Technical architecture of Arkangel AI for AI medical audit

Arkangel AI is organized as a three-layer adjudication architecture for Colombian medical claims: 27 administrative rules, 29 clinical rules and 42 financial rules. The layers run independently, then a consolidation step assigns the glosa causal required by Anexo Técnico 6 of Resolución 3047 de 2008.

Research source: Arkangel AI medical-audit architecture for Colombian claims

System flow

Claim package
  ├─ XML / CSV: RIPS, invoice and structured billing data
  ├─ PDF / Word: clinical record, authorizations and support documents
  ↓
Normalization and document indexing
  ↓
Independent audit layers
  ├─ Administrative layer: 27 DAMA UK rules
  ├─ Clinical layer: 29 PERT CLIN rules
  └─ Financial layer: 42 FIN CTR rules
  ↓
Consolidation by invoice line
  ↓
Causal assignment under Resolución 3047 de 2008
  ↓
Human auditor review and explicit release

No audit layer reads the result of another layer. That design prevents clinical reasoning from contaminating financial findings, and it prevents administrative defects from being converted into glosas without positive evidence.

Administrative layer

The administrative layer checks whether the claim package is auditable: patient identity, RIPS structure, authorization, support-document completeness and filing timeliness. If the problem is structural, the account can be returned or observed before clinical or financial adjudication.

Clinical layer

The clinical layer evaluates clinical pertinence, justification in the clinical record and consistency with Colombian Ministry of Health clinical practice guidelines. Causal 2 findings, such as a medication without indication, belong here because they require medical reasoning rather than only transaction logic.

Financial layer

The financial layer compares billed values with the contract, current tariff and benefit plan. It also checks patterns such as upcoding, unbundling and double billing across SOAT, EPS and ARL contexts.

Consolidation and governance

Each rule returns passed, failed or not applicable. A missing document is not automatically a glosa; it is recorded as an observation unless there is positive evidence of breach. The consolidation layer groups findings per invoice line and assigns the corresponding causal.

The human auditor always has the final word. Arkangel AI does not send provider communications autonomously, and every correction is logged with actor and timestamp.