TD Bank is testing whether a foundation model trained on outside data can match the predictive power of models it builds entirely in-house — a meaningful shift in philosophy for a bank that has long prided itself on proprietary AI. 

The bank is running a proof-of-concept test with Prior Labs, a Freiburg, Germany-based startup that pioneered tabular foundation models, a category of AI purpose-built for structured data, Maksims Volkovs, chief AI scientist at TD, told FinAI News.

TD would not say how much the project costs. 

The collaboration puts Prior Labs’ TabPFN models to work on TD use cases within the bank’s agentic treasury and retail banking businesses, Volkovs said. 

“In the past, we’ve preferred to start from scratch and train the predictive tabular models we use strictly on our data; we did that with TD AI Prism,” he said. “But there are some potential use cases we’re exploring that don’t quite have large volumes of data to work with and train the models.” 

Because developing tools for such processes can be tricky without enough data, the bank is exploring how a model that’s trained on data that isn’t specific to TD could identify similar patterns to make accurate predictions, Volkovs said. 

That data scarcity problem is what drew TD to Prior Labs’ TabPFN, Volkovs said.  

TabPFN is trained on hundreds of millions of synthetic databases, enabling it to process data across industries without task-specific training. 

According to Prior Labs’ website, the company can help FIs deploy AI tools for: 

• Credit risk modeling for mortgage approvals and SME banking; 

• Automotive finance fraud detection; 

• Spend forecasting; 

• Portfolio growth forecasting and treasury management; and 

• Algorithmic trading — forecasting directional price movements of assets. 

If the accuracy of Prior Labs’ model holds up, the next phase would address explainability and model validation — prerequisites under TD’s Trustworthy AI governance framework — before it can move toward production, he said. 

The longer-term vision is collaboration, not replacement. Volkovs pointed to a scenario in which TD AI Prism and a generalized pre-trained foundation model operate together, with Prism handling use cases rich in proprietary TD data and the external model filling data. 

A mixed development approach 

Many financial institutions are taking a mixed approach when it comes to developing and deploying AI within their own operations, Ashish Nagar, founder and chief executive of Level AI, a customer-experience automation vendor that has a platform that runs on custom-built models, told FinAi News. 

Data security, latency, accuracy and cost are the top parameters that drive a bank’s decision to build models in-house or rely on external LLM providers, he said. 

“I expect large institutions [that] can spend big on tech to develop small parameter models for operations that are very data sensitive for them,” Nagar said.  

“FIs that cannot spend in training and build their model from scratch will rely on third-party vendors who can ensure data security.” 

Nagar said a combination of both will best serve FIs. 

“It has to be a mixed approach of in-house built, external LLMs, closed source and open source,” Nagar said, adding that Level AI helps banks choose the right model for the right task. 

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