Photo credit: venturebeat.com
Wells Fargo has achieved a remarkable feat in the realm of financial technology by developing a generative AI system, known as Fargo, that operates on a large scale and is production-ready. In 2024, Fargo managed to facilitate an impressive 245.4 million interactions, surpassing initial projections by more than double, all while ensuring that sensitive customer data remains secure and untapped by external language models.
This AI assistant simplifies routine banking tasks through voice and text interfaces, able to assist customers with activities like bill payments, fund transfers, and inquiries regarding account transactions. Users tend to engage repeatedly in sessions, indicating its effectiveness and user-friendly design.
The infrastructure behind Fargo employs a privacy-centric approach. Customer interactions begin within the app, where speech is accurately transcribed using a speech-to-text model operating locally. The transcribed text undergoes a thorough cleansing and tokenization process via Wells Fargo’s internal systems, which include a small language model (SLM) for detecting personally identifiable information (PII). Only after these safeguards is a request sent to Google’s Flash 2.0 model, which identifies user intent and extracts relevant details, without any sensitive data being processed by the model itself.
Wells Fargo’s Chief Information Officer, Chintan Mehta, highlighted the strategic advantages of this orchestration in an interview, stating, “The orchestration layer interfaces with the model, while we act as filters in both directions.” He clarified that the model’s sole responsibility is to ascertain intent and entities based on user input, while all subsequent processing remains within Wells Fargo’s control. “None of our APIs connect directly to the large language model (LLM); they operate independently,” Mehta explained further.
Recent statistics from Wells Fargo showcase a significant increase in usage, with interactions climbing from 21.3 million in 2023 to over 245 million in 2024, totaling more than 336 million since Fargo’s launch. Notably, Spanish usage has soared, making up over 80% of interactions since the assistant’s rollout in September 2023.
This model reflects a key shift in strategy; Mehta noted the bank’s focus on establishing “compound systems” where orchestration layers dictate the appropriate model for each task. While Fargo relies on Google’s Gemini Flash 2.0, the bank is flexible, employing various smaller models such as Llama for other internal functions, along with OpenAI’s solutions as needed.
Mehta asserts that adopting a model-agnostic approach is crucial in today’s landscape, especially since the performance differences among top models have diminished significantly. Still, he acknowledges that certain models excel in specific domains—like Claude Sonnet 3.7 for coding or OpenAI’s models for deep research—but emphasizes the importance of how these models are integrated into operational pipelines.
One notable observation made by Mehta concerns context window size, where he views Gemini 2.5 Pro’s 1 million-token capacity as advantageous for tasks involving retrieval-augmented generation (RAG). He argued that in many cases, the lag introduced by preprocessing unstructured data often negates the benefits of using a model. “Gemini has done exceptionally well in these areas,” he remarked, hinting at the enhanced efficiency in handling high-volume automation without human input—a contrast to competitors like Citi, where concerns about the safety of large language models persist.
Agentic Moves and Multi-Agent Design
Wells Fargo is also exploring greater autonomy in its systems. Mehta shared insights into a project aimed at re-underwriting 15 years’ worth of archived loan documents, utilizing a network of cooperating agents, some based on open-source frameworks like LangGraph. Each agent was tasked with specific roles, ranging from document retrieval to data matching with record systems, culminating in calculations—all traditionally overseen by human analysts. While final outputs are subject to human review, the majority of the process operated independently.
There is ongoing evaluation of reasoning models for internal applications, where Mehta sees differentiation still prevalent. Although most models are proficient in routine tasks, reasoning capabilities represent a distinct advantage, with some models performing better than others through varying methodologies.
Why Latency (and Pricing) Matter
At Wayfair, the Chief Technology Officer, Fiona Tan, has noted the impressive speed of Gemini 2.5 Pro, stating that it can sometimes outperform models like Claude and OpenAI in response times. This efficiency in latency enhances the potential for real-time customer applications. Currently, Wayfair utilizes LLMs primarily in its internal functions, such as merchandising and capital planning, but quicker processing speeds could enable customer-facing applications, including Q&A tools on their product detail pages.
Tan also commented on improvements in Gemini’s coding performance, indicating that it is now comparable to Claude 3.7. Her team has begun testing the model through products like Cursor and Code Assist, allowing developers the flexibility to choose tools that best meet their needs.
With aggressively competitive pricing for Gemini 2.5 Pro set at $1.24 per million input tokens and $10 per million output tokens, Tan believes this pricing strategy, alongside options for reasoning tasks, positions Gemini as an attractive option for future projects.
The Broader Signal for Google Cloud Next
The success stories of Wells Fargo and Wayfair come at a pivotal moment for Google, coinciding with the company’s annual Google Cloud Next conference in Las Vegas. While the AI landscape has seen significant attention on OpenAI and Anthropic recently, enterprise implementations may start shifting favorably towards Google once more.
During the conference, Google is anticipated to introduce a slew of initiatives aimed at enhancing agentic AI capabilities, providing tools for making autonomous agents more effective within business workflows. During last year’s Cloud Next, CEO Thomas Kurian had predicted the development of agents to assist users in achieving specific goals and collaborating with one another—concepts that resonate with the orchestration and autonomy strategies articulated by Mehta.
Mehta expressed confidence in the transformative potential of generative AI for enterprise applications, asserting, “This technology is powerful. I have no doubt.” However, he warned against succumbing to the allure of emerging technologies without a careful evaluation of their practical benefits, recognizing that the true constraints in AI adoption may lie beyond mere computational power. “I believe the real bottleneck will be in power generation and distribution,” he noted, foreshadowing ongoing challenges in deploying these advanced technologies effectively.
Source
venturebeat.com