The care you delivered but never billed

The demo was supposed to be about something else entirely. We were walking a specialty-clinic group through a document pipeline, and to make the walkthrough concrete we ran their own sample chart, de-identified, through an extraction pass. On screen, the note listed a patient's diabetes with early kidney involvement, managed and monitored, documented across three visits in the physician's own words.

The billing file for those visits coded uncomplicated diabetes. Every time.

The CFO leaned toward the screen and asked the operations lead a question neither of them enjoyed: "How often does that happen?" Nobody knew, which was the real answer. The clinicians were doing the work and writing it down. The codes, entered under time pressure by people juggling encounter volume, recorded a simpler patient than the note described. And in every value-based contract and risk-adjusted payment model the group had signed, the payment follows the coded patient, not the treated one.

Care that is documented but not coded is revenue you earned, spent the clinical effort for, and then declined to be paid.

This is not an obscure failure. Risk-adjustment audits routinely find that documented chronic conditions go uncaptured in claims at rates of 20 percent and up; industry estimates put uncaptured risk-adjustable conditions in the hundreds of dollars per member per year for the populations where it matters. The mechanics are mundane. A specialist manages the condition and assumes primary care coded it. A condition captured last year never gets re-documented this year, and in risk models the slate wipes annually. The note says "CKD stage 3 secondary to diabetes" and the code list says diabetes, unspecified, because the specific combination code is the 14th item in a dropdown at 6pm.

None of this is fixable by telling physicians to code better. They are already carrying the documentation load; the gap lives between free text and a terminology, and that translation is not a clinician's job.

The method is clinical NLP grounded in the coding ontologies. A language model tuned for clinical text reads the encounter notes and extracts condition mentions with their attributes: current versus historical, asserted versus negated versus suspected, severity, laterality, causal linkage. "No evidence of retinopathy" must land as an exclusion; "mother had CHF" belongs to family history, not the problem list. Assertion classification and negation handling are decades-old subfields for a reason, and generic text extraction without them produces coding suggestions that are confidently, dangerously wrong.

Each extracted condition then grounds to the terminology: mapped through SNOMED to the specific ICD-10 code the documentation supports, with the combination logic the dropdowns defeat. The output is a gap list per encounter, documented-but-uncoded conditions, each with the exact sentence from the note attached as evidence, routed to the coding team as a review queue, not an auto-commit. The coder stays the decision-maker; the system's job is that nothing documented slips past unread. Uncertain groundings abstain and say so. In a domain where upcoding is a federal offense, the evidence-linked, human-confirmed design is not a nicety; it is the entire difference between revenue integrity and a compliance incident.

(Run in the other direction, the same machinery flags coded-but-undocumented conditions before a payer's auditor finds them. The list is rarely empty, and finding it yourself is considerably cheaper.)

We built this grounding layer, note to ontology with evidence sentences attached, inside our own extraction stack, and the assertion handling consumed more engineering than everything else combined. That is normal. It is also the part that makes the output usable.

The clinic group ran the retrospective: one year of encounters, extraction against claims. Documented, clinically active, uncoded: 23 percent of risk-relevant conditions, concentrated exactly where the demo chart had hinted, in specialist notes for multi-system chronic patients. The recapture, executed through their existing coding workflow with evidence attached, repriced the year by seven figures. Not new care. Not new documentation. The same notes, finally read all the way.

TensorLabs builds ontology-grounded extraction for clinical operators who suspect their notes know more than their claims. If your coding quality checks are periodic and manual, the gap is compounding between audits. Tell us which payment models you sit in, and we will tell you where the documentation-to-code gap usually hides in them.