How Do You Automate Medical Coding (CPT and ICD-10) in a Practice?

We have written a lot about the front and back of the revenue cycle. Charge capture, claims, denials, payments, patient balances, scheduling. There is a step sitting right in the middle of all of it that gets asked about more than almost anything else, and it is the one people get sold the hardest on. Coding. Specifically, can a machine read the note and assign the CPT, ICD-10, and modifiers correctly, so a coder does not have to touch every chart.

The short answer is that yes, autonomous coding is a real, working category now, and for the right kind of practice it saves real money. The longer answer is that the headline accuracy numbers you will see quoted are, in a very specific way, not the whole story. If you only remember one thing from this article, remember this: automation rate is not accuracy rate, and the gap between them is where practices get disappointed.

Coding is not the same job as charge capture

First, a distinction that trips people up. Charge capture answers “did this service get billed at all.” Coding answers “was it billed with the right codes.” A visit can be captured perfectly and still coded wrong, and a perfectly coded chart is useless if the charge never made it out the door. They are two different failure modes and they automate differently. This article is about the second one, the actual assignment of CPT, ICD-10, HCPCS, E/M levels, and modifiers from the clinical documentation.

Why does this matter so much right now? Because the people who do this work are getting harder and harder to find. When MGMA polled group leaders about their hardest roles to fill, medical coders topped the list at 34 percent. The Bureau of Labor Statistics projects roughly 7 percent growth in medical records specialist jobs from 2024 to 2034, with about 14,200 openings a year, and the supply of credentialed coders is not keeping pace. An aging workforce, burnout, and steady turnover mean seats stay open. Every open seat is charts that do not get worked.

What an open coder seat actually costs you

The cost of that gap does not show up as a denial. It shows up as delay. In hospital finance the metric is DNFB, discharged not final billed, the stack of encounters that are done but cannot be billed because they have not been coded. In a clinic it is the same idea by a different name: the charts sitting in a coding queue. Every day a chart sits there is a day added to your days in A/R, which is one of the most watched numbers in the whole revenue cycle. Coding backlogs push A/R up directly. Cash you have earned just sits on the wrong side of the ledger.

So the pitch for autonomous coding is genuinely compelling. Feed the AI the note, it reads the documentation, applies coding guidelines and payer rules, and assigns codes with no human touch on the cases it is confident about. On high-volume, structured encounters this works well. The strongest real-world adoption, and the first KLAS “Autonomous Coding 2025” report backs this up, is in radiology and the emergency department, where documentation is relatively uniform and the code sets are more predictable.

The number that hides the truth

Here is the catch. Vendors love to quote accuracy. You will see 95 percent, 97 percent, even 98 percent. Those numbers are usually real. But read them carefully, because that accuracy is measured on the subset of charts the system chose to code by itself. The high-confidence ones. The system quietly sets aside the messy charts and routes them to a human, and those do not count against its accuracy score.

That is why automation rate matters as much as accuracy rate. Automation rate is the percentage of charts the AI actually handles start to finish. Your real, effective result is roughly automation rate multiplied by subset accuracy. A tool that is 98 percent accurate on the 40 percent of charts it will touch is doing very different work than one that is 95 percent accurate across 90 percent of your volume. One vendor deployment we saw reported by Fathom at a group called Your Health cited a 95.5 percent automation rate at 98.3 percent accuracy across all service lines. That combination, high automation and high accuracy together, is the thing worth paying for. Treat any single number in isolation as marketing. Vendor-reported figures like these should be read as illustrative of what is possible in the right setting, not as a guarantee for yours.

Where it works and where a human still wins

So where does autonomous coding genuinely earn its keep today? High-volume, structured, single-specialty work. Radiology reads. ED visits. Routine outpatient encounters. Repetitive procedure coding. Reported ranges for these sit around 92 to 97 percent accuracy on structured encounters, which is competitive with, and often more consistent than, human coders. Worth knowing: inter-coder agreement among credentialed humans on detailed ICD-10 assignment is often only around 82 to 86 percent, so “match a human perfectly” is a lower bar than it sounds. Machines do not get tired, and they apply the same guideline the same way every time.

Where the human still wins is complex inpatient coding, encounters with multiple comorbidities, unusual specialty cases, and anything where the documentation is thin or contradictory. Reported accuracy on complex inpatient work drops into the 82 to 90 percent range, which is exactly why good implementations route those cases to a person. The right model is hybrid: the AI clears the routine volume instantly, and your credentialed coders spend their time on the hard charts, audits, and clinical documentation improvement instead of grinding through radiology reads all day.

The three things that decide whether it works for you

Before you buy anything, three practical checks. First, documentation quality gates everything. AI codes what the note says, and if your providers document loosely, the AI either guesses or kicks the chart to a human, and your automation rate collapses. Fix documentation first or fix it alongside. Second, you need a real audit trail and compliance oversight. Coding is where fraud and abuse exposure lives, so every automated code needs to be traceable and reviewable, and a human owns the compliance line. Third, ask specifically for automation rate and subset accuracy for your specialty and your payer mix, not a generic case study. If a vendor will only give you one number, that is your answer.

Finally, be honest about denials. Cleaner coding does reduce coding-related denials, and some radiology and ED deployments report large drops. But MGMA data still puts first-submission denials around 15 to 20 percent across all causes, and most denials are not coding errors. Coding automation fixes one lane. It does not fix eligibility, prior auth, or timely filing.

Where to start

If you are drowning in a coding backlog or cannot fill a coder seat, autonomous coding is one of the highest-return automations in the revenue cycle right now, as long as you buy it with your eyes open. Start by measuring your own DNFB and days in A/R, then find out how much of your volume is actually the structured, high-confidence kind that automates well. That single number usually tells you whether this is a quick win or a longer project.

That is exactly what our free 30-minute Waste Audit does. We look at where charts pile up, where cash is stuck, and which of your workflows are genuinely ready to automate versus which need a documentation fix first. If we cannot find real savings, there is no fee. If we do, we take 15 percent of what we document and you keep the rest, with no rip-and-replace of the systems you already run. Book one at /free-audit, see what we automate at /solutions, or read /how-it-works.