Case study · Canadian mortgage lender

Where AI earned its place at a Canadian mortgage lender.

In 2026 I worked with the lender as its AI consultant, across nine teams, and I still advise them today. Every idea was tried on real work first. Some became tools people use. Others stay prototypes until they prove themselves.

Client: A Canadian private mortgage lender and fund manager
My role: AI consultant, then advisor
Teams: Sales, underwriting, servicing, accounting, capital markets, legal, IT, development, executive
Since: 2026, ongoing

Where it started

Lots of AI ideas. Not much time to test them.

Leadership wanted staff to become “doers, not requesters.”The lender’s leadership, on the AI program

Documents everywhere. Staff read and re-keyed deal packages, statements and notices by hand.
Rules in people’s heads. Operating rules lived in spreadsheets and in the memory of a few experienced people.
Reports stitched together. Leaders pieced the week together from emailed reports in different formats.
Data that has to stay safe. Borrower and investor data needed Canadian hosting and tight permissions before AI could touch it.

The work, drawn to scale

Six pieces of work, and where each one landed.

Listed in the order they started. Bars show roughly when each was active in 2026.

Sales & investor relations

AI chatbot for fund questions

Live

The question

Can it answer fund questions without making things up?

New sales and investor-relations staff needed quick, accurate answers about the lender’s funds. Many of those questions went to the CEO.

What changed

An internal AI chatbot in Teams answers from approved fund material and says so when it doesn’t know. It passed 30 tests before launch without a made-up answer, and testing caught a figure in a source file that needed correcting.

Development

Choosing an AI testing tool

Trial first

The question

Which of these vendors is worth a contract?

Software testing was slowing releases, and vendors were pitching AI testing tools.

What changed

I compared more than 15 tools and recommended a short trial of two on the same tests, using made-up data so no client information left the company. A small trial instead of a contract.

Accounting

Card statements, coded

Live

The question

Is this worth automating, and how?

Each month the team typed up company card statements and coded every charge by hand.

What changed

A tool, now in use, reads the statements and suggests a code for each charge from past decisions. An accountant reviews them. We also moved to the bank’s own export, which is more reliable than reading scans. My estimate: about a day of work each month becomes about an hour of review.

Across the business

AI Champions training

Delivered

The question

Will our people actually use it?

AI use was uneven. A few people used it well; most didn’t know where to start.

What changed

Weekly short sessions for a group from underwriting, sales, legal and development, each built on a real task from someone in the room. By the third week, eight people had a win to share.

Executive

The CEO’s weekly brief

Prototype

The question

Can I see the week without reading every report?

The CEO received many reports by email, each in its own format. Finding the blockers took an hour of reading.

What changed

A weekly brief sorts each area into blockers, momentum, issues and decisions, with every point linked to its source. It also caught reports that had stopped arriving and totals that didn’t add up.

Capital markets

Financial lender reports, automated

Live

The question

Can we trust the numbers before they go to our funders?

Monthly reports to the lender’s own funders were built by hand in complex spreadsheets, where one wrong cell is easy to miss.

What changed

Report engines rebuild each report from the loan data and compare it with the current version. The checks surfaced spreadsheet issues that are hard to spot by eye, such as dates stored the wrong way.

Also: an appraisal second opinion that uses the lender’s own property history (tested on a live deal, with read-only summary data), a Canadian-hosted cloud setup for these tools, and workflow mapping in servicing, insurance and legal.

“Working with NEOEDGE AI didn’t just give us a single-use AI tool; it changed the way we think about our workflows and how we can utilize AI at our core.”
Senior executive, Canadian mortgage lender

What another lender can take from this

Three things I’d do again.

Test it before anyone relies on it.

The fund chatbot answered 30 test questions before launch. That testing also caught a wrong figure in a source file before any staff member saw it.

Build the checks in.

The lender reports compare themselves with the current version, and the weekly brief flags reports that stop arriving. A tool that checks its own numbers earns trust faster.

Start small, and not always with AI.

A two-tool trial instead of a vendor contract. The bank’s own export instead of reading scans. The simplest fix that works is usually the right first step.

Your turn

Have one of these on your desk?

A Decision Sprint tests one AI decision on your real cases in about two weeks. You leave with a written call: build it, buy it, fix it, or stop.

Book a fit call

Sam TalasilaFounder, NEOEDGE AI. Former Head of LLMs & Data Products, Wealthsimple.