Technical Documentation · Credit Underwriting AI V4.0
Full Agentic AI Architecture
Application Ingestion → Spreading → Scoring → Decision → Structuring → Fair Lending. Seven parallel sub-agents. Five peer source tools. Every loan type. Every regulation.
SR 11-7 Model Risk
ECOA · Reg B
HMDA · CRA
CECL Provisioning
PCATS · Durable DAG Orchestration
SBA SOP 50-10
FIRREA Appraisal
nCino · Encompass · Finastra · Baker Hill
The Platform Beneath · PCATS
PCATS Applied to Credit Underwriting
Everything on this page is the credit-domain instantiation of the CAIBots Perpetual Cognition & Autonomous Transaction System. Five planes govern every run; the seven-stage workflow below is the Coordination Plane's DAG, and every decision travels the closed loop — Intent → Cognition → Policy Gate → Execute → Validate — before the system rests. Candidate decisions are frozen and hashed before policy evaluation; model output never executes without passing the deterministic gate. The system law: "Cognition proposes, Control governs, State remembers."
CONTROL PLANE
Versioned, hashed credit policy · gate verdicts (allow / HITL / dual / committee / block) · authority matrix · kill switch per lane
INTELLIGENCE PLANE
Seven parallel sub-agents · deterministic math in code, models for judgment & narrative · candidate decisions as schema-validated JSON
COORDINATION PLANE
Durable workflow DAG — the 7 stages below · retries, timeouts, compensation, HITL gate insertion, SLA escalation
EXECUTION PLANE
LOS / notice / document writes with idempotency keys, read-back confirmation, lineage tags, compensation on partial failure
STATE PLANE
Hash-chained decision ledger · frozen decision snapshots · policy version registry · S(t+1) validation after every write
Framework note: agent-internal reasoning may use LLM frameworks (ReAct-style toolformers); cross-system execution runs on the durable workflow engine — external writes are never delegated to an LLM loop.
01
📨
Application Ingestion — Multi-Channel Credit Request
Credit Application Ingestion — Multi-Channel, Multi-Product
Online · Branch · Broker/Correspondent · RM-Originated · Renewal · Portfolio Monitoring · Secondary Market
▼
Credit requests arrive across seven channels with dramatically different data completeness and document sets. Digital consumer applications arrive structured with bureau pull permission — agent initiates bureau pull, income verification (Finicity/Plaid), and employment check immediately on submission. Commercial applications arrive with incomplete financial packages — agent identifies missing documents, sends automated information requests, and begins spreading available financials immediately. Broker/correspondent submissions arrive via LOS API (nCino, Encompass, Finastra, Baker Hill) — agent validates eligibility, runs initial AUS findings, and flags overlay policy exceptions. Portfolio monitoring triggers — covenant breach detected, financial statement past due, property inspection due, maturity approaching — initiate automated renewal or remediation workflow. All submissions normalized to CAIBots Credit Request Schema v2.
Digital Consumer — Immediate Bureau Pull
Online Business Application Portal
RM-Originated Commercial (LOS)
Broker / Correspondent (API)
Portfolio Monitoring Trigger
Renewal / Modification Request
Secondary Market Acquisition Review
LOS Integration
nCino · Encompass · Finastra · Baker Hill — All submissions normalized to CAIBots Credit Request Schema v2 before agent pipeline invocation. Application classified → borrower profile bound → loan type determined → complexity tier assigned → workflow instantiated.
↓ application classified · borrower profile bound · loan type determined · workflow instantiated
02
⚙
Credit Underwriting Agent Orchestrator
7 Parallel Sub-Agent Dispatch · Policy Gate · Fair Lending Monitor
Durable Workflow DAG (Coordination Plane) · Exception Tracking · Credit Committee Packet Assembly
▼
Seven specialized sub-agents execute in parallel from application receipt. The Orchestrator manages dispatch, resolves conflicts, enforces SLA timing, and maintains the full audit trail for SR 11-7 compliance.
FS · Financial Spreading
Financial Spreading Sub-Agent
Extracts and spreads financial statements (tax returns, CPA, management accounts). Calculates DSCR/FCCR/EBITDA, normalizes add-backs, builds 3-year trend. Global cash flow across all entities.
CS · Credit Scoring
Credit Scoring & Bureau Sub-Agent
Tri-merge bureau pull, FICO 10T/VantageScore, derogatory analysis, D&B Paydex/Experian Intelliscore, UCC search, judgments/liens, OFAC/bankruptcy check.
CV · Collateral
Collateral Valuation Sub-Agent
Orders/reviews appraisal, validates via AVM, calculates LTV/CLTV/DSCR, monitors market comparables for CRE. Environmental clearance, flood zone, FIRREA compliance.
RP · Risk & Policy
Risk & Policy Sub-Agent
Applies credit policy guardrails, calculates risk grade (1–10), flags policy exceptions, assesses industry/concentration risk, stress tests under rate and vacancy scenarios.
ST · Structuring (Roadmap)
Structuring & Pricing Sub-Agent
Designs loan structure — tenor, amortization, covenant package, pricing grid (SOFR/Prime spread), guaranty requirements. Benchmarks against peer deals and internal pricing history.
CO · Fair Lending
Fair Lending Sub-Agent (ECOA/HMDA)
Runs HMDA and ECOA disparate-treatment screening throughout decisioning (portfolio disparate-impact analytics: Phase 2). Monitors for prohibited basis proxies. Pricing disparity detection. Reg B adverse action generation. HMDA LAR auto-update.
CL · CECL (Roadmap)
CECL / Provisioning Sub-Agent
Calculates expected credit loss, assigns loss given default, probability of default, exposure at default for CECL reserve. PD/LGD/EAD cohort modeling by loan type and vintage.
Orchestration Framework
Durable DAG Orchestration · Policy Gate Enforcement — All 7 sub-agents dispatched simultaneously. Orchestrator enforces the orchestration-compute SLA (~3s target; external source I/O adds provider latency — see the two-clock runtime model in the demo), resolves inter-agent conflicts, and maintains an SR 11-7-aligned audit event log.
↓ application decomposed into parallel underwriting workstreams · loan type routing confirmed
03
◈
Loan Type Classification — Underwriting Methodology & Decision Routing
Product Classification · Policy Mapping · Approval Authority · Decision Routing
Consumer · SBA · CRE · C&I · ABL · Syndicated · Leveraged
▼
Loan type classification determines the entire underwriting methodology and approval authority matrix. Consumer/SBA loans use scorecard-driven decisioning with AUS findings — highest auto-decision eligibility. Commercial Real Estate requires property-level cash flow analysis, appraisal, environmental review — always credit officer review. C&I/ABL requires spreading, industry analysis, borrowing base calculation. Syndicated/Leveraged requires sponsor analysis, credit agreement review, EBITDA addback scrutiny.
Consumer / Retail
FICO-Driven · AUS (DU/LP)
Conforming/Non-QM · HELOC/2nd · Auto · Personal · Credit Card
30–70% auto-decision eligible
Reg B adverse action · HMDA reportable · ECOA monitoring
Instant decisioning via scorecard + policy gate
SBA / Small Business
SBSS Score · SOP 50-10
SBA 7(a) / 504 / Express — SBSS threshold ≥155
Personal financial statement · 3-year P&L spreading
UCC search · SBA eligibility · ETRAN submission
Guaranty percentage calculation · SBA Authorization
Commercial Real Estate
NOI / DSCR · FIRREA Appraisal
NOI/DSCR analysis · AVM validation · Lease roll analysis
Environmental (Phase I/II) · Sponsor guaranty · Stress test
Market study · Construction draw schedule
Always credit officer review — never auto-decision
C&I / ABL / Syndicated
Spreading · Borrowing Base · Credit Agreement
3-year financial spreading · EBITDA normalization
Borrowing base (A/R, inventory) · Field audit program
Leveraged lending guidance (DFAST/OCC) · Syndicate structure
Credit agreement review · Agent bank duties
✅
AUTO APPROVE
Score above threshold · Policy compliant · No exceptions · Consumer/SBA eligible · Fair lending clear · Instant decision · LOS auto-updated
⚠
ASSISTED REVIEW
Minor exceptions · Borderline risk grade · Agent + credit officer co-review · Full package assembled by Agent · Officer provides final judgment
🏛
COMMITTEE
Above officer authority · Material exceptions · Complex structure · Leveraged · Syndicated · Agent prepares full credit memo + presentation
❌
DECLINE
Policy ineligible · Risk grade below floor · Fraud flag · Auto-generate Reg B adverse action notice with FCRA reason codes within 30 days
↓ loan type classified · underwriting workstreams dispatched · all five source tools called in parallel
04
◎
Source Selection — Five Peer Underwriting Intelligence Tools
Vector RAG · API Calls · SQL · Knowledge Graph · Memory Store
All five tools dispatched in parallel for CRE/C&I/Syndicated · Consumer uses RAG + API + SQL · RAG pipeline lives entirely inside Vector RAG tool box
▼
▼ Entire RAG Pipeline Lives Inside Vector RAG Tool Box
The 5-step RAG pipeline (Chunking → Embedding → Vector Store → Retriever → Prompt Assembly) executes entirely within the Vector RAG peer tool at Source Selection. All five tools are peers called simultaneously — not sequential.
↓ all underwriting intelligence assembled · financial analysis complete · risk grade calculated · policy checked · fair lending tested
05
◐
Credit Analysis Assembly — Complete Underwriting Package
Spreading · Scoring · Structuring · Compliance — Risk Grade & Decision Recommendation
Five parallel assembly tracks · Financial · Credit · Collateral · Policy · Fair Lending
▼
Five parallel assembly tracks execute simultaneously. Financial Track: 3-year spreads normalized, DSCR/FCCR/EBITDA calculated with adjustments documented, global cash flow assembled across all entities, debt service coverage tested under stress scenarios. Credit Track: tri-merge bureau analyzed, risk grade assigned (1–10 scale), probability of default calculated, peer comparison benchmarked against RMA industry data. Collateral Track: AVM validated against appraisal (within policy tolerance), LTV/CLTV calculated, lien position confirmed, environmental clearance verified. Policy Track: every underwriting criterion checked against credit policy — exceptions identified, severity classified, required approver level determined. Fair Lending Track: ECOA-protected class screening applied, disparate impact test run, pricing disparity flagged — all clear required before decision output.
1.42×
DSCR
Min policy: 1.20×
3
Risk Grade
Pass · Watch: 6+
1
Exceptions
P1 · Officer req'd
3-year spreads complete
Global cash flow assembled
Risk grade assigned (1–10)
Policy exceptions tallied
Collateral adequacy confirmed
Lien priority verified
Fair lending screen: CLEAR
CECL provision calculated
Structure recommendation assembled
↓ complete underwriting package → LLM reasoning · credit memo generation · decision recommendation · structuring terms
06
🧠
LLM Reasoning — Credit Memo Generation & Loan Structuring
3-Stage Pipeline: Financial Narrative → Decision Recommendation → Complete Credit Package
Credit memo · Term sheet · Covenant package · Adverse action notice · CECL provision memo
▼
Stage 1 — Financial Narrative: Synthesize spreading results, ratio analysis, and trend into coherent financial narrative — identifying key credit strengths, concerns, and mitigants in the institution's voice. Industry peer benchmarking contextualized against RMA data. Stress scenario analysis quantified and explained. Stage 2 — Decision Recommendation: Formulate a specific, policy-grounded credit recommendation — approve/decline/counter-propose — with risk grade rationale, exception justification, conditions precedent, and required approval authority level. Stage 3 — Complete Package Generation: Produce the full credit deliverable suite — credit memo (institution template, committee-ready), term sheet draft, covenant package (financial maintenance + reporting + negative covenants calibrated to risk grade), Reg B adverse action notice (if applicable), and CECL provision memo.
Claude Sonnet (Anthropic) — Default
Primary reasoning model. Credit memo generation, decision narrative, structuring recommendation — grounded in institution templates and policy via retrieval. The platform is model-agnostic; institutions may designate an alternate primary.
Claude Opus (Anthropic) — Escalation Tier
Complex multi-entity global cash flow narratives, syndicated/leveraged analysis, and low-confidence escalations. Regulatory documentation drafting under the same retrieval grounding.
Llama 3 / Mistral (Open-Weight)
On-premise / VPC deployment option for institutions requiring strict data residency. Same retrieval grounding and policy-gate architecture; model choice never changes the governance path.
Institution Credit Memory — Retrieval, Not Training
The institution's credit culture — memo style, covenant construction norms, risk-appetite calibration, prior committee feedback — is carried in a governed retrieval layer, not in model weights. No institution data is used to train or fine-tune models. Adaptation happens through curated knowledge-base updates and offline, change-controlled evaluation under SR 11-7 — every change versioned, tested against a golden-file decision suite, and reversible.
↓ credit memo + decision recommendation + term sheet + covenant package + adverse action → output layer
07
✅
Credit Decision Output — Explainable, Auditable, Fair-Lending-Tested
Approval Gates · Booking · Post-Decision Portfolio Intelligence
Auto-decision lane · Credit officer workbench · Committee package · Covenant monitoring · CECL · HMDA LAR
▼
Outputs routed to three simultaneous destinations. Auto-Decision Lane — eligible consumer and SBA loans: policy compliant, fair lending cleared, no exceptions — approve/decline decision rendered instantly, Reg B notice auto-generated for declines, LOS updated, funding workflow initiated. Credit Officer / Committee Package — complete credit memo in institution template, term sheet draft, covenant package, exception approval memo, risk grade rationale, CECL provision memo — all assembled and ready. Post-Decision — loan agreement covenant monitoring activated, financial statement collection scheduled, HMDA LAR updated, CRA activity recorded, CECL reserve posted.
🔴 CREDIT OFFICER APPROVAL MANDATORY
CREDIT OFFICER APPROVAL IS MANDATORY for all non-auto-decisions. COMMITTEE required above officer authority. APPROVAL REQUIRED for all policy exceptions. Agent never commits the institution without human credit authority sign-off.
Auto-Decision Lane
Policy compliant · Fair lending clear
Instant Reg B adverse action (auto-decline)
LOS updated automatically
Funding workflow initiated
Credit Officer Package
Credit memo: institution template, committee-ready
Term sheet draft with pricing grid
Covenant package (maintenance + reporting + negative)
CECL provision memo
🔴 HITL Gates — Always Required
All non-auto-decisions: Credit Officer
Above officer authority: Committee
All policy exceptions: Approval required
Fair lending flag: FL Officer + Credit Officer
📊
Credit Officer Workbench
nCino / Abrigo / Sageworks integration
Committee-ready memo in institution template
Spreading with all adjustments documented
Exception log with severity scoring
CRE: 3 weeks → 4 days · C&I: 2 weeks → 3 days
⚖
Fair Lending & ECOA
Disparate treatment screen: every decision · DI analytics: Phase 2 (Q4 2026)
Pricing disparity: >25bps flag
Redlining screen: Geographic analysis
HMDA LAR: Auto-updated on every application
Quarterly fair lending report: Auto-generated
📈
Post-Booking Portfolio Intelligence
Covenant breach: Immediate alert + waiver workflow
Annual review: Auto-assembled at anniversary
CECL migration: Quarterly PD/LGD tracking by cohort
Watch list: Auto-identification from deterioration signals
Maturity ladder: 90-day early alert
07 · Performance Targets — Underwriting Velocity & Quality (design targets · validated during pilot)
4 days
was 3 weeks
CRE Underwriting
3 days
was 2 weeks
C&I / ABL
<4 min
was 1–3 days
Auto-Decision (Consumer/SBA)
70%+
Consumer/SBA eligible
Auto-Decision Rate
100%
on every decision
Disparate-Treatment Screen
80%
eliminated by Agent
Spreading Time Reduction
