Why AI Agents, MI Calculators, and Humans All Get Income Wrong. And Why Prism Exists.
Income calculation looks simple.
In real mortgage operations, it is the number one source of repurchase requests, investor stipulations, and post-close defects.
This is not a fringe issue. This is where small decisions quietly turn into large risk.
We recently ran a controlled experiment using the same borrower documents and three different approaches:
- A mocked agentic AI system
- A separate AI system using an MI calculator with light human review
- Prism IncomeXpert, Lender Toolkit’s production underwriting engine built for lenders working in Encompass by ICE Mortgage Technology
All three produced numbers that were close.
Only one produced a number that was defensible.
That difference is where Responsible Mortgage AI lives.
The Borrower Scenario (Intentionally Simple)
This was not a complicated self-employed file. It was a file lenders see every day.
- W-2 hourly borrower
- Same employer across multiple years
- Overtime present on the current paystub
- Overtime not historically broken out by type on W-2s
- Tight DTI where a few dollars per month matters
This is the exact profile that creates downstream chaos if income is not calculated conservatively.
Experiment 1: Mocked Agentic AI System
What it did well
- Correctly identified the borrower as W-2 hourly
- Correctly calculated base hourly income
- Correctly recognized overtime as variable income
Where it went wrong
The system included overtime in qualifying income.
- Base income: about $5,004 per month
- Overtime annualized conservatively: about $33 per month
- Total qualifying income: $5,037.09
Mathematically, this looks reasonable.
From a risk and defensibility standpoint, it is not.
The system did not understand why overtime cannot be used without historical segmentation by type. It reasoned probabilistically, not defensively. It optimized for “can this pass” instead of “will this survive.”
This is how AI systems over-qualify loans. Not maliciously. Just structurally.
Experiment 2: AI + MI Calculator + Light Human Review
A second experiment used:
- A different AI system
- An MI calculator referenced inside the prompt
- A light human review
Outcome
- Overtime was excluded
- Base hourly income only
- About $5,004 per month qualifying income
This was better than the first experiment.
But it still relied on:
- Theoretical hourly extrapolation
- Perfect realization of hours
- No adjustment for YTD drag or income volatility
MI calculators compute. They do not underwrite. They do not reason about stability signals, trend shifts, or audit defensibility.
They help you calculate. They do not help you defend.
Experiment 3: Prism IncomeXpert Output
Prism processed the same documents and produced:
- Recommended qualifying income: $4,995.42 per month
- Overtime explicitly documented but excluded
- Multiple system advisories generated automatically
Prism Advisories
- Stability: Recent increase detected. Underwriter review recommended
- Eligibility: Income must be expected to continue for 3 years
- Observation: Insufficient oldest-year data. Confirm AUS allows YTD plus 1 year analysis
These advisories explain why Prism avoided the higher numbers.
Prism did not ask:
“What is the highest income we can justify?”
It asked:
“What is the highest income that remains defensible under audit and quality control, including standards aligned with Fannie Mae expectations?”
That is Responsible Mortgage AI in action.
Why the Small Delta Is Actually a Big Deal
The difference between:
- $5,037
- $5,004
- $4,995
looks trivial.
It is not.
In tight DTI scenarios:
- $30 to $40 per month can flip a loan from pass to fail
- Over-qualifying income is one of the fastest paths to repurchase exposure
If a loan passes at $5,037 and fails at $4,995, and the system flagged stability concerns that were ignored, that is not just an AI miss.
That is lender liability.
This Is Not Just an AI Problem. Humans Get This Wrong Too.
Different roles in the workflow calculate income differently.
Loan Officer
- Incentivized to qualify
- Likely to use the highest reasonable number
- May include overtime because it exists
Processor
- Focused on document completeness
- May follow a checklist without interpreting risk signals
- Often defers judgment
Underwriter
- Thinks in terms of repurchase, audit, and investor defensibility
- Usually chooses the lowest supportable number
Prism behaves like a disciplined underwriter every time. No emotion. No production pressure. No memory gaps.
The Competitive Reality: Speed vs Survivability
The market is filling with:
- “Good enough” AI income tools
- Chat-based underwriting assistants
- LO-friendly qualification engines
Many optimize for speed and approvals.
Prism optimizes for survivability.
It is part of Lender Toolkit’s end-to-end automation strategy inside Encompass. Not just calculating income faster, but reducing the manual chaos that leads to post-close findings, audit friction, and repurchase risk.
This is the difference between generic AI and Responsible Mortgage AI. One helps you move faster. The other helps you stay in business.
The Takeaway
- AI agents can reason
- MI calculators can compute
- Humans can interpret
- Prism underwrites
Prism:
- Knows when not to use the highest number
- Detects instability signals humans overlook
- Produces an audit trail, not just an answer
- Turns complex income logic into consistent, defensible decisions inside Encompass
In mortgage underwriting, being almost right is often worse than being conservative.
This is not a rare edge case. This is the everyday file that quietly creates real losses.
And it is exactly why Prism exists.
