When an AI medical coding platform cannot finalize a chart, it should stop, record why, and route the chart to a named reviewer with the supporting documentation attached. That handoff is exception routing. It decides which charts a medical coder sees, what evidence appears on screen, and what has to be true before the codes are ready for the claim.
Most coding managers can recite the automation rate a vendor quoted them. Fewer can say what happens to the charts the AI declined to finish, how many sit in that state this morning, or who clears them. That queue is where their team spends the day.
What is exception routing in medical coding?
Four parts make the path work: a trigger, a reason code, an owner, and a resolution standard.
An automation rate tells you how many charts skip this path. Exception routing tells you what happens to the rest.
What triggers an exception?
Four causes sit behind most exceptions, and each needs a different owner. Pooling them into one "needs review" queue is the common design mistake.
Model uncertainty. The platform produced a code it cannot stand behind. Nothing in the chart is wrong or missing.
Documentation ambiguity. The note contradicts itself, omits a required element, or two clinicians describe the encounter differently.
Coding-rule questions. The documentation is clear, but an NCCI edit, an MUE limit or a coverage policy makes the code or modifier a judgment call.
Technical failures. A page never arrived, an integration timed out, or the operative report is missing. No medical coding judgment is involved and none should be attempted.
The table below is a recommended operating design, not a description of any vendor’s product.
| Exception type | Owner | What the reviewer needs on screen | Resolution requires |
|---|---|---|---|
| Model uncertainty | Certified medical coder | Proposed codes, the documentation passage behind each, and the guideline applied | Medical coder confirms or replaces the code and records the basis |
| Documentation ambiguity | Certified medical coder, escalating to CDI | The conflicting passages side by side, with author and timestamp | Clarification from the record, or a physician query when the record cannot settle it |
| Coding-rule question | Coding lead or compliance | The triggering edit or coverage policy, cited | A documented decision against the cited rule |
| Technical failure | Operations or integration owner | What failed, which chart, and what is missing | The missing input arrives and the chart re-enters medical coding, never released as-is |
Every exception needs a reason code drawn from a fixed list. Free-text notes cannot be counted, and a queue you cannot count is a queue you cannot staff.
Why a confidence score is not a verdict
Teams reach for a single number and route anything below a threshold to review. Treat that number with care. Guo and colleagues found that a model’s stated confidence can drift from how often it is right: "We discover that modern neural networks, unlike those from a decade ago, are poorly calibrated."
A confident wrong answer does not raise its hand, so routing alone cannot carry your accuracy program. Any threshold is also a local claim about one model on one chart population. Ask what a score means on charts like yours, and how often that was checked.
Does every exception require a physician query?
No. Treating the query as the default response spends provider goodwill on charts the record already answers.
The ICD-10-CM Official Guidelines set the order, and the order turns on one distinction: whether the provider left something out, or wrote two things that disagree.
When the provider simply did not document laterality, Section I.B.13 allows you to look wider: "When laterality is not documented by the patient’s provider, code assignment for the affected side may be based on medical record documentation from other clinicians."
When the documentation conflicts, that door closes. The same section: "If there is conflicting medical record documentation regarding the affected side, the patient’s provider should be queried for clarification." Section I.A.19 states the general rule in the same terms: "If there is conflicting medical record documentation, query the provider."
One point is easy to over-apply. The warning that unspecified codes "should rarely be used" sits in I.B.13 and applies to unspecified laterality. Section I.B.18 treats unspecified codes generally as having "acceptable, even necessary, uses." Routing rules that treat every unspecified code as an exception will flood the queue.
AI uncertainty is not a provider’s uncertain diagnosis
Keep these apart. A model that cannot decide is a workflow problem. A provider who writes "probable" is a guideline situation with a defined answer.
For outpatient and professional-fee medical coding, Section IV.H says not to code diagnoses documented as "probable," "suspected," "questionable," "rule out," "compatible with," "consistent with," or "working diagnosis." Code to the highest degree of certainty for that visit instead, using symptoms, signs or abnormal findings. We covered the mechanics of that in how coding AI handles thin notes.
That is a rule, not an exception. A platform that treats every uncertain-sounding phrase as a reason to stop fills the queue with charts the guidelines already resolve.
A hypothetical ASC case
The following example is illustrative. It is not drawn from a real patient or a real deployment.
An ASC submits a knee arthroscopy. The scheduling record and the H&P say left knee. The operative report header says left, and the body of the report twice describes work on the right compartment.
The conflict sits inside the surgeon’s own documentation, which settles the routing question. This is not a case where the provider was silent and other clinicians can fill the gap. Under I.B.13 the chart goes to a physician query, and the platform should not pick a side or fall back to an unspecified code while that query is open.
The anesthesia record and the nursing note still matter. They show the reviewer which side the rest of the record points to, so the question reaching the surgeon is specific. They do not resolve it. The chart waits, and the ASC hears that the operative report needs correcting.
How should you measure exception resolution?
Four measures, reviewed weekly:
- Queue age. The oldest unresolved exception, by reason code. A rising figure in one code points at a staffing gap or a physician who needs a conversation.
- Resolution time. Median and 90th percentile from exception to release, split by reason code.
- Reviewer correction rate. How often the medical coder changes the proposed code. A rate near zero suggests over-routing. A climbing rate in one specialty is a signal to tighten what runs autonomously.
- Independent audit of autonomously released charts. Sample what never became an exception. Routing only measures what the platform flagged, and the charts it released without hesitation are the ones nobody has looked at.
The NIST AI Risk Management Framework treats this as governance rather than preference. MAP 3.5 asks that "Processes for human oversight are defined, assessed, and documented," and MANAGE 2.4 asks that responsibilities be assigned "to supersede, disengage, or deactivate AI systems" when performance drifts from intended use.
One boundary to state plainly for your finance team: this is coding accuracy, not payment. A correctly coded chart can still meet a denial on coverage, medical necessity or eligibility.
How MediCodio AI routes exceptions
Documentation Quality Gates read each note for contradictions, such as a site that disagrees with the procedure, and a chart that fails is flagged with the reason named rather than dropped into a general review pile. Routing rules and specialty settings are set per client in the interface, which is how the four exception types reach different owners.
On the guideline side, signs and symptoms and outpatient sequencing each run as their own validation pass, so an office note documenting "rule out" is handled as a coding rule rather than routed to a medical coder as uncertainty.
When the record genuinely conflicts, the Physician Query Portal sends the question to the physician and brings the answer back in the platform, with query, response and chart held together. Prediction Trail opens any code and shows the chart passage it came from and the guideline that justified it, giving a 100% code-level audit trail.
Across 35+ outpatient and professional-fee specialties, complex charts go to assistive coding, where a certified medical coder reviews and finalizes at 98%+ accuracy against AMA, AAPC and CMS guidelines.
Related reading
- How coding AI handles thin notes (pending publication)
- Catching NCCI and modifier errors before submission (pending publication)
- How to evaluate AI medical coding software
- How autonomous medical coding works
- The CODIO AI platform
Sources
- CMS/NCHS, FY 2027 ICD-10-CM Official Guidelines for Coding and Reporting: https://www.cms.gov/files/document/fy-2027-icd-10-cm-coding-guidelines.pdf
- NIST AI Risk Management Framework, Core: https://airc.nist.gov/airmf-resources/airmf/5-sec-core/
- Guo et al., On Calibration of Modern Neural Networks (ICML 2017): https://proceedings.mlr.press/v70/guo17a.html
