Every AI coding vendor says explainable. The only definition that settles anything is the one an auditor uses. How to test the claim in a demo.
Explainable has stopped doing any work in this category. It appears in every AI coding deck, it survives every RFP, and it costs a vendor nothing to say, because there is no agreed definition it can be measured against.
There is one definition that settles the question, though, and it does not come from the technology side. It comes from the audit. A capture is explainable if, three years after it was made, your organization can state why the condition was captured in terms an auditor accepts. Everything else is a demo feature.
That standard is unusually easy to test, and most of the testing can be done in a first meeting.
What Explainability Has to Mean Here
Strip the term back to what an audit response actually requires and it resolves to four things a capture has to carry.
The source. Which document, from which encounter, on which date, supports this condition. Not the chart in general. The specific record the determination rests on.
The support within it. What in that document establishes the condition as monitored, evaluated, assessed, or treated. An auditor does not accept that the condition appears somewhere in a long record.
The decision. Who confirmed it, and when. A capture is a judgment, and judgments belong to people. This is the element most likely to be missing when a tool is described as explainable rather than auditable.
The outcome. Which encounter carried the diagnosis and what CMS did with it. A capture nobody can trace forward is one you cannot confirm ever counted, whatever the coding record says.
Explainability in risk adjustment is simply whether those four survive as a linked record, produced automatically as work happens, rather than reassembled on request years later.
Why a Score Cannot Carry That Weight
A confidence value is a statement about the model, not about the patient. It expresses how strongly a system associates a pattern with an outcome. It is genuinely useful for routing work, for triage, for deciding what a human looks at first.
It is not evidence, and the reason is structural rather than a matter of degree. Evidence points outward to a document a third party can read. A score points inward to the system that produced it. Raising the number does not move it any closer to being the first kind of thing. A capture supported by a very high score and nothing else is a capture supported by an assertion.
This distinction rarely matters in operation and matters entirely at audit, which is why it survives so long unexamined. A tool can run for two years producing scored recommendations that coders accept, and the weakness only appears when someone outside the organization asks a question the score cannot answer.
Want to see what a linked capture record looks like?
Four Questions That Settle It in a Demo
These are answerable live, on the vendor’s own sample data, and the answers are more informative than any accuracy figure.
Show me a capture and take me to what supports it. The useful part is what happens next. Does the tool take you to the source document and the supporting text, or does it describe the condition again in different words. A restatement is not a citation.
Who confirmed this one, and can I see that. A named person and a timestamp, attached to the capture rather than stored in a separate audit log that has to be joined to it later.
Show me one the tool got wrong. Every tool produces recommendations a coder rejects. What matters is whether the rejection is visible in the record and whether you can see why the recommendation was made. A system that only surfaces its accepted output is not showing you how it behaves.
Run it against fifty charts we have already coded. This is the only test that uses your population, your provider mix, and your documentation habits. What it catches that you missed is interesting. What it claims and then cannot point to a source for is the finding.
What It Costs When the Link Is Missing
RADV audits sample members and extrapolate findings across the contract. That mechanic is what turns an evidence problem into a financial one.
A removed capture in a sample does not cost you that capture. It contributes to an error rate applied across a population. So the exposure a coding approach creates is not proportional to how often it is wrong. It is proportional to how often it is wrong in a way nobody can explain, because an unexplained capture cannot be defended and a defended one usually survives.
The operational cost arrives earlier and is easier to overlook. When captures do not carry their own evidence, responding to an audit becomes an investigation rather than a retrieval: staff reconstructing years-old determinations under a deadline, in a record that was never built to answer the question. Organizations that have been through this once tend to describe the reconstruction work as the expensive part, ahead of any repayment.
How Invent Health Approaches It
Invent Health runs risk adjustment as one connected loop rather than three tools that integrate, and it starts with analytics, not coding.
Risk Analytics decides what is worth working. Suspect logic runs across historical, lab, pharmacy, comorbidity, and chart-derived signals, drawing on EMR data through CCD and FHIR. Conditions are ranked by how much they move the score and how likely they are to close, and internal capture is validated against CMS-recognized HCCs in the MOR.
The Coder Workbench confirms the condition is real. AI-assisted coding reads the clinical documentation, shows the supporting evidence, and recommends ICD-10-CM codes for a certified coder to confirm and sign. Invent Health’s NLP detects that supporting evidence with 85 percent or higher accuracy out of the box, and the Coder Workbench averages 95 percent coding accuracy. It is a two-pass, coder-in-the-loop model. The AI surfaces what matters and never assigns a code, a person makes the call, and every code stays linked to the chart for audit-defensible lineage.
Encounter Submissions makes sure it counts, validating 837P, 837I, and DME for Medicare EDPS and Edge Server XML for ACA, flagging missing and non-risk-eligible diagnoses before a file leaves, then reconciling MAO-002 and MAO-004 for Medicare and the Edge Server reports for ACA.
Because the three run on one platform, the four elements above hold as a single record rather than as four systems that have to be joined after the fact. That is the whole design argument, and it is the reason the evidence trail is a byproduct of the work rather than a project that starts when a sample arrives.
Why the Standard Rises in 2027
Two changes make an unexplained capture more expensive than it has been.
The CY 2027 Rate Announcement ends unlinked chart reviews beginning with the 2027 payment year, so a chart review diagnosis has to tie back to an encounter CMS accepted, and the July HPMS memo on CY 2027 risk adjustment implementation sets out what has to be in place. The linkage CMS is asking for is the same linkage an audit response needs, which means the two requirements now point at one capability instead of two.
V28 removed a large share of diagnosis codes from risk-adjustment mapping and re-based values, so fewer conditions carry payment and each surviving capture represents more revenue in both directions. Combined with RADV moving toward every contract every year, the cost of a capture that cannot be explained rises while the number of captures carrying your revenue falls.
An organization selecting an AI coding approach this year is deciding what it will be able to say in an audit three years from now. That is the whole of the explainability question, and it is not a technical one.
Test It on Charts You Have Already Coded
Point any AI coding tool at a period your team has already worked and ask it to produce the four elements for every capture it makes. What it can produce automatically, and what requires someone to go find, is the answer.
Frequently asked questions
What is explainable AI in risk adjustment?
In risk adjustment, an explainable capture is one that carries four things: the source document supporting the condition, the supporting content within it, the person who confirmed it and when, and the encounter CMS accepted. If those hold as a linked record, the capture can be defended years later.
Is a confidence score the same as evidence?
No. A confidence value describes how strongly a system associates a pattern with an outcome. Evidence points outward to a document a third party can read. A high score is still an assertion about the model rather than support a reviewer can independently examine.
How do you evaluate whether an AI coding tool is explainable?
Ask it to take you from a capture to the supporting documentation, show who confirmed it and when, show a recommendation a coder rejected, and run against charts you have already coded. Each is answerable in a demo and none requires trusting an accuracy figure.
What should an AI coding vendor be able to show in a demo?
A path from any capture to the specific document supporting it, the confirming coder and timestamp attached to that capture, and its own rejected recommendations. A restatement of the condition in different words is not a citation.
Why does explainability matter for RADV audits?
Because RADV findings are extrapolated across a contract, so a capture that cannot be explained contributes to an error rate applied to a population rather than costing only itself. Exposure tracks how often a system is wrong in ways nobody can explain.
What is coder-in-the-loop AI?
A design where AI surfaces conditions and their supporting evidence but a certified coder confirms and signs each capture. The AI recommends and the person decides, which preserves the human determination an audit response depends on.
Can AI assign HCC codes without a coder?
Technically yes, but the resulting captures carry no human determination, which is one of the four elements an audit response requires. Under extrapolation, captures that cannot be attributed to a person and a source create exposure disproportionate to their number.
What is the difference between explainable and auditable?
Explainable usually describes whether a system can indicate why it produced an output. Auditable describes whether the organization can reproduce that reasoning years later with a named human decision attached. In risk adjustment, only the second one is worth paying for.
How should accuracy figures be interpreted when evaluating AI coding?
As capability on the vendor’s test set rather than performance on your population. The meaningful test is running the tool against charts your team has already coded and comparing what it caught, what it overcalled, and what it could actually point to a source for.
Does the 2027 chart review change affect explainability requirements?
Yes, indirectly. Starting with the 2027 payment year, chart review diagnoses must tie back to an accepted encounter, and that linkage is the same one an audit response needs. Two separate requirements now point at a single capability.
What makes an audit response expensive when captures are not explained?
The reconstruction work. Staff rebuild years-old determinations under deadline from records never designed to answer the question, which typically costs more in labor and disruption than the repayment attached to any individual finding.
How does Invent Health handle AI explainability?
AI-assisted coding in the Coder Workbench shows the supporting evidence and recommends ICD-10-CM codes for a certified coder to confirm and sign, never assigning a code itself. Every code stays linked to the chart, and because coding, analytics, and submission share one platform, the trail holds through to the CMS outcome.
