Agentic AI in Healthcare: Why Data Foundations Decide the Outcome

Many organizations that claim they are "on an AI journey" are not as concrete as the phrase suggests. On a Signal Room conversation, Gary Cao, with three decades of experience in the field as a chief data, analytics, and AI officer, explained such claims usually consists of a lot of talking and very little tool evaluation, with no comprehensive intention to construct a framework or a pathway for the upcoming three to five years. The hype will always be quicker than the strategy, and agentic AI is widening the gap.

People are right to think of the versions of AI, from analytics to generative models to the present day autonomous agents, as waves the industry is chasing one at a time. But these are in fact, layers of a single evolution. Organizations that have the most trouble incorporating AI are those that have all of the layers, and treat them as distinct and fragmented. At Hutchins Data Strategy Consultants we have noticed that organizations with the most enthusiasm for deploying Autonomous Agents are the most unprepared, and often the most clueless, as to what these Agents will do in organizations with poorly constructed systems and structures, with poor and unintegrated data.

Cao breaks enterprise AI into four interdependent elements: business strategy, data, the analytics and algorithms that turn data into insight, and the technology and execution that put it to work. The failure mode is not the absence of any one of them, but the absence of the connections between them.

Because technology is the most visible aspect of any advancement, it gets the most focus. Data governance, which sits beneath the surface and is hardly ever showcased during a demonstration, gets the least focus. The most troublesome aspect, in his view, is analytics — figuring out how to connect technology and data to a business issue and generate an output that is usable by a business decision-maker. That aspect of connectivity is difficult and often not maintained. This is the reason many organizations possess the best technologies and the most flexible infrastructures and yet, cannot provide a single example of a business decision that has been improved with those technologies and infrastructures.

Agentic AI does not fix any of this. An agent is something that exists at the intersection of all four areas and does something about it. Any type of weakness that does exist in the chain becomes that agent’s weakness and does so in the form of an action, as opposed to a report that can be stopped and questioned by a human.

Rapidly developing technology has also failed to alleviate the burden of data. One of the most notable changes that Cao mentioned was related to the speed of things. Building a data science model would consume hours or even days, weeks, or months. Now, having the right data, confidence in the data, and clarity of the workflow can render that model in a matter of seconds or minutes. This is a true acceleration and has been and will continue to be, widely misunderstood.

It is an error to assume that a faster model equates to less valuable data. In fact, the inverse is true. The rapid development of a model increases the significance of the data’s relevance, trust, and utility. Without a lengthy model development phase, problems with the data may go unnoticed. Rapid development eliminates the buffer. What was previously accomplished with the effort of model building must now be achieved with disciplined framework construction.

For these reasons, governance must evolve. Regarding the new risks that language models create, Cao stated that data governance must evolve beyond the traditional methods. With new risks of information leaks and model failures from unpredicted inputs, he stated that to accomplish this, there must be a culture of discipline and education with a clear priority to reduce the risk of significant and irreversible errors. With an autonomous agent, significant and irreversible errors are no longer an abstraction.

In a Deterministic Industry, with Probabilistic Systems

Healthcare is a highly deterministic environment and AI systems are inherently probabilistic. In order to reconcile these opposing systems, Cao referenced a saying attributed to statisticians, that all models are wrong, but some are useful. Leadership needs to develop an understanding of this concept to allow the integration of AI into highly deterministic environments.

The tolerance for risk should match the expected consequences. Ample variance can be absorbed by the marketing department, but this cannot be said for the risk management department. Even the level of variance absorbed by the operational or logistical activities of the organization cannot be equated to the variance that the clinical decision-making process can tolerate. The same agent architecture can be appropriate or reckless, and the difference is the risk of making a mistake.

Cao has the comforting reassurance to leaders that the unfamiliarity they presume to have in such situations is misplaced. The probability model is simply a much more defined and specific example of the uncertainty that executives have to deal with, and it can, and should, be addressed with the grounded, sound principles with which they have dealt with uncertainty in the past, and with which they have dealt with uncertainty in the past, and with the guardrails they have previously employed.

Adoption Is a People Problem Before It Is a Technical One

In many cases, the biggest barrier to enterprise AI is the workforce. Cao was blunt in saying that the workforce had to be upskilled and retained, and that the perception of AI as a means of job substitution was a guarantee of resistance and, most importantly, a lack of technological adoption. This challenge is as cultural and philosophical in nature as it is technical, and it cannot be resolved with better and more effective models.

It alters how investment should be evaluated. Cao argued that overlooking short-term financial returns when evaluating AI investments is shortsighted. Two factors should be considered. The first is the measurable value derived from AI like increased revenue and reduced costs. The second is the type of value that is strategic and difficult to quantify, like the retention of employees and the trust of clientele. An organization that evaluates AI exclusively according to the first factor will systematically overfund the first dimension while underfunding the second, and will likely be at a competitive disadvantage for reasons that are not revealed in its financial models.

How Hutchins Approaches Agentic and Enterprise AI

Our work starts by connecting the four elements Cao describes rather than buying into any single one. We help organizations build the strategy and roadmap that a real AI journey requires, strengthen the data foundations that determine what an agent can safely do, and evolve governance to match the autonomy and risk of the systems being deployed. We are equally candid about the people dimension — the upskilling and trust without which adoption fails regardless of the technology.

This is continuous with the readiness and governance work that any serious AI program depends on. Agentic AI does not change the fundamentals; it raises the cost of getting them wrong.

These questions run throughout The Signal Room podcast, where leaders who have moved organizations from experimentation to enterprise scale describe what actually carried the weight.

Authoritative sources

• NIST AI Risk Management Framework
• WHO — Ethics & governance of AI for health
• HL7 FHIR interoperability standard
• HIMSS
• NEJM AI