Heather Bassett, MD, is Chief Medical Officer with Xsolis, an AI-driven health technology company fostering payer-provider collaboration.
As a physician, I learned early that a recommendation and an order are not the same thing: one informs judgment, the other changes what happens next. That distinction now matters in how leaders think about AI.
For most of the past decade, enterprise AI stayed on the recommendation side of that line. In healthcare, a model flagged a patient at risk, scored a claim or surfaced a pattern. A person decided what to do next. The technology informed the work, but it did not carry the work forward.
That line is starting to blur. Emerging agentic systems can assemble information, draft correspondence, trigger next steps and follow up when something is missing. Most still operate in bounded workflows under human oversight, and they should. But AI is beginning to move from generating an answer to carrying work forward.
That changes the strategic question. Leaders still need to know whether a model is accurate and useful, and whether it treats people fairly. A strategy built for recommendation tools falls short for systems that can act. It also has to address what the system may do, where it must stop and how we will know if it acts wrongly.
AI strategy has to follow the action.
Organizations have spent years building processes to evaluate decision-support tools. In healthcare, we ask how a model was trained, whether it performs across populations, how its output is explained and who reviews the recommendation. Those questions still matter, but they no longer cover an agent or the strategy around it.
A strategy for agentic systems cannot be just a list of use cases, vendors and models. It has to define how AI fits into daily operations: which workflows are appropriate, which actions may be delegated, how we measure performance and who remains accountable. Governance turns those choices into enforceable boundaries.
Leaders also need to know what tools the agent can access, what data it can change and where a human must intervene. Requiring approval for every low-risk step recreates the administrative burden the technology was meant to reduce. Skipping review on high-consequence actions creates obvious risk. The practical answer is risk-calibrated autonomy.
An agent that gathers records or checks for required information may operate with limited intervention. One that drafts a prior authorization letter may create and queue the document but require approval before it goes out. A system that could alter a clinical order, a patient status or a financial determination belongs in a different risk tier altogether. Set those boundaries deliberately, not by inheriting a vendor’s default configuration.
Start with the work, not the technology.
Whenever a new AI capability arrives, my first question is intentionally boring: What problem are we solving? Agentic AI should not get a pass on that discipline just because the technology is new.
Every industry has work that is high-volume, rules-heavy and spread across too many systems. In healthcare, utilization review, prior authorization correspondence or mid-revenue-cycle documentation are obvious examples. These processes involve multiple steps, but the steps can often be defined, measured and audited. That makes them reasonable places to explore bounded agents.
I did not go to medical school to move the same clinical facts from one screen to another, and I have yet to meet a clinician who finds that work meaningful. The real opportunity is to reduce administrative friction so clinicians and staff can spend more time on judgment, communication and the parts of care that actually require a person.
That also means measuring more than activity. The number of tasks an agent completes tells you little about its value. Before deployment, leaders should define the outcome: faster turnaround, less rework, fewer avoidable escalations or meaningful time returned to the team. If the only metric is that the agent did more things, we may be automating motion rather than solving a problem.
Governance is how autonomy earns trust.
I sometimes hear caution and speed discussed as though they are opposites. With agentic AI, governance is not separate from the strategy. It is what makes responsible speed and scale possible.
A flawed recommendation that a person catches may cost time and attention. A flawed action repeated across thousands of cases before anyone notices can create a much larger failure. That is why autonomy should start in bounded workflows and stay closely monitored.
Monitoring also has to reach beyond model accuracy. An agent depends on data, tools and workflow rules, any of which can break down as business processes, policies and operating conditions change. Leaders need to watch the actions the system takes, not just the predictions it makes—exception rates, human overrides and downstream corrections. Every agent also needs a named owner with the authority to pause it or roll it back.
NIST’s AI Risk Management Framework offers a useful structure for governing AI risk, and ONC’s HTI-1 rule already requires certified health IT to apply risk analysis, mitigation and governance to predictive tools. Agentic systems extend that same work to governing actions, not just outputs.
Workforce readiness belongs inside the strategy. When we deploy AI quietly and meet questions with vague answers, employees find their own tools and build their own workarounds. That silence becomes a governance gap. Be clear about where AI is used, what it can and cannot do, and how employees should report something that looks wrong.
There are three questions your AI strategy should answer.
The strategy becomes actionable when leaders can answer three practical questions about every agent they deploy.
1. What exact actions may the system take, and which actions are off-limits?
2. How will we know it is performing as intended, and what will trigger additional review?
3. Who owns the outcome and has the authority to pause or roll back the system?
Agentic AI is still emerging, but leaders should not wait for widespread autonomy to decide what acceptable autonomy means, or treat every agent as though it were making an independent clinical decision. I believe the advantage will go to organizations that define boundaries for each agent action, match oversight to consequence and revisit those decisions as operations evolve.
Forbes Technology Council is an invitation-only community for world-class CIOs, CTOs and technology executives. Do I qualify?