AI vs. human underwriting: what each catches, and what each misses.

Arithmetic and consistency are AI's clear strengths. Recalculating GDS/TDS at the qualifying rate rather than the contract rate, applying the 50% condo-fee rule the same way on every file, and cross-checking every document against a fixed checklist are exactly the tasks where a tool doesn't get tired, distracted, or inconsistent between a Monday file and a Friday one.

Speed compounds the advantage: a first-pass review that takes a human twenty or thirty minutes can run in a fraction of that time, which matters most at volume — the busier the pipeline, the more that consistency gap tends to widen between a rushed manual review and a checklist-driven one.

Anything that requires weighing context rather than applying a rule stays a human strength. Non-standard income — a self-employed borrower with an unusual T1 pattern, a new immigrant with a short Canadian credit history, a borrower transitioning between jobs — needs a judgment call about what the underwriter will actually accept, not just a calculation.

Story files — a good file with one complicating factor that needs explaining — and exception requests, where a broker is asking a lender to bend a normal rule for a specific reason, both depend on framing and relationship as much as arithmetic. Client conversations about what a stress-tested number actually means, covered in our stress test working reference, are squarely human territory too.

The workflow taking shape across Canadian brokerages runs AI first: documents get parsed, ratios get recalculated at the qualifying rate, and gaps get flagged automatically. A human — the broker, agent, or a fulfillment specialist — then reviews that output, resolves anything flagged, and adds the judgment calls AI can't make: how to frame a story file, whether an exception request is worth pursuing, what to say to the client.

Only then does the file go to the lender, where the actual underwriter — the only party with authority to approve it — makes the final decision. Nothing about that last step changes whether the first pass was run by a person or a tool.

No. Whether a first pass was run manually, by a contract underwriter, or by an AI tool, the credit decision authority stays with the lender's (or insurer's) underwriter under OSFI Guideline B-20, and the brokerage retains its regulatory responsibility for what was submitted regardless of which tool touched the file first.

That's worth repeating in plain terms because AI marketing sometimes implies otherwise: no tool moves accountability off the brokerage, and OSFI's broader direction on AI model governance — covered in our companion piece on AI underwriting tools in Canada — points toward more scrutiny of AI-assisted files over time, not less.

1. 01Start with your highest-volume, most standardized file type. Clean salaried purchase files are the easiest place to see whether an AI first pass actually saves time before trusting it on complex ones.
2. 02Keep a human review step, always. Even a fast AI pass needs a person confirming the flags and adding anything the tool couldn't weigh — that's not a temporary transition step, it's the permanent shape of the workflow.
3. 03Ask what the tool escalates, not just what it catches. A tool that's honest about its limits, on self-employed income or gifted deposits for instance, is more useful than one that confidently handles everything.
4. 04Measure against your current process, not against a perfect one. The relevant comparison is whether AI-assisted review beats your current manual first pass on speed and consistency — not whether it's flawless.

If you'd rather have that first pass run for you as part of a broader service, Treadstone's fulfillment associates combine both — underwriting-style human review, on every file, following the same discipline in the Pre-Underwriting File Review Checklist. If you want to run the AI-first pass yourself, join the early-access waitlist for Engage's AI mortgage underwriting.