AI & CPT Codes
When Precision Doesn’t Mean Predictability
Recently, I called my insurance company to confirm coverage for a specialist visit. They asked for the CPT code. I got it from the provider, called back, and was told the service was covered—but only for part of what might occur during the visit. When I asked what that meant, the answer was: “Do what you need to do.” In practical terms, proceed without a clear view of financial exposure.
This reflects how the system is structured. Coverage determinations sit with the payer. Clinical decision-making and coding sit with the provider. Everything in between—benefits design, utilization rules, call centers, and technology platforms—passes information back and forth without clear accountability for the outcome. When things don’t line up, the risk shifts to the patient.
Operationally, payers adjudicate at the claim line level, not the visit level. Providers deliver care based on clinical need, not pre-defined billing boundaries. The interaction is only fully resolved after the claim is submitted and processed. Anything before that is directional.
Standards Do Not Mean Predictability
CPT codes standardize how services are described and serve as inputs to reimbursement—not definitive answers. A visit can generate multiple codes based on findings, time, and medical decision-making. “Covered” at the code level does not translate into a predictable out-of-pocket cost.
Even when patients follow the expected steps—confirming benefits, validating a CPT code—the result remains uncertain because the system is not designed to produce a single, reconciled answer in advance. Each party may be operating correctly within scope. The gap is in how those stages connect.
This is often framed as a transparency problem, but that’s incomplete. The data and rules may exist. The issue is how they come together in practice. Without real-time connection between benefit design, clinical variability, and coding, the patient becomes the point of integration.
Some complexity is unavoidable. Clinical pathways evolve, coding reflects what occurs, and medical necessity is often evaluated after the fact. At the same time, payer incentives and risk controls limit how definitively coverage can be determined in advance. The result is structural uncertainty.
AI Isn’t a Full Solution
AI is increasingly positioned as the fix. It can ingest benefit structures, coding logic, historical claims, and clinical patterns at scale. In theory, it should improve real-time estimates.
But the limitation isn’t primarily computational—it’s structural.
If AI is deployed within the same fragmented model, it may increase efficiency without resolving the disconnect. The system can become faster and more precise at each step while still failing to produce a coherent answer when the patient needs it.
Real-time benefit tools and cost estimators already exist, but they often return point-in-time answers tied to a single code, without accounting for how visits evolve. They create the appearance of precision without improving predictability.
The question isn’t whether AI can improve parts of the system. It’s whether it will be used to connect those parts while supporting the patient experience.
Used narrowly, AI refines existing processes. Used deliberately, it could link benefit design, likely clinical pathways, and coding variability into a bounded, scenario-based view before care is delivered.
That shift is less about new technology and more about application. It requires aligning incentives, defining ownership across the patient journey, and designing outputs that reflect how care actually unfolds.
My visit turned out fine, but that wasn’t system performance—it was outcome variance. Clinical uncertainty is expected. Financial uncertainty, when inputs are known, is partly a design choice. CPT codes add precision to billing, but without alignment across payer, provider, and patient-facing tools, that precision doesn’t translate into predictability.