Global Exchange

Designing an AI-powered conversational experience for WhatsApp support

Design and evolution of an AI chatbot on WhatsApp to reduce the Customer Relationship Center’s operational workload and improve the support experience.

Conversational AI · Customer Support · WhatsApp

Project Overview

From repetitive queries to conversations that resolve themselves

Global Exchange is a global leader in currency exchange, headquartered in Spain with a presence in 30+ countries.

This project focused on designing and launching a WhatsApp conversational agent to handle informational queries and reduce the operational workload of the Customer Relationship Center (CRC).

The product launched as an MVP in production and evolved based on real usage data and user feedback across different markets.

Product

AI conversational agent (WhatsApp)

Role

UX/UI Design Lead

Skills

Product Strategy

Conversational Design

UX Writing

Stakeholder Management

Data-driven Iteration

The challenge

Every simple query took up valuable human time.

Locations, opening hours and FAQs took up
a large share of support.

The challenge was to design a scalable conversational experience that could resolve these informational needs efficiently, while ensuring regulatory compliance, reliability and a clear handoff to human support in a multi-market context.

High volume of informational queries

Office locations, opening hours, services and FAQs took up much of the CRC’s time.

Overloaded support operations

Informational queries that required human attention, taking capacity away from sales.

Scaling without losing reliability

Automation alone wasn’t enough: compliance had to be ensured across different markets.

Product goals

Before automating, we had to learn

Four goals guided the design, from reducing operational workload to validating the channel before scaling it.

Reduce the CRC’s workload

Handle repetitive, non-sales queries through a conversational channel, freeing up the team’s time.

https://framer.link/isaacdsign

Always-available support

24/7 support on a channel users already use every day, with a mobile-first experience built for quick queries.

https://framer.link/isaacdsign

Validate WhatsApp as a channel

Rather than assuming adoption, the project aimed to validate it through real usage as a scalable touchpoint.

https://framer.link/isaacdsign

Learn before scaling

Before enabling transactional use cases, the focus was on understanding real user behavior and conversation patterns.

https://framer.link/isaacdsign

Scope & constraints

Defining what not to do is also design

A clear MVP scope was key to balancing speed, risk and learning in a regulated fintech environment.

IN THE MVP

✓ Office locations and
opening hours

✓ FAQs, services and available currencies

✓ Language and country
selection

✓ Handoff to a human agent when needed

OUT OF SCOPE

✕ Exchange rates

✕ Order creation

✕ Order tracking

To avoid risk and premature complexity.

REGULATORY CONSTRAINTS

Regulatory compliance

WhatsApp platform limitations

Constraints on
interaction patterns & UI

DELIVERY CHALLENGES

Dependency on an external provider

Coordination across AI,
support and compliance

Initial estimates far
below actual timelines

Managing it required
constant priorization
and cross-team
alignment.

Good design wasn't about deciding what to build, but having the discipline to decide what not to build yet.

DENTRO DEL MVP

✓ Ubicación de oficinas y horarios

✓ FAQ, servicios y monedas
disponibles

✓ Selección de idioma y país

✓ Derivación a un agente humano cuando sea necesario

FUERA DEL ALCANCE

✕ Tipos de cambio

✕ Creación de pedidos

✕ Seguimiento de pedidos

Para evitar riesgos y complejidad
prematura.

RESTRICCIONES NORMATIVAS

Cumplimiento normativo

Limitaciones de la plataforma
WhatsApp

Condiciones en patrones de
interacción y UI

RETOS DE ENTREGA

Dependencia de un proveedor
externo

Coordinación entre IA, soporte y
compliance

Estimaciones iniciales muy por debajo
de los tiempos reales

Gestionarlo requirió priorización continua y alineación entre equipos.

Diseñar bien no fue decidir qué construir, sino tener la disciplina de decidir qué no construir todavía.

DENTRO DEL MVP

✓ Ubicación de oficinas y
horarios

✓ FAQ, servicios y monedas
disponibles

✓ Selección de idioma y
país

✓ Derivación a un agente
humano cuando sea necesario

FUERA DEL ALCANCE

✕ Tipos de cambio

✕ Creación de pedidos

✕ Seguimiento de pedidos

Para evitar riesgos y
complejidad prematura.

RESTRICCIONES NORMATIVAS

✓ Ubicación de oficinas y
horarios

✓ FAQ, servicios y monedas
disponibles

✓ Selección de idioma y
país

✓ Derivación a un agente
humano cuando sea necesario

FUERA DEL ALCANCE

✕ Tipos de cambio

✕ Creación de pedidos

✕ Seguimiento de pedidos

Para evitar riesgos y
complejidad prematura.

Diseñar bien no fue decidir qué construir, sino tener la disciplina de decidir qué no construir todavía.

Solution

A guided conversation, not an open-ended chat

Design decisions to deliver reliable answers in a regulated, multi-market environment.

Guided entry in a regulated context
The MVP was built on a guided entry model to reduce ambiguity and risk. Asking for language and country from the start made it possible to give accurate answers and ensure compliance across multiple markets.
Impact:
More accurate answers from the first message
Regulatory compliance across markets
A predictable, easy-to-follow experience
Intent-driven design, not free text
Instead of encouraging open-ended conversation, the experience was designed around clear user intents.
Impact:
Fewer misunderstandings
Fewer error states
Clear expectations about what the bot can and can’t do
Handoff to the CRC as part of the experience
Handoff to the CRC was treated as a designed outcome, not a system failure, with clear rules to resolve, clarify or connect users with a human agent.
Impact:
Continuity between conversational and human support
Clear rules: automated resolution, clarification or handoff
Users never get stuck without a way out
Guided entry in a regulated context
The MVP was built on a guided entry model to reduce ambiguity and risk. Asking for language and country from the start made it possible to give accurate answers and ensure compliance across multiple markets.
Impact:
More accurate answers from the first message
Regulatory compliance across markets
A predictable, easy-to-follow experience
Intent-driven design, not free text
Instead of encouraging open-ended conversation, the experience was designed around clear user intents.
Impact:
Fewer misunderstandings
Fewer error states
Clear expectations about what the bot can and can’t do
Handoff to the CRC as part of the experience
Handoff to the CRC was treated as a designed outcome, not a system failure, with clear rules to resolve, clarify or connect users with a human agent.
Impact:
Continuity between conversational and human support
Clear rules: automated resolution, clarification or handoff
Users never get stuck without a way out
Guided entry in a regulated context
The MVP was built on a guided entry model to reduce ambiguity and risk. Asking for language and country from the start made it possible to give accurate answers and ensure compliance across multiple markets.
Impact:
More accurate answers from the first message
Regulatory compliance across markets
A predictable, easy-to-follow experience
Intent-driven design, not free text
Instead of encouraging open-ended conversation, the experience was designed around clear user intents.
Impact:
Fewer misunderstandings
Fewer error states
Clear expectations about what the bot can and can’t do
Handoff to the CRC as part of the experience
Handoff to the CRC was treated as a designed outcome, not a system failure, with clear rules to resolve, clarify or connect users with a human agent.
Impact:
Continuity between conversational and human support
Clear rules: automated resolution, clarification or handoff
Users never get stuck without a way out
Guided entry in a regulated context
The MVP was built on a guided entry model to reduce ambiguity and risk. Asking for language and country from the start made it possible to give accurate answers and ensure compliance across multiple markets.
Impact:
More accurate answers from the first message
Regulatory compliance across markets
A predictable, easy-to-follow experience
Intent-driven design, not free text
Instead of encouraging open-ended conversation, the experience was designed around clear user intents.
Impact:
Fewer misunderstandings
Fewer error states
Clear expectations about what the bot can and can’t do
Handoff to the CRC as part of the experience
Handoff to the CRC was treated as a designed outcome, not a system failure, with clear rules to resolve, clarify or connect users with a human agent.
Impact:
Continuity between conversational and human support
Clear rules: automated resolution, clarification or handoff
Users never get stuck without a way out

SWIPE TO EXPLORE

Guided entry in a regulated context

The MVP was built on a guided entry model to reduce ambiguity and risk. Asking for language and country from the start made it possible to give accurate answers and ensure compliance across multiple markets.

Impact:

More accurate answers from the first message

A predictable, easy-to-follow experience

New bullet point

Intent-driven design, not free text

Instead of encouraging open-ended conversation, the experience was designed around clear user intents.

Impact:

Fewer misunderstandings

Fewer error states

Clear expectations about what the bot can and can’t do

Handoff to the CRC as part of the experience

Handoff to the CRC was treated as a designed outcome, not a system failure, with clear rules to resolve, clarify or connect users with a human agent.

Impact:

Continuity between conversational and human support

Clear rules: automated resolution, clarification or handoff

Users never get stuck without a way out

Results

What the MVP taught us, in numbers

The chatbot launched as an MVP in production, allowing us to observe real user behavior and assess the impact of conversational support throughout 2025.

Total conversations handled

Driven by informational queries about offices, opening hours and services, handled through the WhatsApp channel during 2025.

Automatically resolved queries

Resolved in Spanish with no human intervention, reducing the workload of the Customer Relationship Center (CRC).

Conversations handed off to the CRC

Handed off to human agents when queries went beyond the chatbot's informational scope, ensuring users always got an answer.

Product impact

From MVP to the next iteration

The chatbot introduced a new first layer of customer support, redefining how informational queries were handled across channels.

A lighter load for the CRC

Diverted repetitive queries, freeing up CRC capacity for higher-value cases.

Clearer support channels

A clear distinction between automated informational support and human-assisted service.

WhatsApp validation

Confirmed as a reliable, mobile-first support channel for informational queries.

A foundation to evolve

A solid base to take the experience beyond the MVP, backed by real usage data.

These results confirmed that the initial MVP scope was right, while revealing clear opportunities to improve how users start and navigate the conversation.

The next step wasn't adding more features, but refining the conversational flow based on real usage patterns.

Flow 2.0

Data-driven iteration

Learnings from real usage data and user feedback showed that the initial conversational flow, while effective, added unnecessary friction at the entry point. Instead of adding features, the focus shifted to improving how users start and move through the conversation.

What the data revealed

Conversation analysis revealed consistent patterns: users preferred asking directly rather than browsing predefined options, and many conversations followed similar informational paths.

Findings

The problem was how the conversation started
Structured choices slowed down the first interaction

Key product decision

Based on these insights, I led the shift from a menu-based entry model to a conversation-first model, reducing friction at the start without losing control where compliance and human handoff required it.

Impact

Users could express their intent naturally
Laid the groundwork for the next iteration of the conversational system

What the data revealed

Conversation analysis revealed consistent patterns: users preferred asking directly rather than browsing predefined options, and many conversations followed similar informational paths.

Findings

The problem was how the conversation started
Structured choices slowed down the first interaction

Key product decision

Impact

Users could express their intent naturally
Laid the groundwork for the next iteration of the conversational system

System design

Hybrid conversational model

A system that combines free-text input with structured logic to balance flexibility and control in a fintech environment.

The conversational structure was defined and validated before implementation, in close collaboration with internal teams and the AI provider.

How the hybrid model works

Free-text input as the starting point, for a natural interaction

AI-led resolution of informational queries

Structured logic for legal requierements, inactivity and handoff to the CRC.

Why a hybrid approach

Reduces ambiguity without over-constraining users

Improves the speed and clarity of the first interactrion

Keeps behaviour reliable in regulated scenarios

This model lets the chatbot act as a support assistant, rather than a rigid menu or an uncontrolled conversational interface.

Designing an AI-powered conversational experience for WhatsApp support

Designing an AI-powered conversational experience for WhatsApp support

Designing an AI-powered conversational experience for WhatsApp support

From repetitive queries to conversations that resolve themselves

From repetitive queries to conversations that resolve themselves