Health care’s relationship with artificial intelligence (AI) often swings between two extremes: enthusiasm and anxiety. On one hand, AI is helping clinicians reduce administrative work, uncover relevant information buried in medical records, and identify opportunities to improve patient care. On the other, regulators are becoming increasingly concerned about how AI is used in high-stakes decisions that affect payment, compliance, and patient outcomes.

In February 2026, the Office of Inspector General (OIG) issued guidance specifically identifying the use of AI-generated prompts that add risk-adjusting diagnoses as a potentially abusive practice. The concern is clear that when AI flags a diagnosis without adequate clinical support and that code goes unreviewed, it can result in inflated risk scores, improper payments, and ultimately, repayment demands from CMS. For health plans and providers that have rapidly adopted autonomous coding technologies, the message is hard to ignore: Efficiency cannot come at the expense of accuracy and oversight.

As someone who has spent decades working across clinical care, provider operations, health plans, and medical coding, I see both the promise and the limitations of AI every day. The technology is advancing rapidly, but there is a growing misconception that because AI can perform well on some clinical tasks, it can also reliably navigate the complexities of risk adjustment coding.

Clinical expertise and coding expertise are not the same

Diagnosing and treating patients is fundamentally different from assigning codes. Medical coding requires interpreting complex documentation rules and regulatory guidance, and the sheer volume of those requirements, spanning thousands of coding guidelines, payer-specific policies, and evolving CMS directives, makes it one of the most cognitively demanding administrative tasks in health care. That is precisely where AI can add real value, helping manage complexity without replacing the coder, so the human reviewing the work can focus on judgment instead of recall. The efficiency gains are real, but they depend on a trained professional’s final sign-off. AI’s track record in clinical decision support does not automatically transfer to coding accuracy, and the contextual judgment that determines whether a diagnosis is documentable and reportable is still where errors tend to surface.

I have personally seen AI coding fall short in ways that carry real clinical and financial consequences, from attributing a spouse’s diagnosis to the patient, flagging resolved conditions as active, and even misinterpreting abbreviations, creating diagnoses that were never present in the first place. AI can identify information in a patient chart, but determining whether it is clinically appropriate and reportable for risk adjustment still requires human judgment and contextual interpretation.

Part of the challenge is risk adjustment is not simply a matter of finding more diagnoses; rather, it is ensuring all diagnoses are accurate, supported, and compliant. Both CMS and OIG are increasing their focus on coding accuracy and overpayment recovery. Often, coding errors can go undetected until an audit surfaces them years later, and even small inaccuracies can carry significant risk. With this in mind, health plans would be wise to view AI less as a replacement for clinical and coding oversight and more as a tool that supports the decisions driving it.

Real-time support is better than retrospective fixes

AI’s most meaningful contribution to risk adjustment might actually be earlier in the process. Rather than coding charts after a patient encounter, AI can help physicians ensure accurate documentation while they are still with the patient. Health care has long relied on retrospective chart reviews to catch and correct coding gaps, even though that approach is expensive, time-consuming, and places a real burden on providers. Chart retrieval, coding review, validation, and resubmission add up, and when physicians receive documentation questions weeks or months after a patient encounter, it contributes to the frustration and burnout already straining the workforce. By surfacing relevant information, flagging documentation gaps in real time, and supporting clinical decision-making during the encounter itself, AI can help providers get it right the first time, cutting down on retrospective corrections without adding to their workload. That is where the efficiency gains become tangible, and where AI earns its place in the workflow.

This kind of real-time support is especially valuable for complex diagnoses like cancer. AI can quickly identify documentation related to a patient’s cancer history, but it cannot reliably determine whether the documentation supports reporting active cancer for ICD-10 coding. For example, a physician may document “prostate cancer” while ordering a PSA test to monitor a patient’s history of disease. Clinically, that may be appropriate, but monitoring alone does not satisfy coding requirements for active cancer. Presenting that nuance to the provider during the visit, while a human still validates the coding decision, is far more effective than trying to correct it months later through retrospective chart review.

The future is human-guided AI

While there’s a lot of buzz around AI, we won’t see the true benefits in health care by replacing human expertise, but by making the technology sharper. The organizations that will see the greatest success with AI are those that balance innovation with oversight. As adoption accelerates, competitive advantage will come from using AI to augment clinicians, coders, and health plans, enabling more informed decisions while maintaining trust, compliance, and high-quality patient care.

Kate Marta is a health care executive.