The Darwin Factor | Survival and Synergy in the AI Ecosystem

Session Focus

The discussion examined healthcare AI as it enters a more competitive and mature phase. The panel looked at which applications are beginning to demonstrate real value, what makes an AI company or capability defensible, how health systems should decide whether to build, buy or partner, and how governance and regulation may shape adoption.

Fargnoli framed the market as a competition between emerging AI companies, EHR vendors, hyperscalers and other established technology platforms. Rather than focusing on ambient documentation, the panel looked for the next areas where healthcare AI could create sustained clinical or operational value.

Where AI Is Showing Early Traction

Ada Glover said some of the strongest early opportunities are in well-defined operational tasks such as intake, referrals, prior authorizations and identifying opportunities in value-based care.

For those applications to scale, organizations need a clearly defined task and business owner, consistent and reliable AI performance, traceability when an output needs review, protection against model drift as underlying technology changes, and a measurable return on the cost of deploying AI.

Her point was that some of the less visible administrative work in healthcare may offer better near-term opportunities than more ambitious applications, because the problem, workflow and economic value are easier to define.

Will Gordon MD said scalable technologies tend to be the ones people begin using naturally. If a product requires extensive training, support or behavior change, that can be a warning sign.

He used an Epic feature that automatically creates discharge summaries as an example. He discovered it without being told and began using it because it immediately saved time. It is low risk because the physician still reviews the output, but it removes work from the clinical workflow.

What Radiology Has Taught Healthcare About AI

Terrence Chen PhD said radiology illustrates an important lesson: the difficult part is not getting an AI algorithm to work, it is getting an AI system to work inside medicine.

Radiology was an early environment for AI because imaging data was already digital and relatively standardized. Yet predictions that AI would replace radiologists have not materialized. Radiologists do more than identify abnormalities in an image. They integrate prior studies, patient history, multiple data types and clinical communication.

Chen argued that model accuracy alone therefore does not demonstrate clinical value. The goal should be for AI capabilities to become integrated so deeply into the clinical workflow that users no longer think of them as a collection of separate AI applications.

He also described how AI can move beyond interpreting existing data. In colonoscopy, AI can help reconstruct the colon in real time and identify areas that have not been adequately visualized. In that case, AI changes how the data is collected and how the procedure is performed.

His broader recommendation was to evaluate the entire workflow and clinical outcome, not simply the model’s technical performance.

Operational AI and Revenue Cycle

Eric Podradchik said revenue cycle has become one of the clearer areas for demonstrating financial return from AI.

Early applications include coding, claims denials and reducing manual work between providers and payers. Because revenue cycle applications can produce a measurable ROI, the financial benefits can help support AI investments in clinical areas where the return is harder to quantify.

He emphasized that the goal is not simply to eliminate jobs. Health systems already have more administrative work than people can reasonably manage. AI can remove repetitive work while helping staff identify opportunities to prevent denials and improve claims processes.

Workflow integration is again critical. A useful capability that exists outside the systems clinicians and staff already use will be much harder to scale.

What Makes an AI Company Defensible

Glover said startups need to think carefully about where established EHR vendors have an inherent advantage.

Documentation, for example, is central to the EHR, which makes it a difficult area for a standalone company to defend over time. More specialized clinical problems or capabilities that operate across multiple systems may offer stronger positions.

Zus Health has focused on interoperability and longitudinal patient data rather than relying on information from a single EHR, which allows the company to work across a broader ecosystem of healthcare data and workflows.

Chen agreed that assembling scattered data is valuable, but argued that one of the strongest positions is being involved in creating the clinical data itself. Imaging equipment, ultrasound, surgical robotics, wearables and other sensors generate data rather than simply consume it.

If a company does not generate data, he said, it needs another source of defensibility, such as proprietary datasets, deep clinical expertise, workflow integration or accumulated clinical evidence. His concise view was that companies closer to the creation of clinical data or the clinical action itself will be more difficult to replace.

How Health Systems Evaluate New AI Vendors

Podradchik said health systems need to understand the roadmaps of their major technology partners, particularly EHR and imaging vendors, before adding another product.

The organization should be able to explain why it is going outside its core platform. The reason might be financial ROI, a clinical outcome or a strategic priority, but there should be governance around the decision rather than individual departments independently adding vendors.

He also stressed that any outside technology eventually needs to integrate into existing workflow. Standalone products may work initially but become difficult to expand across an enterprise.

Gordon distinguished between technologies built around tasks and those willing to take responsibility for outcomes. Established enterprise platforms are well positioned to automate tasks and processes. A newer company may have more opportunity if it is willing to be accountable for an outcome, such as improving a patient’s blood pressure rather than simply supplying another tool.

Patient-Centered Data and Care Outside the EHR

Glover said many newer care models begin working with patients whose records inside their own system are effectively empty. To provide proactive care, they need to assemble information from outside that individual EHR.

She argued that traditional health records tend to be organized around encounters, while patients live most of their lives outside healthcare encounters. That becomes particularly important when multiple organizations are involved in a patient’s care.

Transitions and referrals are still often managed by effectively throwing data over the fence and hoping the next organization receives it. A more patient-centered architecture would allow multiple organizations to collaborate around the same longitudinal record.

The growth of patient-facing AI may complicate that further. Patients may increasingly arrive with information or recommendations generated by their own AI tools, requiring health systems to determine how that information connects with the established medical record.

Build, Buy or Partner

Podradchik described a future in which health systems continue to rely heavily on major platforms such as Epic while selectively bringing in outside capabilities where those platforms are not strong enough.

The decision will not be permanent. A specialized vendor may provide significant value today, but if the core platform later offers the same function more cheaply and more naturally inside the workflow, the health system will have to reassess whether the additional technology is still justified.

Gordon argued that patient-facing technology may be particularly well suited to outside partnerships because EHRs are fundamentally designed around healthcare encounters rather than the patient’s broader life.

Glover cautioned health systems about building too much internally. Creating a prototype has become relatively easy. Maintaining it, adapting it as models change, controlling costs and supporting it reliably at enterprise scale are much harder. Internal development makes the most sense when an organization is deliberately trying to push beyond what the market currently offers.

Governance, Regulation and Risk

Gordon described two realities that exist simultaneously: healthcare is deeply regulated, and existing regulatory frameworks are not well designed for rapidly evolving AI.

Large platforms may be well positioned because they have the resources to manage legal and regulatory complexity. New entrants may have an advantage because they are less constrained by existing products and business models.

He pointed to regulatory sandboxes as one possible approach for testing agentic or autonomous clinical AI in a controlled environment while maintaining patient safety.

Chen said healthcare is too large and complex for one company to provide every capability. He expects major platforms to increasingly serve as orchestration layers, bringing different technologies together while establishing governance, continuous performance monitoring and clear accountability among the platform, model provider and health system.

Glover emphasized continued investment in traceability and proof, while Podradchik noted the tension between AI capabilities advancing rapidly and healthcare’s existing regulatory and risk frameworks.

The panel did not resolve that tension. The shared expectation was that healthcare AI will continue moving quickly, but successful deployment will require clearer accountability, tighter workflow integration and collaboration across an ecosystem rather than reliance on a single vendor.

Key Takeaways

1. Workflow matters more than the algorithm alone. AI has to fit naturally into clinical or operational work to scale.

2. The clearest early wins are often practical. Revenue cycle, documentation and defined administrative tasks offer measurable value.

3. Technical accuracy is not the same as clinical value. Outcomes and the full care process matter.

4. Defensibility will come from more than having an AI model. Proprietary data, clinical expertise, workflow integration, data generation and evidence create stronger positions.

5. Health systems will use a mix of core platforms and specialized vendors. The right answer may change as enterprise platforms expand.

6. Patient-facing AI remains a major opportunity. Traditional EHRs are organized around encounters rather than the patient’s full experience.

7. Governance has to evolve with the technology. Regulation, monitoring, accountability and liability remain unresolved as AI becomes more autonomous.