Savi Agentic AI
August, 2025
Project Focus
Enhancing the AI chat experience in Savi Finance
Project Scope
My Role
Design Process
Design the Feature
The Next Step Forward
I worked as a team member of 4 designers at Savi Finance, a platform that helps users manage finances across institutions, track spending and savings, and monitor financial health. We initiated the development of an AI chat assistant to provide solutions, with the challenge of making interactions feel more natural and human.
As a UX designer on the team, I contributed to shaping the user experience of the AI chat assistant. My work centered on designing the chat interface, exploring personas for the AI agents, and contributing to the flow of interactions.
I focused on secondary research on Voice UX, human AI interactions, big companies’ design principles and competitor analysis
I created 5 principles for human-AI communications based on research
I created major user flows and Information Architecture
I created AI agent personas
Research
Define
Ideation
Design
Duration
4 weeks
Team
4 UX Designers
Tools
User Research
My research began with a review of conversational design and responsible AI guidelines from major industry leaders, including Google, Microsoft, and Nielsen Norman Group, focusing on how AI personality, tone, and communication style can create more human-like interactions.
My idea for AI personas came from a simple analogy: when people meet different bank agents, they naturally connect with the one whose communication style feels right, and switch if the interaction style does not suit them.
Sometimes users do not know what questions they should ask. Co-creating the question, or repair conversation is the key
Creating tones and personality will make the conversation more human
Applying this concept to chatbot design, I created AI personas that allow users to select an agent whose style aligns with their preferences—whether they value conciseness, creativity, or collaborative problem-solving.
To bring the idea of different AI agents to life, I treated them almost like real people, giving each one a distinct identity.
Based on the personas I created for the agents, I made three cards on the greeting section of the AI chat first screen.
Flexibility
User choice
Trust
My design Principles
I also explored real-world apps for inspiration on making AI interactions more engaging, human-like, and responsive.
Conversational Persona
Helps AI maintain a consistent personality and tone to build trust and feel more human.
Human-Centered AI
Emphasizes empathy & personal connection in AI interactions to foster meaningful engagement.
User-Centric Prompts
Provides frameworks for crafting clear and effective AI responses aligned with user expectations.
Duolingo
Learning app
Duolingo uses haptics, animations, and playful sounds to make users feel acknowledged and cared for, creating engaging interactions.
Erica
(Banking app)
Erica demonstrates how users can interact via chat or voice, providing flexibility and creating a sense of personalized assistance
Replika
AI chat platforms
Replika demonstrates how AI can make conversations feel more human and emotionally engaging with defined personas
Hi Emily,
We are here to help you!
We choose an agent based on input, but you can switch anytime!
I am Marcus
I’m the analyst!
I give you straight answers,
backed by numbers and facts
My name is Evelyn.
I am your mentor
I guide you with experience
and help you stay on track.
I’m Sophia
a visual explainer~
I make complex information
easy to see and understand.
I am Marcus
I’m the analyst!
I give you straight answers,
backed by numbers and facts
My name is Evelyn.
I am your mentor
I guide you with experience
and help you stay on track.
I’m Sophia
a visual explainer~
I make complex information
easy to see and understand.
Agent
Identity
image / Voice
Response Style
Interface Style
Marcus
the Analyst
male in 30s,
technical tone
straightforward,
logical, efficient
minimal text
clean, structured
layout
Evelyn
the Mentor
female in 50s,
warm and reassuring
supportive, patient,
elaborates, reframes questions, co-creates
soft colors, conversational feel
Sophia
the Explorer
female in 20s,
light and fast-paced
energetic, engaging, uses visuals,
creative examples,
dynamic layout with more visuals
After creating distinct AI agents with personalities, I was satisfied that users could choose the style of interaction that suited them best.
But this raised a new challenge:
should users really be responsible for making that choice every time?
I thought back to the interaction with ChatGPT models like 4.0, 5.0, or mini. Early on, many users were confused about which model to pick, or grew tired of comparing them, since choosing a model was an extra cognitive step before even starting the conversation.
I realized my design was running into the same issue—users shouldn’t have to think about which “agent” to use on top of forming their own question.
That insight shifted my design from user-selected personas to AI-driven thinking models. Instead of asking people to choose an agent first, the system now analyzes their question and automatically activates the most suitable agent to respond.
While the AI can automatically select the most suitable agent, users can override the system and freely choose the persona they trust most. This balance respects both efficiency and personal preference.
On the welcome screen, all AI personas appear together to greet the user, signaling that the full team is standing by. When a question is asked, all agents “hear” it, but only the one best suited to the request steps forward to respond. This creates a sense of collective support while keeping the interaction focused and efficient.
This project was completed within a four-week timeframe, which limited the scope to research, brainstorming, ideation, and initial prototyping. If given more time, my next steps would focus on three areas:
User Testing – Conduct usability testing with at least five participants to validate assumptions, observe interaction patterns, and refine the agent-selection model.
Interface Development – Expand on the core design by creating more detailed interfaces for each persona, including options for users to switch agents seamlessly in the middle of a conversation.
Personalization – Explore adaptive interface styles that adjust to different user preferences, ensuring the system feels responsive and tailored.
Secondary Research
The Bank Agent Analogy
Three Personas
The AI Agents
The ChatGPT models
Service from a Whole Team
My Understanding
Competitor Analysis
Prototype
My Area of Design
A Client Project that helps Savi Finance design a virtual financial assistant that communicates with a more human touch
