• Healthcare 150
  • Posts
  • Agentic AI in Healthcare: From Assistance to Autonomy

Agentic AI in Healthcare: From Assistance to Autonomy

Healthcare spent three years piloting generative AI. Now it's starting to delegate decisions. What's actually changing, where the money is, and why the clinical evidence still lags behind the enthusiasm.

Author

Valentin Monza
August 11, 2026

For most of the past decade, artificial intelligence in healthcare was, above all, a promise of assistance. Models that summarized charts, drafted notes, answered questions, and suggested diagnoses—always waiting for a human to press the button. That era isn't over, but it's no longer the frontier. Today the frontier is called agentic AI: systems that don't wait to be asked but instead set sub-goals, execute multi-step workflows, and take real-world actions—querying a database, submitting an authorization, monitoring a patient—with human oversight that grows steadily lighter.

The difference isn't merely semantic. It marks the shift from a tool that responds to a collaborator that acts. And it arrives at a moment when the sector is already saturated with operational friction: according to Salesforce, 87% of healthcare workers spend extra hours each week on administrative tasks. On the patient side, digitization has matured too—BCG estimates that close to 50% of U.S. adults use health apps and roughly a third use wearable monitoring devices—creating the continuous data flow that agents need to operate. Healthcare, one of the most regulated, fragmented, and risk-sensitive sectors in the world, is approaching this transition with an unusual mix of urgency and caution. The most recent surveys from McKinsey, Deloitte, BCG, and IBM agree on one point: adoption has stopped being a laboratory experiment and has become an investment decision. The scale of the bet is hard to overstate—BCG estimates that agentic AI will unlock some $200 billion in net-new value across tech services over the next five years, with more than 40% of large enterprises already scaling agentic implementations. Yet the same scientific literature that should underpin that move—systematically reviewed in npj Digital Medicine—shows that rigorous clinical evidence is still scarce.

This deep dive walks through that tension: what agentic AI actually is, how fast it's being adopted, where the economic value concentrates, who's capturing it, and what separates the organizations already reaping returns from those still watching from the sidelines. And it ends where any honest conversation about healthcare should end: with the evidence.

What Agentic AI Is: The Leap from Assistant to Collaborator

To understand why agentic AI generates so much noise, it helps to place it on a spectrum. At one end sits traditional AI: reactive, good for simple predefined tasks, with no memory and no capacity to adapt. Then came generative AI, creative and surprisingly capable, but prompt-dependent: it delivers output when asked and operates with short-term memory. The "AI agent" adds proactivity and orientation toward specific goals. And at the most autonomous end appears agentic AI: proactive, self-directed, able to orchestrate complex workflows, invoke external tools, collaborate with other agents, and sustain long-term memory.

The systematic review published in npj Digital Medicine proposes three minimum criteria for a system to earn the "agentic" label: autonomous operation (it functions without constant human intervention), goal-directed behavior (it monitors outcomes and adjusts its actions to achieve them), and the ability to initiate action (it invokes APIs, databases, or tools on its own). Reasoning well isn't enough; it has to act. IQVIA's metaphor is telling: if generative AI is like an eager high-school intern, agentic AI is its older sibling with advanced degrees, executing independently instead of waiting for instructions. Or, as another framing puts it, where generative AI behaves like "a very capable assistant that waits to be asked," agentic AI functions "more like a coworker than a tool"—one that sets its own sub-goals and carries a multi-step task through to completion. That shift, from a system you prompt to a system you delegate to, is the conceptual heart of everything that follows.

Adoption Has Stopped Being an Experiment

The adoption numbers tell a story of accelerating maturity. According to McKinsey's survey of U.S. healthcare leaders, the share of organizations reporting they had implemented generative AI reached 50% in the fourth quarter of 2025, up from 47% a year earlier and just 25% in 2023. For the first time, every single respondent said they had at least some plans to move forward with the technology: organizational hesitation has all but vanished.

Agentic AI, far younger, trails behind but with momentum. In the same survey, 19% of organizations already reported having implemented multi-agent workflows, while an additional 51% were building proofs of concept and only 1% said they had no plans to pursue AI agents. Put another way: eight in ten healthcare organizations are already, at minimum, experimenting with agentic AI. The generative-AI curve—from proof of concept to deployment—appears to be repeating itself with agents, only faster.

One caveat for reading these figures: adoption is uneven. Healthcare services and technology (HST) firms lead in maturity, clinical-care organizations sit in the middle, and payers are still below 50% implementation. Each subsector, moreover, pursues a distinct pattern: providers tend toward function-specific multi-agent solutions, while payers aim to automate end-to-end workflows.

Where Agents Are Scaling Today

Experimenting is one thing; scaling is another. When McKinsey asked—in a broader global survey of nearly 2,000 participants—in which industries and functions the use of AI agents had already reached the scaling phase, the map revealed that healthcare doesn't lead, but isn't left out either. The technology sector dominates, especially in software engineering (24% of respondents report scaled use) and IT (22%). In healthcare, the function where agents have advanced most is knowledge management, at 14%, alongside IT roles.

The pattern makes sense. Agents scale first where the ground is digitally mature, the data is structured, and the cost of an error is contained: documentation, information synthesis, repeatable administrative tasks. Other measures confirm it: according to data cited by healthcaretoday from Capgemini, 14% of organizations already run agentic AI at partial or full scale and a further 23% run pilots, while 39% of healthcare executives—per the IBM Institute for Business Value—already use AI for inpatient monitoring and early-warning systems, with the expectation of full agentic implementation in that field within three years. By contrast, high-risk clinical functions—where a wrong autonomous decision can harm a patient—advance far more cautiously. It's the same reason the first major use case organizations cite over and over isn't autonomous diagnosis but administrative efficiency: prior authorization, revenue cycle, medical coding, credential verification. There, McKinsey's promise is striking: a 30% to 60% reduction in cost-to-collect through agentic automation of the revenue cycle.

The Economic Case: Why Organizations Are Investing

If adoption is accelerating, it's because someone expects a return. And here Deloitte's data is telling. According to its 2025 survey on agentic AI in healthcare (100 organizations: 50 health systems and 50 health plans), 61% of executives are already building or implementing agentic AI, or have at least secured budget to do so. Looking ahead, the commitment is even more pronounced: 85% plan to increase their investment over the next two to three years—23% significantly and 62% moderately—while just 15% intend to reduce it or keep it flat.

The flip side of that investment is the expectation of savings, and it's remarkable for how widespread it is. Fully 98% of respondents expect to cut costs by at least 10% thanks to agentic solutions, and more than a third—37%—anticipate savings of 20% or more. In a sector pressed by thin margins, staffing shortages, and cost pressure, that kind of projection explains why agentic AI has gone from an innovation curiosity to a line in the operating budget. The focus, as executives themselves summarize, is beginning to shift "from adoption to scale, and from pilots to real operational impact."

But cost savings are only one face of the economic case; the other is speed. Agentic AI doesn't just do the same thing more cheaply—it radically compresses the timelines of processes that used to take weeks or months. In life sciences, IQVIA documents how commercial planning can shrink from a range of 6 to 18 months down to just 4 or 5. IBM reports even more extreme cases in regulatory operations: time to a new drug application (NDA) submission fell from seven months to two or three, and one internal initiative cut content-publishing time from six weeks to six seconds. In drug development, BCG projects that agents could shorten timelines "from years to months" by generating new molecules and simulating their interactions. And looking forward, Gartner—cited by IQVIA—anticipates that by 2028, 15% of day-to-day work decisions will be made autonomously, from a near-zero baseline in 2024. Speed, in industries where time-to-market is worth millions, may ultimately prove an even more powerful argument than direct savings.

From Spend to Return: ROI Is Becoming Measurable

Expectation is one thing; proven return is another. The good news for the technology's advocates is that the gap between the two is closing. McKinsey's survey found that 82% of healthcare leaders who implemented generative AI expect a positive return on investment—the highest proportion ever recorded in its surveys—and, more importantly, that 45% can already quantify that return, also an all-time high. Among those who quantify it, ROI levels mostly fall in a range from less than two times to four times the initial investment.

The breakdown is instructive. Of the leaders who have already implemented, 37% see a "potential positive return," 25% report an ROI of less than 2x, 16% between 2x and 4x, and 4% above 4x. At the other extreme, just 3% report a negative return and another 3% a potential negative return, while 12% still can't clarify the value. The picture isn't euphoric, but it does show a technology beginning to pay its own bills—something that can't always be said of prior waves of innovation in healthcare.

That passage from promise to measurable number is perhaps the most significant change of the past year. Where generative AI was for a long time an expense justified on faith, agentic AI arrives at a moment when organizations demand—and are beginning to obtain—evidence of return.

Early Adopters vs. Watchers: The Widening Gap

But does everyone capture that value equally? The data suggests not—and that the window to avoid falling behind may be closing. Deloitte segments organizations into three groups: early adopters (34% of respondents, already building or implementing agentic AI), starters (27%, who have secured investment), and watchers (39%, who plan to explore only once they see external evidence).

The difference between these groups isn't just timing but expected reward. Among early adopters, 59% expect to capture savings of 20% or more over the next two to three years. Among watchers, only 13% do. It's what might be called the watcher's paradox: those who wait for certainty before moving are precisely the ones anticipating the least benefit. And there's a structural asymmetry behind it: early adopters are 65% large organizations (revenue above $5 billion) and 82% are betting on multi-agent systems that span several functions. Watchers, smaller and more inclined toward point solutions, start the race from further back.

The risk, for anyone watching from the stands, is twofold: not only do they arrive late to the learning curve, but they also start from a less ambitious architecture. In a technology whose value grows with orchestration—agents collaborating across functions—starting with isolated solutions can become a hard-to-reverse disadvantage.

The Barriers That Are Falling Away

Part of what explains this acceleration is that several of the obstacles that held back AI in healthcare for years are dissolving. When Deloitte asked organizations which challenges that previously limited their use of AI are no longer a problem, the answers painted a picture of maturation. Forty percent said they had overcome a lack of technical talent or skills; another 40%, uncertainty about the return on vendor investment. Thirty-eight percent had left cultural resistance to change behind, 35% a lack of leadership buy-in, 34% regulatory uncertainty, and 32% data quality or access issues.

Only 1% answered that "none of these" had stopped being a challenge—that is, almost every organization feels it has progressed on at least one front. That data point is as important as the adoption figures: it suggests that the friction that made AI in healthcare a perpetually "promising but premature" bet has begun to give way across the board. Talent became more accessible, leaders were won over, data improved, and governance frameworks—though incomplete—advanced enough to unblock projects.

The strategic reading is clear: if the classic barriers are falling for everyone, competitive advantage will no longer lie in clearing them first, but in what each organization builds once they're cleared.

The Risks That Still Slow the Scale

That many barriers are yielding doesn't mean the path is clear. When McKinsey probed the obstacles that today make it hard to scale generative AI, the top one was neither technical nor regulatory, but integration: 59% of leaders cited the difficulty of integrating or adapting tools to existing workflows. Next came concerns about risk (43%), a lack of internal capabilities to leverage AI (31%), and insufficient data or tech infrastructure (31%). The pattern reveals a phase shift: as organizations move past the planning and proof-of-concept stage, the challenge stops being "is it worth it?" and becomes "how do I embed it into legacy, complex, mission-critical systems?"

As for the specific risks that most worry leaders, the list is revealing for anyone who works in healthcare. Sixty-six percent point to inaccuracies, biases, or flaws in the models; 60% to security risks; 52% to regulatory compliance; and 49% to ethical and privacy concerns. These aren't abstract fears: when a language model interprets a pain rating of "10" as "severe pain" in the clinical record, that interpretive leap, multiplied at scale, can create legal liability and erode clinician trust. That's why the serious conversation about agentic AI in healthcare is, inseparably, a conversation about governance: bias mitigation, observability of the agents, and explainability of their decisions—supported by frameworks such as the NIST Risk Management Framework, OWASP protocols, and HIPAA compliance.

The Clinical Evidence Gap

And here comes the necessary counterpoint. The entire conversation above—adoption, investment, ROI, governance—comes from surveys of executives and analyses by consulting firms. It's valuable information, but it measures intent and expectation, not proven clinical outcomes. When you look for that other, harder kind of evidence, the picture changes dramatically.

The systematic review published in npj Digital Medicine screened 984 initial records across five scientific databases. After removing duplicates and applying eligibility criteria, only seven studies remained that evaluated genuinely agentic AI systems in medicine, published between 2017 and 2025. Of those seven, only one was a randomized controlled trial—and with just 42 patients. The other six did not involve real patients: they were experiments or simulations. None of the seven demonstrated sustained adaptive behavior over time, one of the defining capabilities of the agentic paradigm.

The individual results are promising—one system reached 94.1% accuracy in diagnosing hepatocellular carcinoma; a virtual-reality assistant, 87% accuracy—but the base is too narrow to support general conclusions. Six of the seven systems were never deployed in real clinical practice, and the quality of evidence was rated between moderate and critical in risk of bias. On top of this sits a regulatory gap: current frameworks, such as the FDA's PCCP or the European AI Act, still don't account for the particular characteristics of agentic systems, such as their varying levels of autonomy.

The conclusion isn't that agentic AI doesn't work, but that a real gap exists between commercial enthusiasm and clinical validation. The money and the surveys run ahead; rigorous trials, behind. For a sector where the standard of proof is—and must be—higher than in any other industry, that gap is the single most important variable to watch.

What to Watch from Here

Agentic AI in healthcare lives, in 2026, a productive paradox. On one hand, every market indicator points in the same direction: adoption is accelerating, investment is being committed, ROI is starting to materialize, and the classic barriers are yielding. On the other, rigorous clinical evidence is barely being born, regulation is running behind, and the risks of integration, bias, and security remain real. Both things are true at once, and the sector's maturity will be measured by its ability to hold both ideas without abandoning either.

For decision-makers, a few practical conclusions take shape. A useful rule of thumb comes from BCG's work on AI transformations—the so-called 10-20-70 rule: roughly 10% of the effort goes to the algorithms, 20% to the technology and data, and 70% to people and processes. The lesson is that agentic AI is far less a modeling problem than a change-management one; the organizations that treat it as a software purchase rather than an operating-model shift tend to stall at the pilot stage. First, the safest ground to start on remains the administrative—revenue cycle, authorizations, documentation—where the value is tangible and the risk contained. Second, competitive advantage no longer lies in clearing the barriers to entry, which are falling for everyone, but in the ambition of the architecture: those betting on agents that orchestrate across functions expect to capture far more than those deploying isolated solutions. Third, and perhaps most important for the long term, the organization that wins won't be the one that automates fastest, but the one that manages to pair the speed of deployment with the rigor of the evidence. In healthcare, autonomy without validation isn't innovation—it's risk.

The leap from assistance to autonomy is already under way. The question that will define the next stage isn't whether agentic AI will transform healthcare—that looks increasingly inevitable—but whether the sector can prove, by the clinical standard it deserves, that this transformation actually improves patients' lives.

Sources