AI Clinical
Reimbursement
System
A self-optimizing intelligence platform that eliminates the $300B+ administrative waste in healthcare claims processing — turning an adversarial system into a collaborative one.
Executive Summary
The Opportunity
Healthcare reimbursement is the financial backbone of a $4.5 trillion industry (CMS National Health Expenditure, 2023), yet it runs on processes designed in the 1990s. The global Revenue Cycle Management market reached $344 billion in 2024 (Grand View Research) — a market built almost entirely on patching a broken system rather than replacing it.
We propose the first self-optimizing AI clinical reimbursement platform — a system that doesn't just automate claims processing but learns from every decision, becoming more accurate with every claim it touches. This is the compound intelligence play: the same BRNZ DNA that powers KENSAI in cybersecurity, applied to the $344B healthcare admin market.
The $300B Problem
The US healthcare system spends more on administration than it does on heart disease and cancer care combined. Here's where the money goes.
Administrative complexity of multi-payer billing, eligibility verification, claims submission, and payment posting. The JAMA study (Tseng et al., 2018) estimated BIR costs at ~$265.6B annually.
The National Health Care Anti-Fraud Association (NHCAA) estimates fraud accounts for 3-10% of total health spending. At $4.5T total, that's $135-450B. FBI estimates conservatively at $100B+.
Each claim costs providers $40-80 in administrative labor — coding, submission, follow-up, appeals. With billions of claims annually, this is a trillion-dollar friction layer.
According to Change Healthcare's 2024 Revenue Cycle Denials Index, the average initial denial rate is approximately 17%. Some specialties see 20%+ denial rates on first submission.
The ICD-10-CM code set contains over 72,750 diagnosis codes. Add ICD-10-PCS (87,000+ procedure codes), CPT (10,000+), and HCPCS — and you have a coding nightmare for humans.
The US spends approximately 34% of total healthcare expenditure on administration — the highest of any OECD country. Canada spends 17%. The gap is the multi-payer complexity tax.
The Human Cost
- Treatment delays from prior auth — 94% of physicians report care delays (AMA, 2023)
- Surprise bills from denied claims patients thought were covered
- Patients forgo care entirely due to uncertainty about coverage
- Physicians spend 15+ hours/week on prior auth paperwork (AMA survey)
- Small practices hire 1 FTE per 2 physicians just for billing
- Revenue leakage from undercoding — estimated 5-10% of billable revenue
- Major payers employ 10,000+ claims processors each
- Fraud costs exceed $100B/year, requiring massive SIU departments
- Regulatory penalties for improper denials (CMS enforcement growing)
The Core Dysfunction: An Adversarial System
The current reimbursement system is fundamentally adversarial. Providers are incentivized to upcode (bill more); insurers are incentivized to deny (pay less). Both sides employ armies of specialists to fight each other. This is a zero-sum game that patients lose. Our system replaces adversarial dynamics with transparent, evidence-based decisioning that both sides can trust.
Product Architecture
End-to-end intelligence from clinical encounter to payment. Every step is AI-augmented, every decision is explainable, every outcome feeds back into the system.
EHR data ingestion via FHIR R4/HL7v2. Clinical notes, lab results, imaging reports, medication orders.
Clinical NLP extracts diagnoses & procedures. Maps to ICD-10-CM, CPT, HCPCS. Suggests modifiers. Confidence scoring.
Medical necessity check, bundling rules, modifier logic, coverage verification, duplicate detection. Fix before submit.
Policy rules engine + medical necessity AI. Real-time benefit verification. Fraud detection. Explainable decisions.
Auto-payment, ERA posting, patient responsibility calc. Every outcome feeds the self-optimization loop.
Core AI Modules
Eight interconnected AI modules, each specialized but sharing a common intelligence layer. Together they form a closed-loop system that improves with every claim.
Clinical NLP Engine
Reads unstructured clinical notes, discharge summaries, operative reports, pathology results, and radiology impressions. Extracts structured clinical concepts: diagnoses, procedures performed, medications prescribed, lab values, and clinical reasoning.
"Pt presents with acute exacerbation of COPD with hypoxic respiratory failure. CXR shows bilateral infiltrates c/w pneumonia. Started on BiPAP, IV abx (ceftriaxone + azithromycin), systemic steroids."
Dx: COPD exacerbation (J44.1), Pneumonia (J18.9), Hypoxic resp failure (J96.01)
Proc: BiPAP (94660), CXR (71046)
Rx: Ceftriaxone, Azithromycin, Steroids
Auto-Coder
Maps extracted clinical concepts to the correct billing codes across four code systems: ICD-10-CM (72,750+ diagnosis codes), ICD-10-PCS (87,000+ inpatient procedure codes), CPT (10,000+ outpatient procedure codes), and HCPCS Level II (7,000+ supply/service codes). Handles modifier selection (25, 59, 76, etc.), laterality, specificity level, and sequencing rules.
Primary Dx: J44.1 — COPD with acute exacerbation
Secondary: J18.9 — Pneumonia, unspecified
Secondary: J96.01 — Acute resp failure, hypoxic
CPT: 99223 — Initial hospital care, high complexity
CPT: 94660 — CPAP/BiPAP initiation
CPT: 71046 — Chest X-ray, 2 views
Confidence: 97.3% | Flag: Review J18.9 specificity
Pre-Validation Engine
The denial prevention layer. Before any claim is submitted, this module runs 200+ validation checks: medical necessity against payer-specific LCD/NCD policies, NCCI bundling edits, modifier appropriateness, timely filing limits, duplicate claim detection, patient eligibility/benefit verification, and place-of-service rules. Goal: reduce first-pass denial rate from ~17% to <2%.
Adjudication AI
The insurer-side brain. Applies payer-specific policy rules, medical necessity criteria (InterQual/MCG equivalent), benefit plan verification, coordination of benefits, and payment calculation (fee schedule, allowed amounts, deductible/coinsurance application). Every decision includes an explainability layer with specific policy citations — critical for regulatory compliance and provider trust.
Claim: 2025-CLM-847291 | Status: APPROVED
Allowed: $4,287.00 | Patient resp: $428.70 (10% coinsurance)
Reason: Medical necessity met per LCD L35062 (COPD acute exacerbation criteria). BiPAP documented as medically necessary per O2 sat <88% on arrival. Chest X-ray appropriate for pneumonia workup.
Confidence: 99.1% | Processing time: 1.8s
Fraud Detection
Multi-layered fraud detection combining anomaly detection (statistical deviation from peer benchmarks), network analysis (identifying collusion rings between providers, patients, and facilities), temporal pattern analysis (impossible billing patterns like 26-hour days), and clinical plausibility scoring (does the treatment match the diagnosis?). Estimated to catch $3-5 in fraud for every $1 spent on detection.
Prior Auth Bot
Automatically generates prior authorization requests by extracting clinical justification from the medical record, mapping to payer-specific criteria, and predicting approval probability. If probability is <70%, suggests additional documentation or alternative treatments. Integrates with emerging CMS Interoperability rules (CMS-0057-F) mandating electronic prior auth by January 2027.
Appeals Engine
When claims are denied, this module analyzes the denial reason, pulls supporting clinical evidence from the medical record, identifies relevant clinical guidelines (AMA, specialty societies), and generates a structured appeal letter with evidence citations. Predicts overturn probability based on payer, denial reason, and historical patterns. Average appeal success rate target: 70%+ (vs industry average of ~50%).
Self-Optimization Loop
The compound intelligence layer. Every claim outcome — approved, denied, appealed, overturned — feeds back into every other module. The auto-coder learns which codes get denied by specific payers. The pre-validator learns new denial patterns before they're documented. The fraud detector refines its anomaly thresholds. The appeals engine learns which arguments succeed.
This is the moat. After processing 1M claims, the system is categorically different from one that processed 100K. After 10M claims, it's untouchable. Same principle as KENSAI — self-optimizing pipelines via DSPy that get better with scale.
Technology Stack
Purpose-built for healthcare AI. Every technology choice serves accuracy, explainability, compliance, and self-optimization.
Clinical NLP & LLMs
- ▸ Fine-tuned medical LLMs (e.g., based on Llama 3 / Mixtral) trained on MIMIC-IV, clinical notes corpora
- ▸ BiomedBERT / ClinicalBERT for entity extraction and relation mapping
- ▸ DSPy for self-optimizing prompt pipelines — the BRNZ DNA
- ▸ Retrieval-augmented generation (RAG) over payer policies, CMS guidelines, clinical criteria
Medical Knowledge Graph
- ▸ SNOMED CT — 350,000+ clinical concepts with hierarchical relationships
- ▸ UMLS Metathesaurus for cross-ontology mapping (ICD-10 ↔ SNOMED ↔ CPT)
- ▸ RxNorm for medication normalization and drug interaction data
- ▸ Neo4j/TigerGraph for graph-based traversal and fraud network analysis
Data & Integration
- ▸ FHIR R4 / HL7v2 for EHR integration (Epic, Cerner, Athenahealth)
- ▸ X12 EDI (837/835/276/277) for claims transmission
- ▸ Apache Kafka / Flink for real-time claim streaming at scale
- ▸ PostgreSQL + TimescaleDB for transactional + time-series data
Explainability & Compliance
- ▸ SHAP / LIME for model interpretability on every claim decision
- ▸ Audit trail with immutable logging (every decision traceable)
- ▸ HITRUST / SOC 2 Type II compliant infrastructure
- ▸ Zero-trust architecture, end-to-end encryption, PHI segmentation
Competitive Landscape
The RCM market is massive but fragmented. No player offers true end-to-end AI with self-optimization. Most are workflow tools with rule-based automation bolted on.
| Company | Valuation / Revenue | Focus | AI Depth | Self-Optimizing? |
|---|---|---|---|---|
| Change Healthcare (Optum) | Acquired by UHG for $13B (2022) | Full RCM, clearinghouse, analytics | Medium — Rules + basic ML | No |
| Waystar | IPO 2024, ~$3.7B valuation | RCM platform, payment management | Medium — Predictive analytics | No |
| Availity | ~$1.5B (est.) | Payer-provider data exchange | Low — Workflow automation | No |
| Olive AI | Raised $848M, collapsed 2023-24 | Healthcare RPA / AI automation | Medium — RPA + ML | No — Failed |
| Aidoc | $250M+ raised, $1.3B valuation | Radiology AI triage | High — Deep learning | Partial — Clinical only |
| Viz.ai | $250M+ raised, $1.2B valuation | Clinical AI coordination | High — Deep learning | Partial — Clinical only |
| R1 RCM (Acclara) | Taken private for $8.9B (2024) | End-to-end RCM outsourcing | Medium — Analytics + labor | No |
| BRNZ (Ours) | Pre-revenue / Concept | End-to-end AI reimbursement | Highest — LLM + DSPy | Yes — Core DNA |
Why No Competitor Is Self-Optimizing
Rule-based systems (Change Healthcare, Waystar, R1) encode human knowledge as static rules. When CMS updates a policy, a human must update the rules. They don't learn from outcomes — they only do what they're told.
Olive AI's failure ($848M raised, collapsed) is instructive: they tried to automate healthcare workflows with RPA (robotic process automation), which is fundamentally brittle. Screen-scraping and click-automation breaks when UIs change. Real AI requires understanding clinical semantics, not clicking buttons.
Clinical AI companies (Aidoc, Viz.ai) build excellent deep learning for specific clinical use cases (radiology, stroke detection) but don't touch the reimbursement pipeline. They're clinical tools, not financial infrastructure.
Our approach is fundamentally different: DSPy-powered self-optimizing pipelines that treat every claim decision as a training signal. This is the same compound intelligence architecture that makes KENSAI's security scanning improve with every scan. No competitor has this DNA.
Market Opportunity
Multiple overlapping markets, each massive in its own right. The convergence of AI capability and regulatory mandate creates a once-in-a-generation window.
Revenue Cycle Management encompasses claims processing, billing, collections, and payment management across all healthcare settings.
Total US healthcare spending, of which ~34% ($1.5T+) goes to administrative functions. Every dollar of admin spending is addressable.
Healthcare IT including EHR, RCM, analytics, telehealth, and clinical systems. Growing at ~15% CAGR driven by AI adoption and regulatory mandates.
CMS mandated electronic prior auth by January 2027. This creates forced adoption of automation tools across all Medicare Advantage and Medicaid plans.
Professional medical coding services, outsourcing, and coding software. Chronic shortage of certified coders (AAPC estimates 30% shortage) drives AI adoption.
FBI estimates $100B+ in annual healthcare fraud. AI fraud detection saves $3-5 per dollar invested. This alone justifies the platform for insurers.
Regulatory Tailwinds
Mandates electronic prior authorization for Medicare Advantage and Medicaid managed care plans by January 1, 2027. Payers must respond within 72 hours for urgent requests, 7 days for standard. This creates forced adoption of automated prior auth systems.
Requires price transparency and prohibits surprise billing. Creates demand for real-time eligibility/benefit verification and accurate cost estimation — both core capabilities of our platform.
CMS adds ~1,000-2,000 new ICD-10 codes annually. Each update creates coding errors and denials. Self-optimizing systems handle updates automatically; rule-based systems require manual reconfiguration.
The EHDS regulation (proposed 2022, expected adoption 2025) creates a framework for health data interoperability across EU member states — opening the door for AI reimbursement platforms in a $2T+ European healthcare market.
The Self-Optimizing Moat
BRNZ DNA — The compound intelligence play that makes every claim processed a permanent competitive advantage.
This is not incremental automation. This is a system that gets categorically better with scale, creating an exponential moat that no amount of engineering can replicate without the data.
What the System Learns
Scale = Moat
Like KENSAI in security: the more scans you run, the smarter the system becomes. A competitor starting from zero can't buy this data advantage.
Compliance & Ethics
Healthcare AI that makes financial decisions about patient care is the highest-stakes AI application outside of autonomous weapons. We take this seriously.
HIPAA Compliance
- ▸ All PHI encrypted at rest (AES-256) and in transit (TLS 1.3)
- ▸ BAA (Business Associate Agreement) with every customer
- ▸ Role-based access controls with minimum necessary standard
- ▸ Complete audit trail — every data access logged and retained 6 years
- ▸ HITRUST CSF certification target within 12 months
EU Health Data Regulations
- ▸ GDPR Article 9 — special category data (health) processing safeguards
- ▸ European Health Data Space (EHDS) readiness for cross-border interoperability
- ▸ EU AI Act compliance — healthcare AI classified as high-risk (Annex III)
- ▸ Data residency controls — EU data stays in EU
- ▸ Right to human review for any AI-made coverage decision
Explainability — Non-Negotiable
Every decision must be explainable to three audiences:
"Your claim was denied because the documentation did not meet LCD L35062 criteria. Specifically, O2 saturation was not documented. Adding this would likely result in approval."
"Your insurer reviewed your claim and found that the treatment is covered. Your share is $428.70 based on your 10% coinsurance. You have the right to appeal any decision."
"Decision made using model v3.2.1, trained on 2.4M claims. SHAP values: diagnosis code (0.34), procedure alignment (0.28), policy match (0.22). Full audit trail available."
Revenue Model
Multiple revenue streams, each aligned with customer value.
Per-Claim Processing Fee
Compared to $40-80/claim in current admin costs, even $12/claim represents 75-85% savings.
SaaS + Performance
$2,000-$25,000/month based on practice size. Includes dashboard, analytics, reporting.
15-25% of recovered revenue from successful appeals. Typical recovery: $50K-500K/year per mid-size practice.
Health systems and large payers: $500K-$5M annual license for on-premise or private cloud.
Unit Economics (Illustrative)
Go-to-Market Strategy
Start narrow, prove accuracy, expand with data advantage.
Beachhead: Mid-Size Provider Groups 100-500 physicians
Multi-specialty groups that feel the pain most. Large enough for volume (10K-50K claims/month), small enough to decide fast.
Expand: Health Systems & Hospitals 500-5,000 beds
Proven accuracy from Phase 1 becomes the sales weapon. Health systems process 100K-1M+ claims/month. Inpatient DRG coding is highest-value.
Cross the Aisle: Sell to Insurers Payer-side AI
The game-changing move. Once processing for providers, we have data to offer AI adjudication to insurers. Payers save billions.
Endgame: The Neutral Infrastructure Layer Platform play
Both sides use our AI. We become the intelligent clearinghouse — a neutral, trusted layer replacing adversarial dynamics. Think Visa for healthcare payments, but with AI intelligence.
Implementation Roadmap
Aggressive but achievable. Each phase delivers standalone value.
Auto-Coding + Pre-Validation MVP
- ▸ Clinical NLP engine (Module A)
- ▸ ICD-10-CM auto-coding (Module B, partial)
- ▸ CPT auto-coding for top 50 specialties
- ▸ Basic pre-validation (NCCI edits, bundling)
- ▸ FHIR R4 integration with 3 EHR platforms
- ▸ DSPy self-optimization pipeline v1
- ▸ HIPAA-compliant infrastructure
- ▸ 3-5 pilot customers, 10K claims processed
Prior Auth + Appeals Automation
- ▸ Prior Auth Bot (Module F)
- ▸ Appeals Engine (Module G)
- ▸ Full pre-validation with payer-specific rules
- ▸ Revenue share model launch
- ▸ Integration with top 5 clearinghouses
- ▸ Approval prediction model
- ▸ 15-20 customers, 100K claims processed
- ▸ Self-optimization showing measurable improvement
Adjudication AI + Fraud Detection
- ▸ Adjudication AI (Module D) for payers
- ▸ Fraud Detection (Module E)
- ▸ Explainability layer (SHAP/LIME)
- ▸ First payer pilot
- ▸ DRG assignment for inpatient coding
- ▸ Knowledge graph v2 (SNOMED + RxNorm)
- ▸ 50+ customers, 1M claims processed
- ▸ HITRUST CSF certification
Full Ecosystem — The Intelligent Clearinghouse
- ▸ Both-side platform (providers + payers)
- ▸ Real-time eligibility + benefit verification
- ▸ Patient cost estimation
- ▸ EU market expansion (EHDS compliance)
- ▸ API marketplace for third-party integrations
- ▸ 10M+ claims processed, compound intelligence moat
- ▸ SOC 2 Type II certification
- ▸ Series A / growth funding
Healthcare reimbursement is a $344B market running on 1990s technology.
The first self-optimizing AI wins it.
This isn't about making claims processing 10% faster. It's about replacing the adversarial, error-prone, fraud-riddled infrastructure with transparent AI that gets smarter with every claim. The compound intelligence moat makes the first mover the permanent winner.