What Health Care Leaders Have Learned From Deploying AI

There is a version of every artificial intelligence (AI) conversation in health care that stays safely at the level of potential—what the technology might do, what the models are becoming capable of, what the future of the patient journey could look like.

At Asembia’s AXS26 Summit, Ankit Jain, CEO and cofounder of Infinitus Systems, and Santiago Abraham, chief information officer for Optum Rx, opened their session by explicitly refusing to have that conversation.¹ Both speakers came prepared with production data, deployment lessons, and a frank accounting of where AI has and has not delivered. The result was one of the more grounded discussions of AI in health care to emerge from this conference cycle, and one that will be useful to any organization still deciding how to move from experimentation to execution.

Adoption Is Lagging, and the Stakes Are High

Jain opened with a set of statistics designed to locate the audience within the broader AI adoption landscape. Sixty-nine percent of health care and life sciences executives say AI is their top priority in 2026. Yet only 8.3% of those same enterprises have AI operating in production today compared with more than 90% in professional services and 23% in information technology. Sixty percent of health care AI investment to date has concentrated on administrative functions. Among pharmacists specifically, 82% report familiarity with AI, but only 39% have actually used it.

The governance picture is equally uneven. More than 250 AI-related health care bills have been introduced across 34 states. This gap between adoption intent and production reality is not unique. According to a recent industry analysis, AI has reached 75% adoption across health systems—but only 18% of those organizations have formal governance structures in place, with most lacking data policies and staff with the skills to evaluate what they have deployed.² The pattern, Jain and Abraham argued, is not a reason to slow down. It is a reason to be more deliberate.¹

“It’s still very much the Wild West in terms of what’s getting deployed, who’s governing and how,” said Jain. “We’ll see who ends up being the governor of AI over time—whether it’s commercial enterprises, state governments, or federal governments.”

What Is Actually Working: Production Use Cases From Optum Rx

Abraham walked through 3 categories of AI deployment that have demonstrated measurable value at Optum Rx. The first is prior authorization. AI has been deployed to support clinicians through the authorization process, reducing friction and improving turnaround times in a workflow that has historically been a significant source of delay between diagnosis and therapy initiation.

“The clinician is making the decision on the care decisions but really supporting that clinician to really be able to ultimately move, move through that expedited, consistent passion is super important, to really get patients to therapy faster,” said Abraham.

The second is pharmacy SIG—the prescribing directions on medication labels. Inconsistencies in SIG coding represent a significant patient safety risk in high-volume pharmacy operations. AI has been deployed to apply consistency across this process, reducing errors while keeping a pharmacist at the center of clinical review. The third is consumer self-service. Abraham noted a near-universal skepticism toward automated phone systems, including his own, and acknowledged that prior generations of AI-powered self-service were largely inadequate. The most recent deployment generation, however, has produced improvements in the underlying model capabilities.

“I will tell you the newest generation of this stuff is materially better,” said Abraham. “We’ve seen self-service essentially spiking pretty significantly. We’ve reduced the volume of live operator calls by close to 20% just in this past year, leveraging this new technology.”

On the technology side, Jain offered a figure that illustrated how rapidly the development landscape itself is shifting: more than 95% of Infinitus’s code today is written by AI. That was not the case 3 months ago. He added a critical nuance for organizations managing legacy technology environments: AI produces the largest productivity gains when building from a clean slate, because models understand AI-generated code better than human-written code. Retrofitting existing enterprise systems with AI, while still valuable, produces more modest efficiency gains.

Where the Hype Has Outrun the Reality

Abraham was direct about where AI has not yet delivered in patient-facing specialty contexts. Full automation of patient and member-facing interactions remains in the category of “proceeding with caution”—not because the technology is incapable of improvement, but because the stakes of error are categorically different from those in other industries.

“If the shopper agent makes the mistake of recommending the wrong thing, you added the wrong thing to the cart. If we make a mistake in health care with AI agents, there are big stakes at hand,” said Abraham

Jain identified a second failure mode: AI investment directed at low-volume use cases. The economics of AI deployment are frequently misunderstood by organizations that conflate having an AI budget with having a deployment strategy. Building and launching an AI system is not a one-time cost; maintaining it across quarters and years is. When the underlying use case does not generate sufficient volume to justify that ongoing investment, the result is the absence of return on investment, not a failure of the technology but a failure of prioritization.

He offered a practical test for evaluating AI systems before deployment: rather than asking an AI agent a question once and assessing the answer, ask it the same question 5 times and evaluate whether it produces a consistent response. Most AI systems in production today will not pass that test. Consistency, he argued, is what health care operations are actually built on—and it is the most underweighted criterion in AI vendor evaluation.

Governance, Guardrails, and the Question of Transparency

The session devoted significant time to governance. Abraham offered a framework for what governance actually requires in practice: a portfolio of use cases with clear quality criteria, tooling for continuous large language model (LLM) evaluation in production, not just at deployment, and explicit policies about where human judgment must remain in the loop.

Jain introduced transparency as the next evolutionary step beyond guardrails. The prior generation of AI vendors operated as black boxes with inputs in and outputs out, with no visibility into the decision logic connecting them. The emerging generation offers full audit trails of agent decisions, logged in a way that enables systematic review, continuous improvement, and genuine accountability.

“We're early on this journey of AI adoption,” said Ankit. “There's a lot of changes that are happening, but one of the things that I'm starting to see that's changing is we're going from a black box model to a little bit more transparency, to what's happening underneath.”

Getting Out of Pilot Purgatory

One of the session’s most practically useful segments addressed what Jain called “pilot purgatory”—the organizational state in which AI initiatives are perpetually in evaluation, never reaching the production threshold where real value is generated, or real lessons are learned. Abraham distinguished between the legitimate reasons for piloting and the failure modes that trap organizations in indefinite pilot mode, unclear success criteria, governance teams that cannot close on requirements, and data infrastructure that generates hallucinations at a rate that makes production deployment indefensible.

Abraham also described Optum Rx’s approach as a 3-layer program: first, retraining technology teams, followed by executive education focused on the art of the possible, followed by bottom-up idea generation from the operators closest to patients and workflows. A fourth layer, customer education, is frequently overlooked, Abraham noted: organizations that retrain their own teams but fail to communicate transparently with clients about how AI is changing service delivery create a trust deficit that undermines adoption.

“Candidly, the best use pieces come from the bottom,” said Abraham. “The operator, closest to the work, closest to the patient, the client, etc, is really going to understand where those opportunities lie. And that's where I think differentiation happens.”

The Next 3 to 5 Years: Quality, Cost, and Hyperpersonalization

In the session’s closing segment, Abraham placed his bets on 3 areas where AI will deliver the most transformative value in pharmacy operations over the next 3 to 5 years. The first is quality: AI’s ability to apply consistency at scale to clinical processes represents a structural improvement over current human-variable performance. The second is cost, not merely through administrative efficiency gains, which Abraham acknowledged are already well-understood, but through real-time, data-driven trend management that enables daily course correction rather than monthly or quarterly review.

The third, and the one he described as most exciting, is hyperpersonalization: an AI system that accompanies each patient through their entire care journey, functioning simultaneously as a patient assistant, a pharmacist assistant, an operator assistant, and a client-facing tool.

Jain closed with a frame that situated the current AI moment in a historical context. When the web browser arrived in 1997, its first use cases were digital replicas of existing media. When the iPhone launched in 2008, the phone button was just a phone. The native innovations—Uber, DoorDash, the entire mobile-native economy—took years to emerge because they required developers and businesses to internalize what the new infrastructure actually made possible. AI in health care, he argued, is at exactly that inflection point. A technology that most organizations are still using to replicate what humans currently do, on the cusp of enabling things that humans simply cannot.

“The true magical power of AI is only starting to get unlocked,” said Jain. “We're all going to envision a world where AI can unlock things that we couldn't imagine doing before, and that's the potential of that health care journey becoming personalized, of each of us having a health care workforce, AI workforce, supporting us and being more efficient, being more present in a way that just wasn't possible before.”

References

1. Jain A, Abraham S. AI in healthcare: what works now, and what matters next. Presented at: AXS26; April 27-30, 2026; Las Vegas, NV.

2. Healthcare AI hit 75% adoption. Only 18% is governed. Alignmt.ai. April 18, 2026. Accessed April 28, 2026. https://www.alignmt.ai/post/healthcare-ai-governance-gap-2026