I recently attended the KLAS Transformation Leadership Forum, which brought together healthcare executives and professional services partners to share what has actually worked in digital and clinical transformation. Not surprisingly, AI was mentioned in quite a few session discussions. As I reflect on my time at the forum, I’ve gathered some of my learning moments to share.
If I had one sentence to capture the mood at the workshop tables, it would be this: efficiency alone does not establish ROI.
One leader from a large healthcare organization, who oversees the AI portfolio of solutions, stated that efficiencies are “felt” across the organization in terms of time saved and lesser administrative burden. But the challenge comes when finance asks what that portfolio returns, the most concrete numbers available are typically token spend and vendor invoices. Translating those efficiencies into measurable financial return, after accounting for the full cost of the solution, is much harder.
Four challenges stood out to me in those discussions.
1. The benefits land in the wrong column
Ambient documentation gives clinicians more face time and shortens after-hours charting. Chart summarization saves clinicians time before a visit. Patient-facing assistants reduce anxiety before an appointment. All real improvements, however their financial impact depends on how those gains change day-to-day operations.
A CMIO, CNO, or CIO builds the case around burnout, retention, and experience. The case reaches the CFO and goes sideways, because retention savings are uncertain in timing and size while implementation costs are certain and arrive now. Most AI tools in production today are one to two years old, still absorbing integration and training costs, and sitting on top of an unchanged cost structure.
2. We have seen consumption pricing before
Traditional enterprise software came with a predictable license. AI arrives with a meter. Leaders described per-message, per-token, and per-transaction pricing, including one conversational tool at roughly 40 cents per message. While this may seem trivial for a single interaction, it becomes material when considering a health system’s monthly or annual volume, which could reach hundreds of thousands of interactions.
Anyone who carried a mobile phone in the late 1990s will recognize the moment. Carriers billed by the minute, and people waited for nights and weekends to make personal calls. Within a decade minutes became a commodity and competition moved to networks and devices. Token costs are falling on a similar curve, which is a real reason for optimism. The complication is that competitive pressure to deploy exists now, and nobody knows when the curve flattens.
Metered pricing also interacts badly with default settings. One organization activated an AI coding capability native to its EHR and learned months later that it had been consuming tens of thousands of tokens a month (a significant expense, especially if some users do not use or unaware of those AI-enabled capabilities in their workflows). This translated into a conversation about the visibility of AI-related activities and costs before governance. Leaders are craving platform-level usage dashboards because they currently lack a single view of where AI runs in their organizations and what it costs. Cost also follows ownership, and at least one organization is moving token spend out of IT and into the department using the capability.
3. A market too unstable for long commitments
Twelve months came up repeatedly as the longest AI commitment several organizations would sign. One leader framed it as a car payment: when a cost can move from three percent of a budget to 30 percent in a year, a long commitment stops looking like a discount. Several said plainly that they do not assume a given AI vendor will exist in its current form next year, and switching mid-workflow carries its own disruption. Longer terms still buy better pricing and justify deeper integration work, so the tension has no clean answer.
4. At-risk contracts and the attribution trap
In discussions about AI, healthcare executives were concerned about implementation timelines, upfront costs, and risks involved in realizing ROI behind those investments. More concretely, attributing the ROI to a single clear initiative is a challenge in itself. A health system runs dozens of concurrent initiatives, and isolating one tool’s contribution to denial rates, throughput, or cash acceleration is close to impossible.
What stood out to me is who was on stage. Most of the strongest transformation stories were presented jointly by a health system executive and a partner . Accountability worked when that partner brought two things at once: deep operating knowledge of the function being changed, whether revenue cycle, access, or clinical operations, and the capability to deploy and support AI inside it. That combination lets both sides commit to outcomes rather than activity, which is the fair alternative version of at-risk that avoids potential year-long arguments about credit.
Setting the clock correctly
The value discussion converged on a few practical points.
- Separate the types of value before writing the business case. Financial return, operating capacity, workforce sustainability, and patient experience require different evidence and persuade different executives. One blended ROI figure satisfies nobody.
- Match the ROI timeline to the solution. An ambient tool with light integration and one workflow can show results in a quarter. An enterprise revenue cycle deployment that touches multiple systems, requires data cleanup, and changes how several teams work will take far longer. A payback period set without accounting for scope and integration depth is a broken promise waiting to happen.
- Use comparisons rather than forecasts. Adopters versus non-adopters on documentation timeliness or patient experience is a simpler conversation with finance than any projection.
- Apply audit-grade rules. In the most rigorous program presented, one-time benefits did not count, only recurring P&L improvement counted, and finance audited every claim. Several of the clearest wins also came from retiring systems rather than buying new ones.
One debate stayed open: whether vendors should arrive with a defensible model of return, or whether the buyer owns that math. The organizations making visible progress decided who owns the value question before signing. The ones still struggling assigned it afterward.
Questions I am still sitting with
If AI becomes inexpensive per interaction, does total spend fall, or does it rise as usage expands into workflows nobody would automate at today’s prices?
Where does competition move once price stops differentiating? Integration depth, clinical validation, liability coverage, something else?
Does department-level cost ownership produce disciplined adoption, or does it quietly shut down experimentation in the departments with the tightest budgets?
I’d welcome your perspective on these questions.