TD is using agentic AI for labor-intensive mortgage lending tasks including information synthesis and same-day mortgage approval. 

TD’s agentic AI tool, built by the bank’s in-house AI center, Layer 6, compresses information synthesis from 15 hours of work to roughly three minutes, Chad Koziel, associate vice president of AI product at Layer 6, told FinAi News.

The tool ingests all the documents behind a mortgage application — ID photo, a PDF bank statement, scanned purchase agreement — and turns them into a single record for a loan officer to review, Koziel said. 

Along the way, the tool calculates income, flags discrepancies such as a mismatched name or address between documents and checks for fraud, he said. 

But then it stops. 

“It doesn’t make a recommendation to the loan officer; that is strictly 100% within the loan officer’s purview,” said Koziel said. 

Asked why the agentic tool doesn’t make recommendations, Koziel pointed to the nature of the product.  

“Mortgages are complex, the circumstances of each mortgage are human and complex,” he said. “We want a human to make that judgment.” 

The tool has doubled the bank’s same-day approvals, meaning customers who walk into a branch in the morning with their documents have a strong chance of an answer by end of day, he said. The bank didn’t quantify how many originations it does for same day-approvals. 

The agentic AI tool was introduced to customers in January after about six months of development and works for any applicant, not only existing TD customers, Koziel said. 

Multi-model approach 

Layer 6 mixes model types depending on the task, Koziel said. 

“Sometimes we train our own foundation models from scratch,” Koziel said, citing TD’s in-house TD AI Prism.  

In other cases, the team fine-tunes open-source models such as Gemma or Llama or reaches for off-the-shelf options from OpenAI or Anthropic — combining them, he said, “in novel ways to achieve performance.” 

Register here for the FinAi Lending Summit, set for Oct. 7-8 in Las Vegas.