Confidential , Internal Document

AI Systems Implementation Plan

Phase 1 & 2 Proposal
Geneva Financial, LLC

AI-powered underwriting assistance and document intelligence for internal operations.

Dhairya Kamalia
April 2026
Draft , Pending Requirements
v0.1
!
All cost figures are indicative estimates based on initial scoping. Final numbers depend on exact requirements, loan volume, and API usage. Key inputs are editable in the Cost Calculator section below.

Executive Summary

This document outlines the initial two phases of an AI integration initiative for Geneva Financial. The goal is to layer intelligent automation on top of the existing Encompass LOS and nCino infrastructure , reducing manual effort in the underwriting and document review workflows while keeping loan officers in control of every final decision.

Phase 1 focuses on building an AI-powered underwriting assistant that pulls borrower data from Encompass, analyzes eligibility across loan programs, and surfaces a clear recommendation with reasoning. Phase 2 extends this with a document intelligence pipeline that automatically parses, extracts, and validates borrower-submitted documents before a human touches them.


1

AI Underwriting Assistant

Intelligent loan eligibility analysis powered by Encompass data

The system connects to Encompass via API, pulls the full structured borrower profile , income, credit score, DTI ratio, employment history, loan type requested , and passes it through an AI reasoning layer. The output is a plain-English recommendation that tells the loan officer whether the borrower qualifies, for which program (Conventional, FHA, VA, USDA), and under what conditions. Everything stays inside their existing workflow.

System Flow
Phase 1 system flow: Encompass API to LO dashboard Four steps: Encompass API pulls loan data, data extraction cleans and structures it, AI reasoning layer runs GPT-4 analysis, then outputs to the LO dashboard as a recommendation. Encompass API borrower data pull Data extraction clean + structure AI reasoning layer GPT-4 analysis LO dashboard recommendation
Estimated Cost Breakdown
$200 – $350
OpenAI, scales with volume
$150 – $300
Hosting, monitoring
Estimated Timeline
API integration
2–3 weeks
AI reasoning layer
3–5 weeks
Dashboard / UI
2–3 weeks
Testing + deploy
1–2 weeks

Total estimated delivery: 14–20 weeks with 1 developer, or 7–10 weeks with 2 developers working in parallel.

2

Document Intelligence Pipeline

Automated parsing, extraction, and validation of borrower documents

Borrowers upload income docs, bank statements, tax returns, and pay stubs through the nCino app. Right now, someone on the ops team manually reviews all of it. This phase builds an AI pipeline that reads every uploaded document, extracts the relevant data fields, cross-references them against what Encompass already has on file, and flags any discrepancies , before a human looks at anything. Based on prior experience, document processing time can be reduced by 40–60%.

System Flow
Phase 2 system flow: nCino upload to ops alert Four steps: nCino upload receives borrower documents, PDF parsing extracts structured fields, AI cross-check validates against Encompass data, then flags discrepancies or approves for ops team. nCino upload borrower documents PDF parsing extract fields AI cross-check vs Encompass data Flag or approve ops team alert
Estimated Cost Breakdown
$400 – $750
Vision + extraction APIs
$80 – $180
Document storage + DB
Estimated Timeline
nCino integration
2–3 weeks
PDF parsing engine
3–4 weeks
AI extraction layer
3–4 weeks
Compliance checks
2 weeks
Testing + deploy
1–2 weeks

Total estimated delivery: 16–22 weeks with 1 developer, or 9–13 weeks with 2 developers. Can start in parallel with Phase 1 after week 4.

Year 1 Cost Projection

Combined estimate across both phases, including development and 12 months of operations.

Phase 1 , Low estimate Phase 1 , High estimate Phase 2 , Low estimate Phase 2 , High estimate
Phase 1 dev: $15k–$28k. Phase 1 ops/yr: $6k–$14.4k. Phase 2 dev: $20k–$36k. Phase 2 ops/yr: $7.8k–$18.6k.
$910 – $1,800
At ~500 loans/month
$11k – $21.6k
API + infra + storage

Subscriptions & Running Costs

No labour costs. These are the actual tools, APIs, and infrastructure bills you'd pay monthly.

Adjust monthly loan volume
500 loans/mo
ItemLowHigh
OpenAI API , Phase 1
Underwriting analysis, ~2–3 calls per loan
$200 $350
OpenAI Vision API , Phase 2
Document extraction, ~20–30 pages per borrower
$400 $750
Cloud hosting (AWS or Azure)
Compute, networking, load balancer
$150 $300
Database (PostgreSQL / RDS)
Borrower data, loan records, embeddings
$80 $180
Document storage (S3 or Blob)
PDFs, parsed outputs, audit trail
$30 $120
Monitoring & logging
CloudWatch, Datadog, or similar
$50 $100
Total per month
$910 $1,800

Encompass API and nCino API access are assumed to be included in Geneva's existing licenses. Confirm with IT before finalising.

API cost breakdown: how the numbers are calculated
Phase 1 , Underwriting API
GPT-4o pricing: $2.50 / 1M input tokens, $10 / 1M output tokens.
Each loan triggers 2–3 API calls: profile analysis, recommendation generation, and dashboard formatting.
A typical call uses ~2,000 input + 600 output tokens, costing ~$0.011 per call.
~$0.40–$0.70 per loan processed.
Phase 2 , Vision & Document API
GPT-4o Vision processes each document page at 765–1,105 tokens per page at high detail.
Average borrower submits 5–8 documents, 20–30 pages total.
Extraction alone costs ~$0.50–$1.20 per borrower, plus a cross-reference call against Encompass.
~$0.80–$1.50 per loan processed.

Estimated delivery time
1 developer
2 developers
7–10 weeks
AI underwriting assistant
9–13 weeks
Document intelligence
12–17 weeks
Phase 2 starts at week 4
2 developers cuts total delivery time by roughly 12–17 weeks compared to a solo developer , getting Geneva live and generating ROI significantly faster at no additional software cost.

Phases can overlap after week 4 of Phase 1, cutting overall calendar time without any added cost.

Open Questions

These need answers before final scoping and pricing can be confirmed.

Does the AI recommendation need to trigger any automated actions in Encompass, or is it advisory only? This significantly changes the architecture.

Is there existing historical loan data (approved, denied, conditioned) that can be used to calibrate the AI model's recommendations?

What is the average monthly loan volume across all branches? This directly determines AI API costs.

Which document types are highest priority , pay stubs, W2s, bank statements, tax returns? Parsing complexity varies significantly by type.

Are there specific compliance rules or internal underwriting guidelines already documented? These can be encoded directly into the AI prompts.

Who on the internal team will be the point of contact for Encompass API access and nCino credentials during development?

Is there a preference for cloud provider , AWS, Azure, or GCP , for hosting the AI pipeline and data processing infrastructure?