AI customer service assistant

Giving a collections chatbot a voice customers trust

Customers behind on their car payments needed help at all hours, and agents couldn't always give clear answers under pressure. I led design for an AI chat assistant that could take payments, set up payment plans, and explain the account in plain language. I built its persona from our best agents' calls and tested its tone with the agents themselves.

Role: Design lead
Year: 2024
Scope: Conversation design, AI persona, research, UI
Platform: Mobile web chat

Four chat screens from the assistant: welcome messages, adding a card, the menu with a live agent option, and feedback.

At a glance

persona traits, each tested at both extremes

call center agents in the tone study

AI assistants reviewed

Challenge

Customers who fall behind on a car loan face high-stakes decisions, often without understanding their options. Help depended on call center hours, and even strong agents made mistakes under pressure. The business wanted a 24/7 assistant that could help before a missed payment turned into repossession.

The hard part was the voice. A collections bot that sounds robotic feels cold. One that sounds too cheerful feels fake.

Collaboration

I led design for the assistant: research, persona, conversation flows, and the chat interface. I worked with product, operations, process, and compliance to define what the assistant could and couldn't do.

Designing for two sides

The assistant had to calm customers and protect the business.

Customers

Behind on payments and stressed, often short on time and unsure what their options mean.

The business

Needed consistent, compliant answers around the clock, without adding call center load.

Research

I listened to calls from our five top-performing agents and reviewed six AI assistants from banks and tech companies.

Top agents lead

They speak first, explain what happens next, and never leave dead air.

Plain language wins

The best agents explained interest and repossession dates in simple terms.

People want an exit

Assistants that hid the live agent option frustrated users most.

Trust needs control

People trust AI more when they understand how it works and can always reach a person.

Key insight

The assistant should sound like the best version of our top agents, the ones customers already respond well to.

Design process

1. Listen
- Top-agent calls
- Competitor review
2. Define
- Persona and traits
- Chat rules
3. Design
- Conversation flows
- Chat interface
4. Validate
- Tone study with agents
- Scripted edge-case tests

Defining the persona

Six traits, each with a way to fail on both sides. The goal was the balance point.

Too littleTargetToo much

DisrespectfulRespectfulObsequious

ReservedEnthusiasticOverzealous

IncoherentArticulateVerbose

ArtificialNaturalOverly formal

DistantEmpatheticOver-involved

IncompetentExpertKnow-it-all

Designed for

People with busy lives who are struggling to stay on top of their finances. Often one unexpected bill put them behind, and they need clear help fast, without judgment.

Chat rules built for trust

People should always know they're talking to AI, what it can do, and how to reach a person.

Speak first

The assistant greets, says what it can help with, and sets expectations.

Short and scannable

Brief messages, with quick replies for common answers.

Always an exit

A live agent is one tap away whenever the assistant can't help.

Honest about being AI

It says it's an AI assistant and that chats may be recorded.

Testing the voice

Seven call center agents compared three tones across six moments, from the first message to the closing recap.

Transactional

I'm an AI chatbot. You owe $750.00 and are 64 days delinquent.

Agent feedbackAgents mostly chose the other two.

Balanced

Hi, I'm your AI assistant. I can help with payments and payment plans. Let's take care of that payment today.

Agent feedbackSeems to be in the middle of transactional and over affectionate.

Warm

Hey there! You're not alone. Let's handle this together.

Agent feedbackHas a personal touch. The last one does sound a little corny.

Votes split between balanced and warm. Agents wanted it human and reassuring, but brief. The final voice kept the balanced structure and added warmer moments, like telling customers when a payment stopped a repossession.

What agents said

Isn't too automated sounding.
Letting the customer know that they prevented repo is a great idea.
The recap is great because that is the way a call with an advisor would end.

I also wrote scripted tests for each trait, from a complex loan question to a customer in crisis, so the team could check the model stayed in character.

Key features

Everything a customer needs to resolve a late payment, inside one chat.

Start screen introducing the assistant and what it can help with, text blurred. — **Clear start.** Says what it can help with first.

Welcome messages with quick replies, text blurred. — **Guided next step.** Quick replies for common answers.

Adding a new card without leaving the chat, text blurred. — **Pay in chat.** Use a saved card or add one.

Chat menu with download chat, mute, call live agent, and end chat. — **Always an exit.** Call a live agent from the menu.

Thumbs down feedback with reasons to select, text blurred. — **Feedback loop.** Ratings with reasons help train the model.

End of chat recap and overall rating, text blurred. — **Closing recap.** Ends like a call with a top agent.

Placeholder text blurred.

Where it landed

The project was paused before launch, with the persona, chat rules, conversation flows, and test plan complete.

Ready to build

Persona, rules, and flows documented

4 test methods

Scripts, agent testing, sentiment analysis, monitoring

Reflection

What I learned: trust comes from clarity and control, so the assistant always explained what it could do and offered a person. What I'd do differently: test the voice with customers as well as agents, and tailor explanations to different levels of financial literacy.