01 · claims-adjudication
Claims adjudication assistant
Classify this claim as Approved / Denied / Pending with
concise clinical and policy reasoning in our company voice.
Redact all PHI. Output { decision, reason, confidence }.
For US health insurers
Private AI for health plans. Compile once. Run forever. Zero PHI risk.
Compile claims adjudication, prior-auth review, fraud detection, member support, risk-adjustment coding, and provider credentialing into signed, auditable .kolm artifacts that run inside your environment. PHI stays inside the customer-hosted bridge your VPC. BAA countersigned within 48 hours.
HIPAA BAA
48-hour countersign
48-hour countersign
PHI mode
K-score gate ≥ 0.95
K-score gate ≥ 0.95
Customer-hosted bridge
data stays in your VPC
data stays in your VPC
HMAC receipt chain
append-only, verifiable
append-only, verifiable
SOC 2 Type II
in progress 2026-Q3
in progress 2026-Q3
Why health plans choose kolm
Three architectural facts your compliance officer actually wants to read.
PHI stays in your control boundary
The customer-hosted compile bridge runs in your environment. Member records, claim notes, EOB text — none of it touches kolm infrastructure. Only the manifest hash crosses the boundary, and only after PHI redaction has run inside your environment.
Audit-ready by design
Every artifact carries a signed receipt chain (HMAC-SHA256, 4 rings) and a K-score ≥ 0.95 in PHI mode. Export the full chain for NCQA, CMS, state DOI, or internal compliance review.
Up to 28× cheaper at claims-volume scale
One-time compile cost (your frontier-teacher bill) vs. per-call cloud inference forever. At a typical 1.2M claims/yr volume, payback is <8 weeks; perpetual local inference after that. Run the numbers.
Six ready templates for health plans
Each template ships with 8–12 anonymized examples, a held-out eval set, automatic PHI redaction, and a forced K-score gate of 0.95. Import to your private namespace in one click.
02 · prior-auth-review
Prior authorization reviewer
Review medical-necessity documentation and output a structured
prior-auth recommendation. Redact all PHI. Output
{ recommendation, criteria_met, redacted_summary }.
03 · fraud-detector
Fraud, waste & abuse detector
Analyze claim for FWA indicators. Flag and explain anomalies
while fully redacting PHI. Output
{ risk_level, flags, explanation }.
04 · member-support-triage
Member support triage
Answer member inquiry using only approved language. Never send
or store PHI. Route to human if confidence < 0.90. Output
{ response, route_to_human, confidence }.
05 · risk-adjustment-coder
Risk adjustment coding
Extract valid HCC codes from notes. Output structured list
with confidence per code. Redact PHI. Output
{ hcc_codes: [ { code, confidence } ] }.
06 · provider-credentialing
Provider credentialing assistant
Summarize credentialing documents and flag missing or expired
items. Redact PHI. Output
{ status, flags, summary }.
ROI calculator · tuned for claims volume
Compares one-time kolm compile cost vs. per-call cloud LLM inference at your annual claims volume. Real arithmetic, no marketing math.
—
frontier annual spend
—
kolm year-one total
—
year-one savings · multiple
Self-serve onboarding for health plans
Zero sales calls. Compliance officer + IT lead can go from "interested" to "first compiled artifact running in our VPC" in <15 minutes.
1Click Start Enterprise on /pricing · pay annually, monthly, or via Net-30 invoice
2E-sign the BAA (PHI Schedule pre-filled) · countersigned instantly on kolm letterhead
3Deploy the customer-hosted compile bridge to your VPC · one Docker / Helm / Terraform command
4Import any of the 6 templates · PHI mode auto-enables · first compile in <5 min
5-question HIPAA assessment → tells you in 60 seconds whether you need the BAA sign-first path or whether the free tier covers initial training.
What we will not pretend
SOC 2 Type II report is mid-audit (target 2026-Q3). HITRUST r2 self-assessment in progress. The customer-hosted compile bridge ships v1; full on-prem control plane is on the roadmap but not GA. K-score is kolm-specific and not yet a recognized regulatory standard — map it against your existing model-risk-management framework (e.g. NIST AI RMF, SR 11-7) and treat the gate as one input, not the only one. Full security posture →