The Problem
A loan officer reviews dozens of applications a day. Each one means manual credit scoring, income verification, and a lookup against the lending policy. It is slow, and worse, it is inconsistent — and inconsistency in lending is not just inefficiency, it is fair-lending risk. Under the Equal Credit Opportunity Act, you must be able to prove you treated similar applicants the same way. "The officer used their judgement" is exactly the answer that loses an ECOA examination.
Why a Black-Box LLM Fails Here
Feeding applications to an LLM and asking "approve or decline?" is the worst possible architecture for lending:
- Unexplainable decisions are illegal here. ECOA requires specific adverse-action reasons. "The model declined with 0.71 confidence" is not a reason a regulator accepts.
- Non-determinism is discriminatory by accident. If the same applicant can get different answers, you cannot prove consistent treatment — the core ECOA requirement.
- No audit trail of the rule. You cannot show which policy criterion drove a decision.
The Governed Architecture
- The LLM reads. It extracts structured fields from the application — income, credit score, requested amount, debt-to-income ratio, verification status. It never decides whether to approve.
- A DMN policy table decides. Your lending policy, expressed as a deterministic decision table. Every applicant with the same inputs gets the same output. That is mechanical proof of consistent treatment.
- BPMN routes. Approve, decline, or refer to a human underwriter, with the full audit trail.
Try It Live
Go to archiet.com/agents and open the Loan Adjudication agent. Type:
Applicant Sarah Chen. Annual income $87,000, verified. Credit score 712.
Requested loan amount $185,000 for a primary residence purchase.
Debt-to-income ratio 38%.Click Run governed agent:
1. The LLM extracts (shown read-only):
{
"applicant": "Sarah Chen",
"annual_income": 87000,
"income_verified": true,
"credit_score": 712,
"requested_amount": 185000,
"dti_ratio": 0.38,
"purpose": "primary_residence"
}2. The DMN policy table evaluates and fires: credit score 700–749 AND DTI < 40% AND income verified → approve up to $200,000 at standard rate. Rule ID LN-RULE-04.
3. BPMN routes to the approval path (requested $185,000 is within the $200,000 ceiling).
4. The audit trail records the firing rule, the exact conditions met, the LLM's extracted JSON, and the versioned policy table.
If a different applicant submits the same numbers, they get the same outcome — and you can prove it.
The Regulatory Angle: ECOA and EU AI Act Article 22
ECOA requires consistent treatment and specific adverse-action reasons. A DMN table delivers both: consistency is structural (deterministic rules), and the firing rule is the reason ("declined: DTI exceeded 43% threshold").
EU AI Act Article 22 (and credit scoring is explicitly an Annex III high-risk use) requires that automated decisions affecting individuals be explainable and subject to human oversight. The DMN table is the explanation; the BPMN human-referral path is the oversight. A neural-network credit model cannot offer either in a form a regulator will accept.
Generate This for Your Own System
Describe a lending or credit workflow in Archiet's Blueprint Wizard and the governed loan agent is generated with your policy thresholds extracted from your PRD — your score bands, your DTI limits, your loan ceilings. The decision engine, the policy model, the API endpoints, and the agent UI come wired and auditable.
The same governed pattern powers invoice approval, FNOL triage, prior authorisation, refund/retention, and KYC/AML in the gallery.