Every risk adjustment leader knows the suspect-list problem. Your NLP tool flags a hundred possible HCCs. Your coders confirm a fraction of them. The rest sit in a queue, or worse, get captured without the evidence to back them up. More suspects did not make you more accurate. It made your team busier.
Accuracy that pays answers three questions. How much risk did you capture? How much of it can you defend when the auditor samples your members? And how much is your current process missing? NLP either moves those three numbers or it adds noise. Here is how to tell the difference.
The Fastest Test: Run It Against Charts You Already Coded
You do not have to take any vendor’s word on accuracy, including ours. Coder Workbench runs a vendor-agnostic second pass, which means it can read the output of the coding tool or vendor you already pay and show you what they missed and what they cannot prove. Same charts, same members, and at the end a number you can act on. That is the cleanest way to judge any coding tool, and it is how we suggest you judge this one.
Where Most NLP Tools Fall Short
Most tools are built to find suspects. They read the chart, match the language to codes, and hand your coders a longer list. That looks like progress, but it moves the work instead of shrinking it. Your coder still opens a 120-page chart and hunts for the one line that proves the suspect is real. When the tool flags a code but cannot show the evidence, the coder does the hardest part alone. Volume goes up. Accuracy stays flat.
A suspect without its proof is a guess with a code attached, and a guess does not survive an audit.
What NLP Has to Do to Lift Accuracy
Four things separate the tools that raise accuracy from the ones that add noise.
First, it surfaces the evidence, not just the code. The tool should point the coder to the exact line in the chart that supports the condition. Invent Health’s NLP detects that supporting evidence with 85 percent or higher accuracy out of the box, so the coder stops searching and starts deciding.
Second, it recommends and never assigns. A certified coder confirms the code, checks the evidence, and signs the chart. Every capture traces back to a person and a line in the chart, which is the trail an auditor actually follows.
Third, it reviews every chart twice. Inside Coder Workbench, each chart runs through a blind two-pass process. The first pass codes the chart. A second, independent pass validates what was captured without seeing the first result. That catches both kinds of error: the condition that was missed and the code that should not be there.
Fourth, it gets faster because of accuracy, not instead of it. When the evidence arrives with the suspect, coders move 30 to 40 percent faster. The chart hunt disappears. The judgment stays.
On the headline number: Coder Workbench averages 95 percent coding accuracy. What makes a capture defensible is not the score, though. It is that every code carries its evidence and a coder’s signature.
Can You Trust AI to Code HCCs?
This is the question every risk leader is really asking, and it deserves a straight answer with two parts.
You should not trust AI to code on its own. A tool that generates diagnoses by itself is a black box, and when an auditor asks why a code was captured, “the model decided” is not an answer you can defend. That is RADV exposure dressed up as efficiency.
You can trust AI that explains itself and leaves the decision with your coder. Invent Health’s AI surfaces what matters and shows the evidence for it, but it never invents a diagnosis and never assigns a code. Your certified coder makes the call. That is the design in one sentence: explainable suspecting, not black-box automation.
As our founder Rohit Bhardwaj puts it:
“A plan gets paid for the risk it can prove, judged on the quality it can show, and audited on the evidence it can produce.”
Rohit Bhardwaj Founder and CEO, Invent Health
Evidence-first NLP with a coder’s signature is how you prove all three.
How Coder Workbench Does It
Coder Workbench reads clinical documents, recommends ICD-10-CM codes, and shows the evidence behind each one. A certified coder reviews, confirms, and signs. It processes more than 10,000 documents a day and supports the models plans actually run: CMS-HCC V24 and V28, ESRD, RxHCC, and HHS-HCC for ACA, with new CMS and HHS models typically implemented within 30 to 60 days of advance notice.
A Signed Chart Is Not the End of the Loop
Here is the part most coding conversations skip. A confirmed, evidenced code still pays nothing until it lands on an encounter CMS accepts. If your coding tool, your analytics, and your encounter submissions live in three systems that do not talk to each other, you can run a flawless coding operation and still lose the code somewhere between the signed chart and the CMS response file.
That is why Coder Workbench runs beside Invent Health’s risk analytics and Encounter Submissions in one platform. Invent Health aims for more than 99 percent first-pass acceptance across Medicare and ACA from day one, so the codes your coders proved are the codes CMS takes. The chart, the evidence, the diagnosis, and the CMS acceptance live in one place, so capture accuracy gets measured where it actually pays: at acceptance, not at “sent.”
Why Defensible Capture Decides 2027 Payment
CMS has spent this year telling plans exactly where this is going. The CY 2027 Rate Announcement, finalized in April 2026, ends unlinked chart reviews starting with the 2027 payment year. The July HPMS memo on CY 2027 risk adjustment implementation lays out what plans need to have ready. And RADV audits now run every year for every plan.
Be precise about what an annual audit means. Auditors sample members and extrapolate what they find. One indefensible code in the sample does not cost you one code. It costs you that error rate applied across the population. At the same time, V28 lowered risk values, so fewer codes carry the revenue and every one of them has to hold.
The plans that come out ahead in 2027 are not the ones capturing the most. They are the ones capturing what they can prove, on encounters CMS accepts, with no audit scramble, because the evidence was attached from the start.
See It on Your Own Charts
The fastest way to judge a coding tool is to run it against charts you have already coded. See what it catches, what it can prove, and what your current process missed.
Frequently asked questions
How does NLP improve HCC capture accuracy?
NLP improves HCC capture accuracy by reading clinical charts, matching documentation to conditions, and surfacing the evidence behind each suspected code. The strongest tools point the coder to the exact supporting line and keep a coder in the loop to confirm it, which raises capture and defensibility together instead of trading one for the other.
What is HCC capture accuracy?
HCC capture accuracy measures how correctly a plan captures the conditions its members actually have, and whether each capture is documented and defensible. Under-capture leaves earned revenue behind. Over-capture creates audit exposure. High accuracy means solving both at once, not producing more codes for their own sake.
Can AI code HCCs on its own?
No. AI should recommend codes and surface evidence, but a certified coder should confirm and sign each one. Fully automated, black-box coding creates RADV audit risk, because captures cannot be traced to a human decision and a documented source. Coder-in-the-loop AI keeps accuracy and defensibility intact.
Is AI coding safe for RADV audits?
AI coding is audit-safe when every code traces back to documented evidence and a human decision. Tools that show the supporting chart line and keep a coder in the loop produce defensible captures. Black-box tools that auto-assign diagnoses do not. And because RADV findings are extrapolated across the population, a small number of indefensible codes can drive a large repayment.
What is coder-in-the-loop coding?
Coder-in-the-loop means AI surfaces suspected codes and the evidence for them, but a certified human coder makes the final call. The AI recommends, and the coder decides and signs. This keeps the speed of automation while protecting accuracy and audit defensibility, since every capture has a human owner.
What is a two-pass coding model?
A two-pass model reviews every chart twice. The first pass codes the chart. A second, independent pass validates what was captured without seeing the first result, which is why it is called blind. Together the two passes catch both misses and overcalls, which single-pass review often leaves behind.
How accurate is Invent Health’s Coder Workbench?
Coder Workbench averages 95 percent coding accuracy, with NLP evidence detection at 85 percent or higher out of the box. Because a certified coder confirms every code against its evidence, captures stay defensible in audit. Plans also use its vendor-agnostic second pass to validate the output of coding tools they already run.
How much faster is NLP-assisted coding?
When evidence is surfaced alongside each suspected code, coders move 30 to 40 percent faster. The speed comes from removing the chart hunt, not from cutting review steps. Coders spend their time deciding instead of searching, so throughput and accuracy rise together.
Which risk models does Coder Workbench support?
Coder Workbench supports the models health plans run across lines of business, including CMS-HCC V24 and V28, ESRD, RxHCC, and HHS-HCC for ACA populations. New CMS and HHS models are typically implemented within 30 to 60 days of advance notice.
Does a confirmed HCC always count toward payment?
No. A confirmed code only counts once it lands on an encounter that CMS accepts. Codes tied to rejected or voided encounters never reach the risk score, no matter how well they were documented. That is why capture accuracy should be measured at CMS acceptance, and why Coder Workbench runs beside encounter submission and risk analytics in one platform. Invent Health aims for more than 99 percent first-pass acceptance across Medicare and ACA, so accepted codes are the norm.
Does NLP replace medical coders?
No. NLP removes the manual chart hunt and recommends codes with evidence, but certified coders still confirm and sign every capture. The goal is to make skilled coders faster and more accurate, not to take human judgment out of a process that CMS audits on documentation.
