There is no single good automation rate, because an automation rate is a ratio and the vendor chooses the denominator. A number is only comparable when it states four things: what went in the denominator, the accuracy floor it was measured at, what counts as a human touch, and the time window. Without those, the percentage is marketing.
The first question in a vendor conversation is definitional, not numerical. Our guide to measuring coding automation rate and accuracy sets out the same discipline.
The denominator does most of the work
Take one platform, one hospital, one month, and follow the charts.
Ten thousand charts go to the platform. Twelve hundred come back untouched, out of scope or missing documentation, leaving 8,800. The platform returns codes above its confidence threshold on 6,600 of those. A coder edits 900. The remaining 5,700 reach the bill with no human change.
| What you divide by | Calculation | Reported rate |
|---|---|---|
| Every chart released to the platform | 5,700 / 10,000 | 57% |
| Charts in scope for the platform | 5,700 / 8,800 | 65% |
| Charts in scope, counting any coded output | 6,600 / 8,800 | 75% |
| Charts the platform accepted | 6,600 / 6,600 | 100% |
Four defensible-sounding numbers from one month of identical work, and every one can be published without lying. Two vendors quoting 80 percent are often not describing comparable performance, which is why the rate alone cannot rank them.
Coding audits have the same problem, and AHIMA documented it
This is not a failure mode invented by AI vendors. Coding audits have always had it.
In a Journal of AHIMA analysis, Mary H. Stanfill ran one audit of 25 records through four accepted calculation methods. Non-weighted code-over-code diagnosis accuracy came out at 92.42 percent, weighted code-over-code at 87.68 percent, weighted record-over-record at 91.6 percent, and non-weighted record-over-record, the all-right-or-all-wrong method, at 28 percent.
Same records. Same errors. A spread of 28 to 92.42 percent. The article's conclusion applies cleanly to automation rate: results "may or may not be comparable depending on which method is used."
Automation rate is a dial, not a capability
Any platform with a confidence threshold can hit almost any automation rate you name. Lower the threshold and more charts clear it. Accuracy on the marginal charts falls, because the charts that were sitting just below the line were there for a reason.
That cuts both ways. A high rate at an undisclosed accuracy level tells you nothing good, and a modest rate at a high accuracy floor may be the better platform. Automation rate and accuracy are a pair quoted at a stated threshold, or they are not a measurement.
Measure against a real baseline. CMS reported a Medicare fee-for-service improper payment rate of 6.55 percent, or $28.83 billion, for fiscal year 2025, down from 7.66 percent the year before, and is explicit that this "is not a measure of fraud." Human coding does not run at zero error either, so the standard for automation is your own audited baseline, not perfection.
The four things a number has to state
| Element | The question to ask | Why it changes the number |
|---|---|---|
| Denominator | Which charts are in the base, and what was excluded before counting? | Out-of-scope exclusions moved the example above by 8 points on their own |
| Accuracy floor | At what audited accuracy was this rate achieved, and who audited it? | The rate is tunable against accuracy through the confidence threshold |
| Human touch | Does a chart still count as automated if a coder opened it, or only if nobody edited it? | Separates true direct-to-bill from coder-in-the-loop review |
| Window | Which months, and was the ramp period included? | Early months in a deployment run lower and are often excluded |
Chart mix drives the rest. Template-driven specialties with a narrow code range automate further than case mixes dominated by rare diagnoses, the pattern found by a 2026 systematic review of automated ICD-10 coding. A rate quoted for one health system's radiology volume does not transfer to your inpatient queue.
Define it before the pilot, not after
Write the denominator, accuracy floor, human-touch definition and measurement window into the pilot agreement before charts start moving. A rate defined after the results exist is defined to flatter them. Our 90-day implementation playbook covers how to structure that window.
Automation rate is an operational metric, not a compliance one. In Medicare Advantage, for example, OIG guidance issued in February 2026 notes that organizations certify submitted risk adjustment data as accurate, complete and truthful. No automation rate changes who signs.
This article is educational and does not constitute a coding determination for any specific patient encounter.
Related questions
Is 80 percent a good automation rate?
Unanswerable as asked. Eighty percent of charts in scope at an audited 95 percent accuracy is a different product from 80 percent of charts the platform accepted at an undisclosed accuracy. Ask which denominator produced the 80 before comparing it to anything.
How is automation rate different from accuracy?
Automation rate measures how much work the platform completed without human change. Accuracy measures whether the codes were right. A platform can raise one by lowering its confidence threshold, which usually costs it the other, so the two are only meaningful when quoted as a pair.
What should we measure in a pilot?
Automation rate and audited accuracy on the same chart population, segmented by chart type and specialty, with the ramp period reported separately rather than excluded. Segment reporting is what tells you whether a headline rate is carried by one easy chart type.
Does a higher automation rate reduce audit risk?
No. It changes where review effort goes. Responsibility for what is submitted still sits with the organization, so a higher rate raises the importance of documented sampling and review of automated output.
