Wells Fargo’s assistant, powered by Google’s AI, poised to hit 100 million interactions annually

Wells Fargo’s CIO Chintan Mehta divulged details around the bank’s deployments of generative AI applications, including that the company’s virtual assistant app, Fargo, has handled 20 million interactions since it was launched in March.

 “We think this is actually capable of doing close to 100 million or more [interactions] per year,” he said Wednesday evening in San Francisco at an event hosted by VentureBeat, “as we add more conversations, more capabilities.” 

The bank’s traction in AI is significant because it contrasts with most large companies, which are only in the proof of concept stage with generative AI. Large banks like Wells Fargo were expected to move particularly slowly, given the massive amount of financial regulation around privacy. However, Wells Fargo is moving forward at an aggressive clip: The bank has put 4,000 employees through Stanford’s Human-centered AI program, HAI, and Mehta said the bank already has “a lot” of generative AI projects in production, many of which are serving to make back-office tasks more efficient.

Mehta’s talk was given at the AI Impact Tour event, which VentureBeat kicked off Wednesday evening. The event focused on how enterprise companies can “get to an AI governance blueprint,” specifically around the new flavor of generative AI, where applications are using large language models (LLM) to provide more intelligent answers to questions. Wells Fargo is one of the top three banks in the U.S., with 1.7 trillion in assets.

Wells Fargo’s multiple LLM deployments run on top of its “Tachyon” platform

Fargo, a virtual assistant that helps customers get answers to their everyday banking questions on their smartphone, using voice or text, is seeing a “sticky” 2.7 interactions per session, Mehta said. The app executes tasks such as paying bills, sending money and offering transaction details. The app was built on Google Dialogflow and launched using Google’s PaLM 2 LLM. The bank is evolving the Fargo app to embrace advances in LLMs and now uses multiple LLMs in its flow for different tasks — “as you don’t need the same large model for all things,” Mehta said.

Another Wells Fargo app using LLMs is Livesync, which provides customers advice for goal-setting and planning. That app launched recently to all customers, and had a million monthly active users during the first month, Mehta said.

Notably, Wells Fargo has also deployed other applications that use open-source LLMs, including Meta’s Llama 2 model, for some internal uses. Open-source models like Llama were released many months after the excitement around OpenAI’s ChatGPT started in November of 2022. That delay means it has taken a while for companies to experiment with open-source models to the point where they are ready to deploy them. Reports of large companies deploying open-source models are still relatively rare. 

However, open source LLMs are important because they allow companies to do more tuning of models, which gives companies more control over model capabilities, which can be important for specific use cases, Mehta said.  

The bank built an AI platform called Tachyon to run its AI applications, something the company hasn’t talked much about. But it’s built on three presumptions, Mehta said: that one AI model won’t rule the world, that the bank won’t run its apps on a single cloud service provider, and that data may face issues when it is transferred between different data stores and databases. This makes the platform malleable enough to accommodate new, larger models, larger models, with resiliency and performance, Mehta said. It allows for things like model sharding and tensor sharding, techniques that reduce memory and computation requirements of model training and inference. (See our interview with Mehta back in March about the bank’s strategy.)

The platform has put Wells Fargo ahead when it comes to production, Mehta said, although he said the platform is something that competitors should be able to replicate over time. 

Multimodal LLMs are the future, and will be a big deal

Multimodal LLMs, which allow customers to communicate using images and video, as well as text or voice, are going to be “critical,” Mehta said. He gave a hypothetical example of a commerce app, where you upload a picture of a cruise ship, and say “Can you make it happen?” and a virtual assistant would understand the intent, and explain what a user needed to do to book a ride on the cruise ship. 

While LLMs have been developed to do text very well, even cutting-edge multimodal models like Gemini require a lot of text from a user to give it context, he said. He said “input multimodality” where an LLM understands intent without requiring so much text, is of bigger interest. Apps are visual mediums, he said.

He said the core value of banking, of matching capital with a particular user’s need, remains relatively stable, and that most innovation will be on the “experiential and capability end of the story.” When asked where Wells Fargo will go here, he said that if LLMs can become more “agentic,” or allow users to go do things like booking a cruise by understanding multimodal input and leading them through a series of steps to get something done, it will be “a big deal.” A second area is around providing advice, where understanding multimodal intent is also important, Mehta said.

Slow regulation has made AI governance a challenge

On the topic of governance of AI applications, Mehta said that the bank’s answer to this has been to focus on what each application is being used for. He said the bank has “documentation up the wazoo on every step of the way.” While most challenges around governance have been dealt with, he agreed that areas around the security of apps, including cybersecurity and fraud, remain challenges.

When asked what keeps him up at night, Mehta cited banking regulation, which has increasingly fallen behind technology advances in generative AI, and areas like decentralized finance. “There is a delta between where we want to be and where the regulation is today. And that’s historically been true, except the pace at which that delta is expanding has increased a lot.”

Regulatory changes will have “big implications” for how Wells Fargo will be able to operate, including around economics, he said: “It does slow you down in the sense that you have to now sort of presume what sort of things have to be addressed.” The bank is forced to spend a lot more engineering time ”building scaffolding around things” because it doesn’t know what to expect once applications go to market.

Mehta said the company is also spending a lot of time working on explainable AI, an area of research that seeks to understand why AI models reach the conclusions they do.