← All Works
9:41 AI Service Navigator GPT-4 · City of Toronto Today 9:41 AM I need shelter tonight AI Nearest shelter found. Red Cross · 0.4km Merci, je vais là-bas Message... I M P A C T 77.8% fewer unnecessary follow-up questions 88.9% of responses rated clear and helpful 100 survey responses across Toronto U S E R S NEWCOMER IRCC · Settlement STUDENT Bus Routes · Food SENIOR Services · Support SERVICE NAVIGATOR VOICEFLOW GPT-4 CITY OF TORONTO 100 SURVEYS · 2024
UX Research  ·  Interaction Design 2024

AI-Powered Service Navigator

Designing a chatbot that guides newcomers, students, and families to the services they need in any language, at the moment they need it.

Voiceflow GPT-4 100 Surveys Usability Testing Multilingual City of Toronto
The Problem

Finding help should not feel like homework.

When a crisis hits, the last thing someone needs is five government websites in a second language. A newcomer navigating IRCC queries for the first time. A student trying to figure out bus routes in an unfamiliar city. A family looking for emergency shelter on a Tuesday night.

The information exists across dozens of portals. Getting to it is a completely different story. The question was not whether we could build a chatbot. It was whether we could build one that actually worked for the people who needed it most.

IRCC.ca
toronto.ca
211.ca
redcross.ca
cra-arc.gc.ca
acces.ca
? Where do I even start?
Who It Was Built For

The people most in need are usually the last ones designed for.

This tool was designed for people navigating unfamiliar systems, often under pressure, often in a language that is not their first.

🌍 Newcomer

Navigating IRCC, settlement services, and legal aid for the first time in a new country.

🎓 Student

Looking for bus routes, food programs, and community resources in an unfamiliar city.

🏠 Family in Crisis

Urgently searching for emergency shelter, food access, or financial assistance.

100 Survey Responses

Conducted across diverse user groups in the City of Toronto before any design decisions were made.

Research

100 surveys. One clear message.

Before touching the interface, I ran 100 surveys to understand how people currently find community services. A few things became clear very quickly.

Most people had no idea which organisation to contact first. Non-native speakers reported feeling completely lost on government portals. Emergency users gave up after two or three failed attempts. And elderly and low-literacy users relied entirely on word of mouth rather than any digital tool.

These were not usability problems. They were access problems. The chatbot needed to do the navigation work so users did not have to.

Before: What Was Broken

Three patterns that were failing real users.

Problem 01

The Question Loop

An emergency user needed shelter that night. Instead of help, they got questions. Four of them. Before a single useful answer appeared.

Service Navigator
I need shelter tonight
What type of shelter are you looking for?
Emergency
What area of the city are you in?
Downtown
Do you have any children with you?
...why won't it just help me?
4 questions in. Still no answer.
Problem 02

Dead End Responses

Many queries returned a vague message about connecting to a resource, with no link, no phone number, and nothing to act on.

Service Navigator
Where can I get food tonight?
I'll connect you to a resource.
●   ●   ● No link. No number. Nothing to act on.
User left without any help.
Problem 03

No Urgency Detection

The system treated a crisis the same as a casual query. It could not tell urgent from exploratory, or emergency from curiosity.

"I need food tonight"
Generic response
=
"I'm looking into food programs"
Generic response

Both got the exact same reply

The Fix

Rebuilding the decision tree around one principle: urgency first, clarification second.

If someone's phrasing signalled a crisis, they received a direct answer before being asked anything. That single shift eliminated the loop entirely.

User Message
Immediate Emergency
Helpline instantly
Chatbot-Assisted Emergency
2-step + links
Borderline Clarification
1 question then help
General Support
Full conversation
Category What it means What happens
Immediate Emergency Life or safety at risk right now Helpline number instantly with zero follow-up questions
Chatbot-Assisted Emergency Urgent but not life-threatening Two-step response with direct resource links
Borderline Clarification Needs one question to narrow down One short question, then a direct resource
General Support Exploratory, not urgent Full guided conversation
One Big Addition

Any language. No translation required.

Before this project, the chatbot worked in English only. For newcomers and immigrants, that was a barrier before the conversation even started.

After integrating GPT-4, users can type in any language they are comfortable with. The chatbot understands and responds in kind. For someone navigating IRCC queries or looking for settlement services for the first time, this is not a small feature. It is the whole point.

Hindi मुझे आज रात आश्रय चाहिए
French J'ai besoin d'aide ce soir
Arabic أحتاج مساعدة الليلة
English I need help tonight
Service Navigator understands all of them
Results

Before and after usability testing across 100 participants.

77.8% Fewer follow-up questions in emergency scenarios
83.3% Users confirmed location-based responses were accurate
84.4% Rated their overall experience as highly satisfying
88.9% Said responses were clear, relevant and helpful
Try the Live Prototype  →
← All Works Next Project →