ASC orthopedic AI coding pilot: 79.5% to 86.1% on the client's own QA, scored by their auditor, not by us
ASP-RCM engagement records
The client's QA team scored every code; we did not grade our own work
An orthopedic and pain ASC with its own coding team and QA auditor; the engine coded live operative notes alongside the coders.
Measured on the production engine
- QA accuracy 79.5% to 86.1%, scored by the client
- Auditor reconciliation on 184 and 230 charts
- Descriptor and laterality checks, 0 false positives
- Knowledge base and payment-indicator match
Measured on a substitute model
- The 75-chart, three-day production run (CPT 97%, device 78%, ICD precision 35 to 54%)
- Its timing: 111 s per chart, about 2.4 charts a minute at 6 concurrent
- 1Jul 16, 2026Pilot opens: old engine coded alongside the coding team, QA team scores
- 2Jul 2026Old engine baseline: 287 errors on 1,402 codes, 79.5%
- 3Aug 2026Upgraded engine: 50 errors on 360 codes, 86.1%
- 4Aug 2026Auditor reconciliation: 93.6% defect-free lines on 184 charts; 0 upcoding on 230
- 5Aug 24 to 26Deterministic descriptor and laterality checks added: 7 findings, 0 false positives
- 6Aug 21, 2026Pilot scoring closes; production run on a substitute model follows
Strong where the dollars are decided, weak on secondary diagnoses
Accuracy by category on the upgraded engine, client's QA scoring, July to August 2026
Pilot stage. Procedure codes, payment indicators and implants above 95 percent; diagnoses pulled the total down.
The engine is strongest exactly where the dollars are decided in an ASC: the procedure code, the payment indicator and the implant. Its weakest category, secondary diagnosis at 53.5 percent, is the one a coder must still own.
Disagreements went to the auditor, not to a vote
Auditor verdicts, August 2026
An earlier 59.5 percent defect-free figure was a scoring inversion on our side; the corrected figure is 93.6 percent.
| Pass | Charts | Result | Upcoding |
|---|---|---|---|
| Auditor reconciliation | 184 | 380 of 406 lines defect-free (93.6%) | Not scored separately |
| Second audit pass | 230 | 416 of 481 lines upheld (86%) | 0 findings |
| Coder-proposed changes | within the above | Auditor sided with the engine 280 of 298 (94%) | n/a |
75 real charts, on a substitute model
After the pilot closed, 75 real charts ran over three production days against the coding team's final codes on a substitute language model, so these figures are not comparable with the QA pilot. Mean 111 seconds per chart, about 2.4 charts a minute at 6 concurrent; the substitution is disclosed because it changes the result.
Match against the coding team's final codes, substitute model
Five changes, one caught an error a human missed
- 1
Load the reference tables
- Code universe 121 to 274 procedures
- HCPCS device table 0 to 9,109 rows
- 122 of 126 ASC payment indicators matched to the quarterly addenda [1]
- 2
Add deterministic checks
- Descriptor and laterality rules run before any model output is shown
- 7 findings, 0 false positives on 136 to 184 audited charts
- 3
Route, do not auto-accept
- A routing layer with 7 checkers sends clean charts to a fast lane and flagged charts to a coder
- 0 false positives
- 4
Let the auditor arbitrate
- Every disagreement goes to the client's QA auditor
- Where coders proposed changes, the auditor upheld the engine 94% of the time
- 5
Hold the ICD list
- Secondary diagnoses shown as candidates with evidence sentences; the coder decides
- Precision of 35 to 54% is not good enough to bill from
| Reference | Before | After | Check |
|---|---|---|---|
| Procedure code universe | 121 | 274 | Client scope |
| HCPCS device table | 0 rows | 9,109 rows | Device match 0% to 78% |
| ASC payment indicators | n/a | 122 of 126 matched | Against the quarterly addenda [1] |
| Payment-indicator mismatches | n/a | 0 over 2,396 coder-final lines | Full book |
| Wrong-side claims caught | 0 | 1 | Missed by the human auditor |
Knowledge base and payment indicators, August 2026, ASP-RCM engagement records
A pilot is not a production outcome; four things remain unproven
- ICDDiagnosis precisionICD-10 precision of 35 to 54 percent means the engine proposes codes the chart does not support. Secondary diagnosis accuracy of 53.5 percent on the QA pilot points the same way.
- LLMSubstitute modelThe 75-chart production run did not use the pilot engine. Until it is repeated on the production model, its 97 percent CPT figure is indicative, not a result.
- 360Pilot scale360 codes scored on the upgraded engine is enough to see direction, not to publish a confidence interval. The 1,402-code baseline is the larger sample.
- $Cash impactNo denial-rate, days-to-bill or payment outcome has been measured for the center. Accuracy on QA scoring is the only outcome we can show.
Frequently asked questions
Who scored the pilot?
The surgery center's own QA team, line by line, against its coders' final codes. Disagreements between coders and the engine went to the client's auditor, who sided with the engine on 280 of 298 disputed lines.
Why does the page quote two different accuracy figures?
Because they are different tests. The client's QA scoring moved from 79.5 percent (287 errors on 1,402 codes) to 86.1 percent (50 errors on 360 codes). The separate 75-chart production run, on a substitute model, matched CPT on 89 of 92 codes but had diagnosis precision of only 35 to 54 percent.
Did the engine ever upcode?
Not in the 230-chart audit pass, which recorded 0 upcoding findings. The deterministic checks run in the other direction: they caught one wrong-side claim the human auditor had missed.
Is this a finished outcome?
No. It is pilot-stage. Some figures ran on a substitute model, no payment or denial outcome has been measured, and secondary diagnosis coding at 53.5 percent still needs a coder.
- CMS: ASC Payment Rates, Addenda (quarterly payment indicator files)
- CMS: National Correct Coding Initiative (NCCI) edits
Checked October 25, 2026. Engagement figures are from ASP-RCM records for the periods stated; rules change, confirm against the source before relying on them.
Run the same pilot on your operative notes
Your QA team keeps score. We show the category split, the auditor verdicts and the deterministic checks, and say which figures are pilot-stage.