How Safe Is TD’s New Agentic AI? Inside the 3-Minute Mortgage

When big news breaks around artificial intelligence, it doesn’t often feature Canadian companies, let alone our geriatric banks.

So the announcement of TD’s new agentic AI, which aims to reduce mortgage underwriting from 15 hours to three minutes, was both refreshing and exciting. A quicker turnaround at this stage of the process means less anxious waiting for the bank’s mortgage applicants.

But the risks associated with AI are worrying. Does the increased efficiency promised by TD come at a cost home buyers are willing to pay?

AI evangelists tout the technology’s ability to create a more efficient and profitable world, but it’s still an experimental tool that’s prone to errors and hallucinations. About half of Google AI Overviews in late 2025 and 2026 contained facts not supported by their cited sources, according to an analysis conducted by AI startup Oumi.

It’s also fair to wonder whether siccing AI on a mortgage file will strip the process of its humanity. A quick turnaround is great, but is that speed the result of ignoring the subtleties that make each borrower unique?

Let’s find out.

Where’s the human-in-the-loop?

For home buyers who don’t fit the big banks’ strict lending criteria — and there are plenty of them in Canada — the mortgage process isn’t as simple as sending over a pile of documents. Earn a non-traditional income or be in recovery mode from a past credit hiccup and the narrative behind your finances can become a critical factor in getting approved.

“When it comes to underwriting, there are nuances to it,” says Frances Hinojosa, CEO, co-founder and principal mortgage broker at Tribe Financial Group.

“You have to be able to understand the story of the client, the situation they're in and where they want to move forward. Pure data is not necessarily going to paint the full financial picture of the consumer that's in front of you.”

According to Chad Koziel, associate vice president of AI product at Layer 6, TD’s AI research centre, TD’s agentic AI is only concerned with the data piece of the puzzle. It scans an applicant’s submitted documents and summarizes the pertinent information.

“That's a lot of work,” Koziel says. “You're combing through cell phone photos, scanned documents, downloaded PDFs, some of which might be many pages.”

TD’s AI doesn’t recommend or decide. But it does speed up the decision making process by providing loan adjudicators the details they need.

“At TD, we remain focused on having that human take into account all of the myriad factors associated with this frankly emotional decision,” Koziel says.

Just how intelligent is it?

John Vo, a mortgage broker with Premiere Mortgage Centre in Halifax, applauds TD’s efforts to hurry things along for borrowers.

“Every hour that a client waits to hear back feels like a day, or a week,” he says. “If the banks can shorten the time that they take to get to a yes, then it's going to be a better client experience.”

But Vo wonders if TD’s AI is ready for prime time.

“I've worked for many of the larger banks. I've seen them adopt technologies and then undo them. I've seen systems go down, crashes, stuff like that. So I'm a little cynical,” he says.

Vo says he may even consider sending clients to other lenders with similar product offerings while TD “works out the kinks.”

That’s somewhat of a moot point. The bank’s AI agent is not being used on broker-submitted files, just those that are created directly with a TD mortgage specialist. But Vo’s concerns around a new technology’s potential mistakes are understandable.

Koziel is confident that TD’s multi-stage monitoring of the AI’s performance will reduce errors from the jump.

“When the model is wrong, and it's wrong some very small percentage of the time, we ask, ‘What is the nature of the error that the model made?’ We don't simply ask, ‘Did it produce an accurate summary at the end?’ We ask at each stage, ‘What is this model doing?’” he says.

Koziel says that when significant errors occur they are hunted down “viciously”. But he says those cases are rare. Many of the errors his team has found can’t be blamed on AI, but rather the humans responsible for the initial data entry.

What’s been done to prevent bias?

Channarong Intahchomphoo is an adjunct professor at the University of Ottawa’s School of Engineering Design and Teaching Innovation who has addressed the United Nations about the risks of bias inherent in AI models. He worries that the data used to train TD’s AI agent could be rife with historical examples of bias.

“Humans are biased. The human system of banking is biased already. And this is translated into data training the model,” he says.

But Intahchomphoo believes that TD’s IT team is up to the challenge of reducing bias. He says it has a reputation for doing good work, and the bank’s status as a federally regulated financial institution means mistakes could be costly.

Koziel was unable to detail Layer 6’s proprietary strategies for addressing bias, but he says AI models that the lab deployed elsewhere were tested using a method known as perturbation.

A typical perturbation test might involve entering different demographic data — age, language, sex — into a model and observing whether it spits out consistent answers or shifts its advice as the inputs change.

So long as TD’s AI agent is tasked with organization and not decision making, the potential for bias should remain low.

» MORE: How gender bias can hurt your wallet

The verdict

Knowing that TD’s AI agent is tasked only with mind-numbing grunt work comes as a relief.

No one is going to mourn the loss of manual data retrieval the agent makes possible. A streamlined underwriting process could save borrowers days of waiting, and ensure that the interaction at the heart of a mortgage deal — humans helping humans — survives.

TD’s foray into AI underwriting gets a passing grade, but not all lenders have the same resources, or the same Big Six responsibility to appease regulators and shareholders. As more of these tools hit the market, it’s important that they’re evaluated for security, effectiveness and fairness.

Consider it a stress test for lenders. One they shouldn't be allowed to fail.