The Bridge Decision: Walking the AI Tightrope as a Mortgage Professional

Standing at the Edge

Imagine standing at the edge of a bridge you need to cross. An AI system — fast, confident, and drawing on an enormous body of knowledge — tells you the bridge will hold. But you can see rust at the joints. You can hear a faint creak when the wind shifts. The AI can't see what you see. It can't hear what you hear. And it will not be the one falling if the bridge fails.

That image captures something I've been working through for the past several years as both a mortgage loan officer and as the architect of an AI-assisted mortgage planning platform. The question isn't whether AI is useful — it clearly is. The question is: who decides when to trust it, and who bears the consequences when the answer is wrong?

"

AI provides capability without investment. You provide investment without certain capabilities. The combination is powerful — but only if you consciously understand when to trust each type of intelligence and never forget who bears the consequences.

In mortgage lending, that question has a very specific answer: the loan officer whose name is on the disclosure.

What AI Does Brilliantly in Mortgage

I want to be clear that I'm writing this as a genuine user of AI in mortgage work — not as a skeptic. Over the past several years I've built a comprehensive mortgage planning software suite, the Mortgage Update platform, that relies heavily on AI-assisted analysis. Its capabilities are substantial:

These tools are genuinely powerful. They compress hours of analysis into seconds and allow a loan officer to have strategic conversations with borrowers that weren't possible before. AI makes them work.

And yet — there is a specific, predictable, and consequential way that AI fails in mortgage work. Understanding that failure is what led me to build the platform the way I did.

The Rust on the Bridge: AI Hallucination in Mortgage Calculations

AI systems are trained to produce plausible answers. In most domains, a plausible answer is a useful answer. In mortgage lending, a plausible answer that is technically wrong is a compliance violation waiting to happen.

Here is a concrete example from the development of the Mortgage Update platform.

Most people — including many AI systems — would assume that private mortgage insurance (PMI) is calculated the same way across all loan types: as a percentage of the original loan amount, applied as a flat annual premium. That assumption is intuitive. It's simple. And for conventional loans, it's roughly correct.

It is wrong for FHA and USDA loans.

FHA mortgage insurance and USDA guarantee fees are calculated based on the average outstanding balance, not the original loan amount — and they recalculate every year as the balance decreases. The premium is not static. It changes annually. Additionally, some forms of mortgage insurance cancel automatically when the borrower reaches a certain loan-to-value threshold. Others remain in force for the life of the loan regardless of how much equity the borrower accumulates. The rules are loan-type specific, and they are not intuitive.

An AI system asked to calculate a monthly payment for an FHA loan will often produce a number. The number will look precise. It will appear in a clean format. It will probably be wrong by a small but meaningful amount — and more importantly, if used in a disclosure document, it may be out of compliance with federal requirements.

"

No one is going to blame AI if my numbers are wrong. It's my reputation on the line, and my liability if I'm wrong.

This is not a hypothetical risk. Mortgage disclosures are legal documents. The figures that appear on them are representations the loan officer makes to the borrower and to regulators. A convincing-looking wrong answer, reproduced from an AI system without verification, is not a defense. It is evidence of negligence.

The Solution I Built: Enforced Precision

The Mortgage Update platform addresses this with a firm architectural rule: all mortgage calculations are performed by a thoroughly tested Python engine using Decimal-precision arithmetic. AI is not permitted to estimate, approximate, or improvise financial figures. If the engine hasn't calculated it, it doesn't appear.

This distinction matters. Python's Decimal library eliminates the floating-point rounding errors that accumulate across 360 months of amortization calculations. The engine is tested against known-correct values for every loan type, every mortgage insurance variant, and every edge case we've encountered across 35 years of origination experience. When it produces a number, that number is correct — not plausible.

AI's role in the platform is analytical and strategic: helping loan officers understand scenarios, identify opportunities, and communicate clearly with borrowers. The moment a specific dollar figure is required, the tested engine takes over. That is the correct division of labor.

This is not AI-skepticism. It is the application of a framework I've come to call the Bridge Decision.

The Bridge Decision Framework

Every time a professional uses AI to inform a decision, four questions should be running in the background. Not as a formal checklist — but as an internalized discipline.

Let me walk the PMI example through this framework explicitly.

The answer to the last question is the most important one. Decision rights follow consequence. If you bear the consequences, you own the decision — and that means you cannot outsource verification to a system that produces plausible answers rather than correct ones.

The Right Partnership

None of this is an argument against AI in mortgage work. The Mortgage Update platform exists precisely because AI enables analysis and client communication that wasn't feasible before. I use it every day.

The argument is for conscious partnership — understanding what each party brings, and building systems that respect that division clearly. AI brings speed, pattern recognition, and analytical capacity at a scale no human can match. Domain expertise brings the judgment to know when the pattern doesn't apply, the experience to recognize the edge case, and the accountability that makes the answer mean something.

The bridge metaphor holds. When the AI tells you the bridge will hold, that information is valuable — it's one data point from a system that has processed a great deal of relevant information. But you're the one who can see the rust. You're the one who can hear the creak. And you're the one who will be crossing it.

"

The fastest path to the right answer is often a wrong proposal. The fastest path to the wrong disclosure is an unverified one.

In mortgage lending, the Bridge Decision isn't a thought experiment. It's a professional obligation.