Wells Fargo
conversational banking assistant at scale, architecture with no PII to the LLM
Fargo, the conversational assistant in the Wells Fargo app launched in 2023, went from 21.3 million interactions in 2023 to 245.4 million in 2024 and passed 1 billion cumulative in less than three years, with an architecture that never sends sensitive customer data to the LLM.
Key points
- Fargo conversational assistant in the Wells Fargo Mobile app, by voice or text.
- In-house orchestration plus Google Gemini Flash 2.0, PII masked before any call to the LLM.
- From 21.3 million interactions in 2023 to 245.4 million in 2024, more than 1 billion cumulative.
- Evidence level A, confirmed active status.
Objective
Handle the customer's everyday banking need in self-service (balance, transactions, Zelle transfers, bill payment, routing numbers, spending insights) inside the app, without going through a costly human channel, and anchor the use of the mobile app as the primary relationship channel.
The deployment
Fargo is the conversational assistant of the Wells Fargo Mobile app, launched in 2023. The customer queries it by voice or text to send money with Zelle, pay a bill, find a routing number, check a balance, or understand their spending. Its architecture is the notable point of the case: speech is transcribed locally, the text is cleaned and tokenized by the Wells Fargo stack, a small model detects PII, and only the intent and entities go to an LLM (Google Gemini Flash 2.0 as the main model). No sensitive customer data passes through the model. Usage jumped from 21.3 million interactions in 2023 to 245.4 million in 2024, more than twice the initial projections, passing 1 billion cumulative interactions in less than three years (March 2026 announcement). More than 3 million Spanish-speaking customers have used Fargo, with more than 160 million interactions.
Results Proof A
Volumes published by Wells Fargo itself in an official release (1 billion interactions, 33 million mobile active users, March 2026, T1), consistent with the internal figures reported by the established tech press (21.3 million in 2023, 245.4 million in 2024, VentureBeat). Several concordant sources including a primary one from the subject brand, holding over several years, raise the level.
How it works
Documented architectureThe stack in detail
- outil Fargo (orchestration maison) Assistant integrated into the Wells Fargo Mobile app. The customer's voice is transcribed locally, the text is cleaned and tokenized by the bank's stack, a small model detects PII; only the intent and entities are then sent to the LLM. Detokenization and API calls stay on the bank's side.
- llm Google Gemini Flash 2.0 Main model called by the orchestration layer to extract the intent and entities of the request, without ever receiving raw data or PII. Wells Fargo describes a compound system where the orchestration chooses the model based on the task (other models such as Llama or OpenAI can be mobilized).
- outil Petit modele de detection de PII (in-house) Reduced language model that spots and masks personal data before any call to the LLM, key to the privacy-first architecture.
- plateforme Application Wells Fargo Mobile Single channel for Fargo (voice and text), integrated with accounts and transactions; the app has passed 33 million monthly active users.
How it runs, concretely
For ops teams-
1Customer request customer
The customer speaks or types their request in the app (Zelle transfer, bill, balance, routing number, spending).
-
2Local transcription and masking AI
Speech is transcribed locally; the text is cleaned and tokenized; a small model detects and masks PII before any external call.
-
3Intent extraction by the LLM AI
The orchestration layer sends the masked text to the LLM (Gemini Flash 2.0), which returns the intent and entities, without ever seeing raw data.
-
4Execution on the bank side data team
Detokenization, API calls to accounts, and the action (transfer, response, insight) happen entirely on Wells Fargo systems.
-
5Monitoring and extension data team
The in-house team tracks volumes, resolution, and misunderstood cases to widen the scope and adjust the model choice in the compound system.
The intent expressed by the customer, extracted after PII masking. The system optimizes on intent understanding and the human-free resolution rate; if masking or detokenization break, the response can no longer be served without exposing data.
How your customers perceive this type of use
Sourced studiesLes consommateurs n'acceptent pas les chatbots par defaut : 64% prefereraient que les entreprises n'utilisent pas d'IA dans leur service client (Gartner, 2024) et pres d'un utilisateur sur cinq du service client par IA n'en retire aucun benefice (Qualtrics, 2025). L'acceptation se construit sur trois conditions mesurees par Salesforce : savoir qu'on parle a une IA, pouvoir escalader vers un humain, comprendre la logique de l'agent.
Acceptance conditions
- Etre informe qu'on parle a une IA et non a un humain (pres de 75% le demandent, Salesforce 2024)
- Un chemin d'escalade clair vers un agent humain (45% plus enclins a utiliser l'agent IA, Salesforce 2024)
- Une logique de l'agent clairement expliquee (44% plus enclins, Salesforce 2024)
Red lines
- Rendre l'humain injoignable : c'est la premiere inquietude des consommateurs sur l'IA dans le service client (Gartner 2024) et 50% craignent que l'IA les coupe du contact humain (Qualtrics 2025)
- Remplacer le service client par l'IA sans alternative : 53% envisageraient de partir chez un concurrent (Gartner 2024)
Sources: Salesforce 2024 · Gartner 2024 · Qualtrics 2025
How to replicate
Inference, not sourcedData prerequisites
- Real-time account and transaction data
- History of customer intents to frame the covered cases
- Reference set of PII to mask (names, account numbers, sensitive amounts)
Org prerequisites
- Dedicated AI product and engineering team over time
- Banking security and compliance framework
- Governance of model choice in a compound system
Possible stack
- Small PII detection and masking model on the enterprise side
- Orchestration layer that routes to the right LLM based on the task
- Third-party LLM (Gemini Flash type) called only on masked text
- Detokenization and API calls kept in-house
The plan, step by step
- Step 1Map the most frequent customer intents (transfer, balance, bill, spending insight) from support and app logs.Deliverable: Reference set of intents prioritized by volume.
- Step 2Build the masking layer: local transcription, tokenization, small model that detects and replaces PII before any external call.Deliverable: Tested masking pipeline, zero outbound PII measured.
- Step 3Plug in an orchestration layer that sends the masked text to an LLM and only retrieves the intent and entities.Deliverable: Intent extraction prototype on masked data.
- Step 4Execute the response and actions (account API calls) entirely in-house, after detokenization, with logging.Deliverable: Full loop in test, each action traced on the enterprise side.
- Step 5Open a beta to a segment of customers in the app with transparency on the AI nature, then widen the scope and adjust the model choice.Deliverable: Assistant in beta, human-free resolution rate tracked.
First step: Map the most frequent customer intents in the app, then set up the PII masking layer before any call to an LLM: it is what makes the architecture transposable in a regulated environment.
The same pattern, proven in other industries
ADT
genAI customer agent
Air India
genAI customer agent (automated customer service)
AT&T
Governed enterprise genAI assistant: an LLM connected to internal knowledge, open to tens of thousands of employees and called at scale by automated systems
Sources
- S1 Wells Fargo Reaches Major Digital Milestones Primary archive pending
- S2 Wells Fargo's AI assistant just crossed 245 million interactions - no human handoffs, no sensitive data exposed Established press archive pending
- S3 Wells Fargo's assistant, powered by Google's AI, poised to hit 100 million interactions annually Established press archive pending
An error, newer info, a source?
This page lives on its accuracy. If a figure has moved, if the deployment has changed, or if you have a higher-quality source, tell us. Every sourced correction is verified before publication.