TD Quietly Put an Agent in Charge of Its Mortgage Pipeline

While the industry discourse focuses on consumer-facing AI agents, Canada's largest bank by mortgage book just deployed an agentic AI model into the application workflow for mortgages and home equity lines of credit. No press tour. No demo video. Just an announcement and a deployment.

There is a particular kind of bank story that gets ignored because it lands without theater. TD Bank confirmed last week that it has built and deployed an agentic AI model to automate the application process for mortgages and home equity lines of credit. The bank's own framing was modest. The story is not.

Canada's Big Five banks have been deploying production AI in customer-facing workflows for longer than most North American peers. RBC has had Aiden in production trading for years. CIBC runs AI in fraud and KYC at scale. TD has multiple production AI systems in retail banking. What is different about this deployment is the asset class. Mortgages are not a low-value, low-risk experiment surface. They are the largest single loan asset on the bank's balance sheet, and TD is the biggest residential mortgage lender in Canada by some distance.

The discourse keeps assuming agentic AI in banking will arrive in the consumer chatbot first. It is arriving in underwriting workflows first. The consumer chatbot is the last mile, not the first.

We covered the broader pattern in agentic AI finding its verticals earlier this year. The thesis was that incumbent financial services would deploy agentic systems into the back office before the customer surface. TD's mortgage move is the cleanest proof point in retail banking so far.

What TD actually shipped

The bank's announcement names two workflows: new mortgage applications and HELOC applications. The agent handles document collection, identity verification routing, income validation, property data lookup, and pre-underwriting checks. It does not, per the bank's framing, make the underwriting decision itself. A human underwriter receives the agent's compiled file and makes the credit call.

That distinction is doing a lot of work. The Office of the Superintendent of Financial Institutions (OSFI) in Canada has been clear that AI used in lending decisions is subject to model risk management under Guideline E-23 and broader fairness expectations under recent guidance. By keeping the underwriting decision in human hands, TD avoids the regulatory weight that comes with a fully automated lending agent. The agent's job is to compress the time-to-decision from weeks to days, not to take the decision out of the bank's hands.

The numbers TD has hinted at publicly suggest a 40 to 60 percent reduction in application processing time. Concretely, that means the borrower's experience moves from "two to four weeks of back-and-forth" to "two to four days, with most touchpoints proactive." That is not a marginal improvement. It is the kind of improvement that flips a bank's mortgage acquisition funnel.

Why mortgages, why now

Three reasons, in order of how much weight they deserve.

The first is competitive. Canadian mortgages are an oligopoly, but inside that oligopoly the experience is similar across all five major banks. Differentiation has historically come from rate, branch network, and broker relationships. TD is making a bet that processing speed becomes the new axis of competition. If TD can close in days while RBC and BMO close in weeks, the broker network reroutes to TD on every margin-thin deal. The downstream effect on origination volume is measurable.

The second is operational. Mortgage origination is one of the most labor-intensive workflows in retail banking. Document collection alone consumes an enormous amount of underwriter time that has nothing to do with credit decisions. A well-built agent that handles the document workflow frees the underwriter's day for the parts of the job that require judgment. The bank's cost-to-originate drops without reducing headcount, because the freed time is reallocated to volume.

The third is preparatory. Once an agent is operating reliably in the application workflow, the next move is wiring it into underwriting recommendation. TD has not announced this, but the architectural pattern points there. Build the agent in the lower-risk surface, watch it for six to twelve months, then extend it. The mortgage application agent is a foothold, not a destination.

As Theodora Lau has written about agentic AI in banking, the institutions that move first into back-office workflows accumulate an operational learning curve their competitors cannot replicate quickly. TD is buying that curve now.

The MM Liability Gap, lending edition

We use the MM Liability Gap framework to evaluate situations where new capability creates ambiguity in who bears responsibility when something goes wrong. Lending is the canonical test case.

When a human underwriter declines a borrower for reasons that turn out to be discriminatory, the bank is unambiguously liable. The decision-maker is identifiable. The reasoning is documented. The Canadian Human Rights Commission has clear procedures.

When an agent compiles the file and a human signs off on the decline, the chain of accountability blurs. Did the agent surface the right documents? Did it route the right income validation queries? Did it apply implicit weighting that influenced the human's call? The human signs the decision, but the agent shaped the inputs. If the agent's training data contained pattern biases, those biases now sit upstream of the human decision and the human cannot necessarily see them.

This is the part of the deployment that will get tested first, and it will be tested by a borrower who is declined, suspects the agent shaped the outcome, and has lawyers willing to pull at the thread. TD has presumably done the model risk work to defend this. The interesting question is what happens at the second-mover bank that does not.

The broader point covered in The AI Fraud Paradox applies here too. The same AI that compresses operational cost also compresses the audit trail. Banks need to invest in the explainability layer at the same time they invest in the agentic layer, or the regulator will find them later.

What other Canadian and US banks will do

The next 90 days are predictable.

Royal Bank of Canada will announce a comparable program before Q3 close. They have the Borealis AI bench, they have the in-house tooling, and they cannot afford to let TD have the speed-to-close talking point through fall mortgage season. BMO will follow within a quarter, likely partnering with an external vendor rather than building. Scotiabank will brief analysts on its own program even if the deployment is months away from production.

The American picture is different. The US mortgage market is far more fragmented and the regulatory posture is less unified, with the Consumer Financial Protection Bureau, the FDIC, the OCC, and state regulators all having a say. A US bank that ships the equivalent of TD's program will spend twice as long on model risk work and three times as long on disparate impact testing. JPMorgan, Wells Fargo, and Bank of America all have the engineering capacity. None of them has the regulatory cover to move at TD's pace.

The bigger picture is that agentic AI in regulated lending is starting now, in Canada, in the largest loan asset class, at the largest mortgage lender. That is a meaningful data point about where the production deployments land first. It is not a chatbot. It is not a marketing experiment. It is the bank's primary product, with an agent inside it. Watch how it performs over the next two reporting cycles.

The first agentic AI deployment in production retail lending at a major North American bank is not in a chatbot. It is in the workflow that decides who buys a house and on what terms. Which of your bank's workflows is next?

Charlie Major is a Product Development Manager at Mastercard. The views and opinions expressed in Major Matters are his own and do not represent those of Mastercard.