results have built the latter.