AI mortgage underwriting is the use of software to turn mortgage documents and property evidence into a consistent review package. It is an underwriting aid, not an automatic credit committee.
Start with evidence, not a black-box decision
A lender usually begins with an application, an appraisal or valuation material, a credit report, title information, and supporting documents. An AI-assisted workflow can classify those files, extract fields, calculate relationships such as loan-to-value (LTV), and surface missing or conflicting information for review.
The useful output is a traceable file: the source documents, the extracted facts, the calculations, and the reasons a risk signal was raised. A score is a way to prioritize attention; it is not evidence by itself.
- Extract borrower, mortgage, and property fields from supported documents.
- Compare credit information, mortgage history, income, net worth, and employment stability.
- Review appraisal values, comparable sales, location and property-type signals.
- Apply lender-configured rules and present red flags beside the deal.
A simple LTV example
Suppose a proposed mortgage advance is $450,000 against a supported value of $750,000. The standalone proposed-advance-to-value ratio is 60%. That calculation is helpful, but it is not the whole recovery question. If an existing $120,000 second-ranking charge is retained ahead of the proposed advance, the aggregate prior-ranking-charge-plus-proposed-advance amount is $570,000 and the aggregate ratio is 76%.
A lender should confirm which value is being used, whether the prior charge is actually registered and current, and whether sale costs, construction exposure, condition, and market liquidity change the recovery picture. AI can make the arithmetic and inconsistency visible; the underwriter decides which evidence is acceptable.
Where human judgement remains essential
An underwriter still validates identity and documents, chooses the valuation basis, interprets repayment conduct, assesses collateral liquidity, tests the exit strategy, and decides whether an exception is reasonable. A clean extraction cannot cure a stale appraisal, an undisclosed prior charge, or an unsupported repayment plan.
The right operating model is human-in-the-loop: software compresses repetitive review and makes risk signals consistent, while the lender owns the final approval, decline, conditions, and exception record.
Questions to ask before adopting it
- Can reviewers open the source evidence behind an extracted field?
- Does the workflow show standalone proposed-advance/value and aggregate prior-charge exposure separately?
- Can a lender configure policy rules without hiding the underlying facts?
- Are exceptions recorded for a person to approve, rather than silently overridden?
- Does it support Canadian documents and the lender's own review process?