Case Study ✷ 03 of 03

Bringing generative AI to Wells Fargo's virtual assistant

Audience: B2C
Role: Senior Content Designer
Team: 1 Product Manager, 1 Conversation Designer
Timeline: January 2025 to June 2025
Tools: Figma, Dialogflow

Context

Fargo, Wells Fargo's virtual assistant, originally relied on a limited set of human-written responses. Off-script questions got a generic fallback: "I'm not able to help with that yet, but I'm always learning." For this MVP, we focused on two high-confusion topics: Passkey sign-in and Save As You Go.

Challenge

Help users get helpful, accurate answers within the assistant, without needing to call support.

Solution

Use generative AI to scale Fargo's response coverage, dramatically reducing fallback messages and improving the customer experience. Launched in phases, from internal pilots to a customer pilot, validating and improving at each step.

✷ Training the AI with grounding documents

To ensure AI-generated answers were accurate and aligned with policy, I reviewed help content and user logs to identify common questions, then created grounding documents: source material the AI relies on to generate safe, accurate, on-brand responses.

Grounding document titled 'Top questions' answering what a passkey is, why it's more secure and convenient than a password, how to set one up, and what it can be used for at Wells Fargo.

✷ Prompt engineering

While testing, Legal flagged responses implying that passwords are "weak" or "easily stolen." I traced it to a vague instruction the LLM was misinterpreting. The grounding documents never said passwords were weak, only that passkeys are more secure. I rewrote the prompt to affirm the benefits of passkeys without downplaying password security, which resolved the issue while staying aligned with both grounding and legal standards.

Comparison of the original vague prompt instruction, 'Do not imply that passwords are inherently vulnerable or can be easily guessed,' next to the rewritten instruction that explicitly avoids implying passwords are insecure while still affirming passkeys are more secure.

Testing the rewritten prompt directly in the playbook editor to confirm the model's output stayed aligned.

Dialogflow playbook editor showing the tool instructions and a test invocation of the query 'are passwords weak,' with the model's response confirming passkeys are more secure without disparaging passwords.

✷ Disclosures

Every AI-generated response included a banner linking to a disclosure. Crafting this required multiple rounds with Product and Legal. Product wanted a friendly, concise tone, while Legal had concerns about overly casual language. Through collaboration, we landed on a version that satisfied both clarity and compliance.

Final version

Final, shorter disclosure: a compact 'Response created by generative AI. Learn more' banner leading to a concise two-paragraph explanation, replacing the earlier dense legal block.

Earlier drafts

Earlier, longer disclosure draft titled 'Generative AI responses' and 'How Fargo uses generative AI,' with a dense Gen AI Beta Terms and Conditions block.

Outcome

The project earned praise from the Head of AI and Design Leadership, and laid the groundwork for expanding generative AI coverage across more assistant topics, making Fargo more scalable and helpful.