Loading…
Loading…
Loan approval decisions need to be explainable to regulators and customers. Archiet generates a governed loan adjudication agent: the BPMN workflow defines the process, the DMN credit policy table defines the approval criteria, and the AI only reads documents — it never decides the outcome.
Consumer lending is one of the most regulated industries for AI use. Under the Equal Credit Opportunity Act (ECOA) and Fair Housing Act (FHA) in the US, the Consumer Credit Sourcebook (CONC) in the UK, and the EU AI Act Annex III (AI in creditworthiness assessment is high-risk), lenders must be able to explain credit decisions to applicants. A machine learning model that produces a credit score based on thousands of features cannot explain itself in the way regulations require. A governed agent with a transparent DMN policy table can: 'Your application was referred to a credit analyst because your DTI ratio of 44% exceeds our 43% threshold for conditional approval.'
The generated loan adjudication agent has three layers. AI extraction layer: the LLM reads the loan application, credit report, bank statements, and income verification documents and extracts structured fields — credit score, loan amount, debt-to-income ratio, employment duration, loan purpose, and collateral value. The AI does not evaluate these fields; it only reads them. DMN policy layer: a deterministic rule engine evaluates the extracted fields against your credit policy table. Rules are explicit: if credit score >= 740 and DTI <= 36% and loan amount <= $500K → auto-approve. If credit score < 580 → decline. Business analysts can edit these thresholds without touching code. BPMN routing layer: the DMN decision drives the process flow — auto-approval queue, senior underwriter assignment, credit committee escalation, or automated decline with regulatory reason letter.
The DMN policy table replaces the 'black box' with a transparent, auditable decision matrix. Rows represent conditions: minimum credit score, maximum DTI, loan amount range, employment duration, and collateral requirements. Each row maps to a decision: auto-approve, conditional approval, underwriter review, credit committee, or decline. The table is stored as a JSON file editable through the operator console. When your risk team adjusts credit policy — tightening DTI requirements after a market event, or adding a new product tier — they edit the JSON. No code deployment required. Every change is logged with timestamp and editor identity for the audit trail.
The generated lending agent produces documentation for ECOA/Regulation B adverse action notices, FCA CONC creditworthiness assessment records, and EU AI Act Article 13 transparency requirements. Every decline generates a structured adverse action notice citing the specific policy condition that triggered it — not a generic 'credit score' explanation. The audit trail records the matched policy rule, the input values, the output decision, and the approver's identity for every application. The AI governance pack includes the risk classification under EU AI Act Annex III, the model card for the extraction LLM, and the algorithmic impact assessment required for automated credit decisions.
A credit scoring model (FICO, VantageScore, or a proprietary ML model) produces a numerical score based on features the model learned from historical data — and cannot fully explain which features drove any individual score. The governed adjudication agent uses the credit score as an INPUT to a transparent policy table — the policy table decides the outcome, not the model. The decision is always explainable: 'Credit score 680 with DTI 41% → conditional approval per rule 3 of the credit policy.'
Yes. The DMN table supports product-level conditions: loan_purpose (mortgage, auto, personal, business), collateral_type (secured/unsecured), and loan_amount_tier. Separate rule rows handle each product's specific approval criteria. The operator console includes a product filter so underwriters can view and edit policy for a specific loan type without seeing all rules.
The default output (the final row of the DMN table) catches all applications that don't match a specific rule — typically routing to a credit analyst for manual assessment. This prevents silent failures: every application receives a routing decision, even if it's 'manual review.'
The generated BPMN process XML opens in any BPMN-compliant tool: Camunda, Flowable, or bpmn.io. The generated Python runtime includes webhook integration points for connecting to LOS platforms (Blend, nCino, Encompass) via API. The agent can be deployed as a sidecar service that receives applications from your existing LOS and returns structured decisions.
Most AI invoice automation is a black box — the model decides, you can't explain why. Archiet generates a governed invoice approval agent: the approval flow is BPMN (auditable), the decision logic is a DMN policy table your finance team can edit without code changes, and the AI only reads documents — it never decides the outcome.
Enterprise procurement teams need approval automation that can be explained to auditors, edited by business analysts, and trusted in regulated industries. Archiet generates a governed procurement agent: BPMN workflow, DMN policy table, bounded LLM for document reading, and a complete audit trail from day one.
Insurance claims triage requires speed, accuracy, and full auditability. Archiet generates a governed claims triage agent: FNOL intake, AI-powered data extraction from claim documents, deterministic DMN severity routing that actuaries control, and a complete audit trail for regulatory examination.
free plan. No credit card required. Generate your first compliant architecture blueprint in under 10 minutes.