On September 30, 2026, revenue cycle outsourcer Savista and autonomous-coding AI vendor Minset announced that their joint deployment had reached up to 93% fully autonomous medical coding in radiology — charts that require zero human touch — alongside 80% automation in evaluation and management (E&M) leveling. Both companies say accuracy held at 100% through initial chart audits, sustained afterward through ongoing 10% sample audits.
For coding and compliance teams fielding vendor pitches built on similar numbers, the announcement is a useful case study in how to read an autonomous-coding accuracy claim rather than just repeat it.
What Savista and Minset Announced
Minset’s product, mCoder, is the autonomous-coding layer inside a broader agentic revenue cycle platform that also includes denials management (mDenials) and patient engagement (m360), all built on what the company describes as a shared reasoning framework. The Savista-Minset partnership pairs that AI with Savista’s credentialed coders — the company cites an average of 7.5-plus years of coder experience — in what both companies frame as a “true human-in-the-loop” model: the AI handles the majority of touchless coding, and human coders provide quality-assurance review on the charts that remain.
Savista CEO Jan Grimm said the partnership “delivers faster turnaround, greater scale and consistent accuracy, without compromising compliance.” Minset CEO Trey McMillian framed it as proof that “autonomous AI can multiply expert team reach while preserving accuracy, oversight and trust healthcare demands.” Savista, which serves more than 800 clients across 49 states and has 35-plus years in RCM, says the deployment is live across multiple specialties with plans to expand to more specialties and client sites.
Why 93% Is a Specialty Number, Not a Portfolio Number
The headline figure — 93% fully autonomous coding — applies specifically to radiology. E&M leveling, a notoriously harder coding domain because it depends on clinical judgment about visit complexity, automated at 80%. That gap is informative on its own: radiology coding tends to map more cleanly from structured orders and discrete findings, while E&M requires weighing documentation against MDM or time-based criteria that still benefits from a trained coder’s judgment.
What “100% Audited Accuracy” Actually Describes
The companies describe accuracy as “initially validated through comprehensive chart audits,” then “sustained over time, scaling to ongoing 10% sample audits.” That’s a meaningful methodological detail for anyone evaluating the claim: a comprehensive audit at launch followed by a 10% ongoing sample is a defensible QA design, but it’s a sample, not a 100% continuous audit of every chart going forward. Coding managers comparing vendor claims should always ask what audit cadence and sample size sit behind an “accuracy” number, and whether that cadence holds steady as volume scales — a 10% sample audited at 100% accuracy on month one doesn’t guarantee the same result at month twelve across a larger, more varied chart mix.
The Role of the Credentialed Coder Baseline
Savista’s 7.5-plus-year average coder tenure is also doing quiet work in this announcement. The quality-assurance layer behind any “human-in-the-loop” autonomous-coding model is only as strong as the coders staffing it, and a partnership built around experienced credentialed coders reviewing the harder charts is a different risk profile than one where QA review is an afterthought staffed by whoever is available.
What Compliance Teams Should Actually Vet
Autonomous-coding accuracy claims are becoming a standard feature of vendor marketing, and most follow a similar shape: a high touchless-automation percentage paired with a round, reassuring accuracy figure. Before taking either number at face value, compliance and coding leadership should ask:
- What counts as “touched” versus “touchless” — does a coder glancing at a chart for five seconds count as human-in-the-loop, or only a full review?
- What’s the audit sample size and cadence behind the accuracy figure, and is it disclosed publicly or only on request?
- Does the accuracy rate hold consistently across specialties, payer mix, and chart complexity, or only in the cleanest subset of cases?
- Who owns audit findings when the AI and a human reviewer disagree, and how is that logged for a payer or RADV audit trail?
- Does the vendor’s automation rate apply to the full chart volume, or only to a pre-filtered subset that excludes the hardest cases?
None of these questions are a knock on the Savista-Minset numbers specifically — both companies disclosed more methodology than many competing announcements do. They’re the standard due-diligence list any health system or RCM leader should run against any autonomous-coding vendor claim, this one included.
The Bigger Pattern: Specialty-by-Specialty Autonomy
The Savista-Minset split between radiology (93%) and E&M (80%) reflects a pattern showing up across the autonomous-coding market generally: vendors are getting comfortable publishing automation rates by specialty rather than a single blended number, because the honest answer is that autonomy varies widely by chart type. That’s a healthier trend for buyers than an undifferentiated “we automate X% of coding” claim, because it lets a health system map a vendor’s actual strengths against its own payer mix and specialty volume rather than buying a headline number that may not apply to the charts that matter most to that system.
What This Means for Teams Evaluating Autonomous Coding
The practical takeaway for coding and compliance leaders isn’t whether 93% or 80% is impressive — it’s that specialty-level transparency and disclosed audit methodology are becoming the baseline for a credible autonomous-coding vendor pitch, not a bonus. Vendors willing to publish per-specialty automation rates and describe their audit cadence, as Savista and Minset did here, give buyers something concrete to diligence. Vendors offering only a blended headline accuracy number without that detail should draw more scrutiny, not less.
Health systems and RCM teams weighing a build-versus-buy or vendor-versus-vendor decision on autonomous coding can apply that same standard internally: track automation and accuracy by specialty, document audit sample size and cadence, and treat any single blended percentage as a starting question rather than a finished answer. That’s the same discipline behind Medikode’s automated medical coding platform, which is built to keep coders and compliance teams in the loop with auditable, chart-level visibility into every AI-assigned code rather than a black-box accuracy score.