SAN DIEGO — Businesses using generative AI for more workplace roles and tasks face a puzzling problem: The technology isn’t very good at predicting what will happen next based on what’s happened before. In response, AI developers are touting a new kind of foundation model that can peer into the future. 

“Predictive tasks are not so easy for vanilla LLMs,” said Azin Asgarian, an AI technical lead in the AI lab at Georgian, a Toronto-based investment firm. Those systems—like OpenAI’s GPT Google’s Gemini—were trained on an internet’s worth of unstructured text data. As a result, the datasets don’t have many examples of financial and other numerical information in tables or chronological time series.

So-called tabular foundation models are trained on very specific data in a structured format, helping them make better forecasts.

Businesses have long used models to predict key trends like long-term customer churn or short-term product popularity. Before tabular foundation models, data science teams had to train a new system for every factor they wanted to forecast.

At TD Bank, that added up to hundreds of individual predictive models, each of which took a team of up to five people some six months to develop and get approved, said chief AI scientist Maksims Volkovs. “It just doesn’t scale.”

To speed up the work, the bank has built its own tabular foundation model called Prism. The system—developed by Layer 6, the bank’s AI unit—is trained on a year’s worth of customer transactions, account and credit bureau data. Bank staff now use it to forecast whether clients will sign up for or drop a service, or when they’ll need an increase in their borrowing limit. 

“One model is able to generalize to a whole bunch of these tasks,” said Volkovs, speaking at the NeurIPS AI research conference in San Diego last month, where Layer 6 was presenting a paper on tabular foundation models. Prism is also more accurate than the individual prediction models it replaced because it’s ingesting much more data, Volkovs said. 

Banks are “sitting on a goldmine of data,” said Greg Mori, vice-president at RBC’s Borealis AI unit. The firm trained and tested its Asynchronous Temporal Model (ATOM) on information from transactions using credit cards, accounts and rewards programs. 

Borealis began working on a foundation model for financial services in 2021. ATOM is now helping RBC chase its target of generating up to $1 billion in enterprise value from AI by 2027. The bank used the model for a new system that helps with decisions about how much credit to extend to clients. Mori said RBC has since lent $7.5 billion more to borrowers than it otherwise would have, generating an extra $375 million in net income before taxes. 

In total, RBC has used ATOM across 15 different services, including fraud prevention and loyalty promotions. It’s also built integrations between the system and its generative AI tools, so staff can ask questions about the data and use it to produce marketing materials or figure out how best to talk to different clients. 

“The depth of information in those transactions really tells a lot about the person,” Mori said. For example, the bank can target its travel-rewards card at customers who frequently spend money overseas, while a loss-protection service might be an easier sell to customers whose account activity suggests they’re risk-averse. 

While generative AI can be a black box, ATOM lets its developers peer inside, said Mori. Banks face legal claims and regulatory scrutiny over their decisions, particularly around credit. “There are explainability requirements—I need to be able to understand how that loan decision was made,” he added.

Large enterprises may have an edge in building tabular foundation models, because they tend to have lots of data in the formats needed to train these systems. Developers are now selling tools that let smaller firms take advantage of the technology. Mountain View, Calif.-based Kumo’s system makes predictions using structured data. At NeurIPS, tech giants like Amazon Web Services and startups like Freiburg, Germany-based Prior Labs presented papers on advances in building and benchmarking tabular foundation models. 

“Companies are trying to automate part of the data science workflow,” said Asgarian, adding that doing so means they don’t need to constantly train, evaluate and test new models from scratch. That could let product managers and other non-technical staff set up their own forecasting systems without needing to rope in an engineer. It could even let AI agents use predictive models to automate more tasks.

Asgarian gives the example of an online store. AI agents can design the web page, write the copy and access all the software systems needed to process orders. “But you also have to predict which items to put on the shelves,” she said. An AI agent plugged into a forecasting system could find and stock the right products, she claimed.

While tabular foundation models aren’t quite as advanced as LLMs yet, the early results are promising, Volkovs said. “Businesses still really care about having access to forecasts about what will happen in the future.”