The AI suggested CPT 99215. The documentation supported 99214. Here’s what the difference teaches us about AI-assisted medical coding.
Artificial intelligence is becoming part of medical coding workflows. Depending on the technology, an AI-assisted system may review clinical documentation, suggest diagnosis or procedure codes, or recommend an evaluation and management (E/M) service level.
These capabilities can be useful. But what happens when the technology recognizes the correct information and reaches the wrong coding conclusion?
That’s a situation coding professionals need to be prepared to recognize.
An AI recommendation may sound convincing. It may even identify the right diagnoses, medications, and tests.
The question is whether the system applied the coding guidelines correctly.
Let’s look at an example.
The AI recommends CPT 99215
The following scenario is fictional and designed for educational purposes.
A medical coding professional is reviewing an established-patient office visit. An AI coding assistant has analyzed the encounter documentation and recommended CPT 99215, an office/outpatient E/M service associated with high-complexity medical decision-making (MDM).
The AI provides this explanation:
“Patient has multiple chronic conditions, requires prescription medication management, and underwent laboratory testing. High-complexity medical decision-making is supported.”
At first glance, the explanation sounds reasonable.
The patient has multiple medical conditions. Prescription medications were addressed. Laboratory tests were involved.
But those details don’t automatically establish high-complexity MDM.
The coder opens the actual encounter documentation to determine what the provider evaluated and managed.
What the physician actually documented
The signed encounter note describes the following:
Hypertension: The condition is controlled. The physician assessed the patient’s response to treatment and continued and renewed lisinopril.
Hyperlipidemia: The condition is stable. The physician assessed treatment response and continued and renewed atorvastatin.
Laboratory review: A previously obtained lipid panel was reviewed.
Laboratory order: A basic metabolic panel was ordered for routine monitoring.
The documentation does not describe severe exacerbations, high-risk management decisions, additional qualifying data activities, or total time supporting time-based code selection.
The AI recommended 99215.
Would you accept that recommendation?
Before answering, the coder needs to examine the three elements of medical decision-making.
Breaking down the medical decision-making
For office and other outpatient E/M services, the visit level may generally be selected using either medical decision-making or the total qualifying time on the date of the encounter.
When using MDM, the applicable CPT guidelines require at least two of three elements to meet or exceed the selected level.
Those elements are the problems addressed, data reviewed and analyzed, and risk associated with patient management decisions.
1. Problems addressed: Moderate
The physician evaluated and managed two stable chronic illnesses: hypertension and hyperlipidemia.
Under the E/M guidelines, two or more stable chronic illnesses addressed during an encounter meet the moderate level for the problems element.
The conditions do not have to be severely exacerbated to support moderate complexity.
At the same time, simply having multiple diagnoses does not establish high complexity.
Supported level: Moderate.
2. Data reviewed and analyzed: Limited
The physician reviewed a lipid panel and ordered a basic metabolic panel.
For this example, assume these are two distinct qualifying tests and that the review and order were not previously counted for another encounter.
Under the MDM data criteria, two qualifying Category 1 activities support limited data.
They do not meet the moderate data threshold by themselves.
A coder must also be careful not to count the same test twice when its review has already been included in the decision-making at the encounter where it was ordered.
The presence of laboratory work does not automatically increase the visit to moderate or high data complexity.
Supported level: Limited.
3. Risk of patient management: Moderate
The physician assessed the patient’s treatment response and made documented decisions to continue and renew prescription medications.
This supports prescription drug management, which is an example of moderate risk in the E/M guidelines.
Notice the distinction.
The medications were not simply listed in the patient’s chart. The physician made documented management decisions involving them.
That matters.
Routine laboratory monitoring does not automatically establish drug therapy requiring intensive monitoring for toxicity, which is a high-risk MDM example.
Supported level: Moderate.
So which E/M code does the documentation support?
Here is the MDM determination:
MDM element
Supported complexity
Problems addressed
Moderate
Data reviewed and analyzed
Limited
Risk of patient management
Moderate
Overall MDM
Moderate
Two of the three elements support moderate medical decision-making.
For this established-patient office visit, CPT 99214 is the better-supported code when selecting the level based on MDM.
CPT 99215 would require high-complexity MDM or documentation supporting selection by the applicable time requirements.
Neither is established in this fictional record.
The problem wasn’t that the AI failed to identify the patient’s medical conditions or medications.
It identified relevant facts.
The error occurred when it interpreted those facts as high-complexity decision-making without sufficient support.
That’s the difference between recognizing information and applying coding criteria correctly.
What should the coder do next?
The coder has identified a discrepancy between the AI-generated recommendation and the documentation.
The appropriate response is to verify the supported E/M level against the complete record and applicable CPT guidance, then follow the organization’s coding review or correction process before claim submission.
In this example, the signed documentation supports moderate MDM. A provider query isn’t necessary simply because the AI recommended an incorrect code.
A query should address a genuine documentation issue requiring clarification. It should not be used to obtain documentation that justifies a higher code when the record already supports a determination.
The coder should also understand the limits of the role. Some organizations authorize coders to correct the code directly; others require a particular review, education, or escalation process.
The important responsibility is to recognize the unsupported recommendation and prevent it from being accepted without proper validation.
Why a convincing AI explanation can still be wrong
This scenario illustrates a particular challenge with AI-assisted coding.
An AI system may identify clinically relevant information and still misapply the rules used to determine a code.
For example, it may:
Give too much weight to the number of diagnoses without determining their complexity.
Treat the presence of prescription medications as proof of a particular risk level without examining management decisions.
Overcount laboratory tests or other data activities.
Confuse moderate MDM elements with high-complexity requirements.
Present a confident conclusion without adequately explaining how the required elements were met.
These are possible failure patterns, not evidence that all AI coding systems behave this way.
Different systems have different capabilities, rules, review processes, and limitations.
What matters is understanding what the technology did and whether its recommendation is supported.
An AI-generated explanation is not a substitute for the encounter documentation or the applicable coding guidelines.
What happens when an unsupported recommendation isn’t caught?
Suppose the coder accepts the 99215 recommendation and the claim proceeds to billing.
If no other review identifies the discrepancy, the organization may submit a service level that the documentation does not support.
Depending on the circumstances, that can create payment errors, denials, overpayment concerns, additional audit work, or compliance exposure.
This does not mean every incorrect AI recommendation constitutes fraud.
A coding error and intentional misrepresentation are different matters.
But organizations remain responsible for the accuracy of claims submitted, whether a person or a technology-assisted system recommended the code.
CMS reinforces the principle that the level of E/M service reported must be supported and medically appropriate.
The objective is not to reject every recommendation from technology.
It is to prevent unsupported recommendations from becoming unsupported claims.
What this means for medical coding and billing workforce readiness
Learning E/M guidelines is essential. Knowing how to identify the correct code remains an important professional skill.
But actual workplace performance requires another layer of judgment.
A coding professional may receive a recommendation that appears complete, logical, and authoritative.
The professional still needs to recognize what the recommendation does and does not establish.
In the fictional example, the coder demonstrated several workplace capabilities:
Reviewed the original documentation rather than relying exclusively on the AI explanation.
Evaluated the three MDM elements using applicable coding criteria.
Identified why high-complexity MDM was unsupported.
Determined the better-supported code.
Recognized that a provider query was not automatically necessary.
Identified the appropriate correction process and potential downstream billing risks.
Those are the kinds of decisions coding professionals may need to make when technology assists their work.
And they’re important skills to practice before encountering these situations on the job.
Technology can recommend. The documentation must support.
AI-assisted coding has the potential to help professionals review information and identify possible coding options.
But speed and confidence are not the same as accuracy.
In this case, the AI recommended 99215.
A careful review of the documentation supported 99214.
The distinction wasn’t obvious from the AI’s explanation alone. It required someone to apply the MDM guidelines and examine the evidence.
At Queens Code Healthcare, our Medical Coding & Billing Workforce Readiness approach focuses on helping learners make those decisions: reviewing technology-assisted recommendations, identifying unsupported conclusions, explaining their reasoning, and recognizing when correction or escalation is appropriate.
Because the workplace doesn’t just require someone who can recognize a code.
It requires someone who understands whether the code is supported.
AI assists. Evidence rules. Humans decide.
This article is for educational purposes. The clinical scenario and AI recommendation are fictional. Coding determinations must be based on the complete documentation, applicable CPT guidance, payer requirements, and organizational policies.
Learn more: Queens Code Healthcare — Medical Coding & Billing Workforce Readiness
