Design concept · conversational UX

What if Félix had
an AI assistant?

A UX exploration by Tom Oliva — showing how conversational AI could extend Félix's WhatsApp-first experience to the web, helping users understand the service before their first transfer.

Try the assistant ↓ See the case study
🇲🇽 México 🇬🇹 Guatemala 🇭🇳 Honduras 🇨🇴 Colombia 🇸🇻 El Salvador 🇩🇴 Rep. Dominicana 🇪🇨 Ecuador 🇳🇮 Nicaragua

Ask anything about sending money.

This assistant knows Félix's service end to end — how transfers work, which countries are supported, pickup locations, payment methods, and security. It responds in English or Spanish, based on what you write.

guardrail

The bot never pretends to process real transfers. It always redirects to WhatsApp for actual transactions — transparent about its limitations.

language

Detects and responds in the user's language automatically — key for a product serving Spanish-speaking migrants from a US-based interface.

suggested replies

Reduces the cognitive load of a blank input — especially important for users who may be unfamiliar with chatbots.

trust & clarity

Every response about security is direct and reassuring — matching the tone Félix uses across their brand ("confiable 100%").

F
Félix Assistant
Online · EN / ES
Demo concept · Real Félix transfers at felixpago.com

The UX thinking behind this concept.

Félix is already a conversational product — its core experience runs on WhatsApp. This concept explores how that same conversational model could extend to the web, helping users understand the service before their first interaction. The design challenges here are identical to the ones I solve in fintech UX every day.

01

Trust is the primary UX problem

Sending money internationally requires complete trust in the service. Every UX decision — from response tone to error handling — needs to reinforce reliability. I applied the same principles I used designing Shell Box's payment flows.

02

The blank input is a barrier

Félix's users may not be tech-savvy. A blank chat input with no guidance creates friction before the conversation starts. Suggested replies lower the entry barrier — same problem I solved in loyalty redemption flows.

03

Guardrails build confidence

A bot that tries to do everything fails at everything. This assistant is explicit about what it can and can't do — and redirects clearly. That transparency increases user confidence, not decreases it.

04

Language switching is a UX decision

Félix serves Spanish-speaking users from an English-language infrastructure. The assistant auto-detects language per message — no toggle, no friction. The channel adapts to the user, not the reverse.

05

The channel is WhatsApp — this extends it

This isn't a replacement for Félix's WhatsApp flow — it's an extension. Users discover and understand the service here, then complete the transfer where Félix already lives.

06

Built in days, not months

This concept was designed and deployed as a working prototype in under a week. The speed came from applying existing UX knowledge to a new channel — not from learning everything from scratch.

Tom Oliva · Senior Product Designer

7+ years designing complex fintech products — payments, loyalty, remittances, and backoffice tools. This concept was built to demonstrate that conversational UX is a channel, not a separate discipline. The thinking is the same. The interface is different.

Conversational UX · Case Documentation · 2025

Félix Pago — Conversational Design Documentation

End-to-end CxD documentation: intent architecture, dialogue flows, voice & tone, metrics framework, Wizard of Oz testing methodology, and Voiceflow-ready specification.
Designer
Tomás Oliva
Role
UX / Conversational Design
Channel
Web chat + WhatsApp-native logic
Deliverable
Working prototype + CxD documentation
01 Context 02 Intent Architecture 03 Dialogue Flows 04 Voice & Tone 05 Metrics 06 Wizard of Oz 07 Voiceflow Spec 08 Reflection
01

Context & Problem

Félix Pago is a remittance service that operates 100% through WhatsApp — Latino immigrants in the US send money to Latin America without downloading any app. The product is conversational at its core: WhatsApp is the interface, not just the channel.

The design opportunity: potential users need to understand the service before initiating their first WhatsApp transfer. The existing help center was article-based — passive, search-dependent, not optimized for someone new to the product.

Why this is a CxD problem, not just a UI problem

In a traditional UI, users navigate menus and forms. In Félix's model, the interface is dialogue. The same UX principles — progressive disclosure, error prevention, trust signals — apply to conversation design. Instead of screen states, you design turns. Instead of button labels, you write utterances. The translation is direct.

02

Intent Architecture

12 core intents mapped from the actual Félix Pago help center and user journey, covering the full transfer lifecycle plus support scenarios.

Intent IDUser utterances (examples)EntitiesPriorityFallback path
HOW_TO_SEND"How do I send money?" / "¿Cómo envío dinero?"HighGeneral_Fallback
FEE_INQUIRY"How much does it cost?" / "¿Cuánto cobran?"country, amountHighFee_Clarification
TRANSFER_TIME"When does the money arrive?" / "¿Cuánto tarda?"delivery_methodHighGeneral_Fallback
COUNTRY_CHECK"Can I send to Peru?" / "¿Envían a Colombia?"countryHighCountry_NotAvailable
PAYMENT_METHOD"How do I pay?" / "¿Aceptan efectivo?"payment_typeHighGeneral_Fallback
SECURITY_TRUST"Is Félix safe?" / "¿Es seguro?"MediumGeneral_Fallback
SEND_LIMIT"How much can I send?" / "¿Cuál es el límite?"verification_levelMediumGeneral_Fallback
CANCEL_REFUND"Can I cancel?" / "¿Puedo cancelar un envío?"MediumEscalate_Agent
PICKUP_LOCATIONS"Where can they pick up cash?" / "¿Dónde retiran?"country, cityMediumGeneral_Fallback
TRACK_TRANSFER"Where's my transfer?" / "¿Ya llegó?"RedirectEscalate_WhatsApp
CONTACT_AGENT"Talk to a person" / "Hablar con agente"RedirectEscalate_Agent
OUT_OF_SCOPE"Can I send a top-up?" / "¿Hacen recargas?"FallbackOOS_Clarification
Design decision — redirect intents

TRACK_TRANSFER and CONTACT_AGENT are redirect intents, not informational ones. The bot cannot access real account data — the correct UX decision is transparency about that limitation and a direct route to WhatsApp. Pretending to have access would erode trust.

03

Dialogue Flows — Turns & Fallbacks

Flow 1 — Happy path: Fee inquiry with country entity resolved
Intent: FEE_INQUIRY · Entity: country=Mexico · Turns: 2
USER
How much does it cost to send money to Mexico?
Turn 1 · Intent: FEE_INQUIRY · Entity: country=Mexico ✓
FÉLIX
💰 Sending to Mexico costs $2.99 USD for bank deposits, or $4.98 USD for cash pickup at OXXO, Walmart, Elektra, and 40+ other locations.

Your first transfer is always free. The fee is shown before you confirm — no surprises.

→ Start your first free transfer: felixpago.com
Turn 2 · Informational · CTA: product redirect · End of turn
SYSTEM
Suggested replies: "How long does it take?" · "Can I pay with cash?" · "Which banks can they receive from?"
User-perspective suggestions — not bot-perspective
Flow 2 — Clarification path: Missing entity (country not specified)
Intent: FEE_INQUIRY · Entity: country=null → clarification required · Turns: 4
USER
¿Cuánto cobran de comisión?
Turn 1 · Entity: country=null → clarification needed
FÉLIX
💰 La comisión depende del país destino. Para los más comunes:

México y Colombia: $2.99 USD a cuenta bancaria · El Salvador: $3.99 + 1.25% · Honduras: varía según el monto enviado

¿A qué país querés enviar?
Turn 2 · Clarification · Shows top options, asks for entity · No personal data requested
USER
A Honduras
Turn 3 · Entity: country=Honduras ✓
FÉLIX
💰 Para Honduras la comisión varía según el monto: hasta $200 → $5.99 · hasta $400 → $7.99 · hasta $1,000 → $11.99 · hasta $3,000 → $27.99.

La comisión siempre se muestra antes de confirmar. El primer envío es gratis.
Turn 4 · Intent resolved · 4-turn clarification complete
Flow 3 — Fallback & escalation: Out-of-scope + agent handoff
Intent: OUT_OF_SCOPE → CONTACT_AGENT · Escalation path
USER
Can I send a phone top-up to my brother in Guatemala?
Turn 1 · Intent: OUT_OF_SCOPE · Topic: mobile top-up
FÉLIX
❌ Phone top-ups aren't available with Félix right now — only bank deposits and cash pickup.

Want me to explain how sending to Guatemala works instead?

→ Or chat directly: wa.me/16693333549
Turn 2 · OOS with redirect · Honest, offers alternative, no dead end
USER
I need help with a transfer that's stuck
Turn 3 · Intent: TRACK_TRANSFER → CONTACT_AGENT
FÉLIX
📞 For a stuck transfer I need to connect you with the Félix team — I can't access real account data from here.

→ Message Félix support: wa.me/16693333549

Type "hablar con agente" and a real person will help right away. Available 24/7.
Turn 4 · Graceful escalation · Transparent · Clear CTA · No loop
Fallback design principle

Every fallback ends with a concrete next action — an alternative within scope, or a direct path to a human agent. The bot never leaves the user in a conversational dead end. This directly addresses the core finding from the LATAM diagnostic: "no escalation at the right time" is the primary driver of conversational frustration.

04

Voice & Tone

Félix Pago serves Latino migrants in the US — people sending money home, often with anxiety about whether the transfer will arrive safely. The voice needs to be warm and trustworthy, not patronizing or corporate.

Voice principles
Tone comparisons
❌ Avoid
"Hi! Welcome to Félix! I'm here to help you with all your money transfer needs. Our service is available 24/7! What would you like to know today? 😊"
✓ Félix voice
"Hi — I can help you understand how Félix works. Transfers, fees, pickup locations, anything else. What do you need?"
❌ Avoid — Error
"We apologize for the inconvenience. Our system is unable to process this request. Please try again later or contact our customer service department."
✓ Félix voice — Error
"Connection error — try again in a moment. If it keeps happening, the Félix team can help: wa.me/16693333549"
❌ Avoid — OOS
"I'm sorry, but unfortunately that feature is not currently available within our service offering at this time."
✓ Félix voice — OOS
"Phone top-ups aren't available right now — only bank deposits and cash pickup. Want me to explain how cash pickup works?"
Microcopy rules
05

Metrics Framework

Target KPIs defined for post-launch measurement. Since this is a working prototype, these represent the measurement methodology I'd implement and present to stakeholders.

Task Completion Rate

≥ 80%
Users who get a satisfactory answer without escalating. Measured by session end state.

Fallback Rate

≤ 15%
% of turns triggering a fallback. High fallback = gaps in intent coverage or training data.

Escalation Rate

≤ 20%
% of sessions routed to human agent. The goal is right-time escalation, not zero escalation.

Avg. Turns to Resolution

≤ 3
More than 4 turns suggests unclear intent design or missing entity handling.

Containment Rate

≥ 70%
Sessions resolved within bot. Informational intents should approach 100%.

CSAT Post-session

≥ 4.2
Single question post-session rating (1-5). Triggered after session ends or idle.
Confusion Matrix — Intent classification accuracy
Predicted →
HOW_TO_SEND
FEE_INQUIRY
CANCEL_REFUND
OUT_OF_SCOPE
HOW_TO_SEND
✓ 92%
5%
2%
1%
FEE_INQUIRY
8%
✓ 88%
3%
1%
CANCEL_REFUND
3%
6%
✓ 85%
6% →OOS
OUT_OF_SCOPE
4%
2%
1%
✓ 93%
Key insight

CANCEL_REFUND has the highest error rate — 6% of real cancellation requests route to OUT_OF_SCOPE. This is a high-stakes misclassification. Fix: expand cancellation utterance examples and add a disambiguation prompt when confidence is below 0.7 threshold.

06

Testing Methodology — Wizard of Oz

Before deploying the AI-powered prototype, I'd validate conversation design using Wizard of Oz (WOz): a researcher manually responds to user inputs in real time using a prepared response library. The user believes they're interacting with an automated system. This surfaces dialogue issues before expensive model training.

1
Prepare response library
Map all primary intents to pre-written responses. Include: happy paths, clarification prompts, fallbacks, and escalation paths. Researcher has instant access during the session.
2
Recruit participants
5–8 participants matching target persona: Spanish-speaking immigrants in the US, ages 25–45, who regularly send money to Latin America. Include first-time and experienced senders.
3
Run scenario tasks
"You want to send $200 to your mom in Mexico. She'll pick it up at OXXO. Use the assistant to understand the process and cost." Observe without guiding. Log silence >5s, re-phrasing attempts, and abandonment points.
4
Debrief and synthesize
5-minute post-session: "What was unclear?" / "What did you expect it to say?" / "When did you feel confused?" Synthesize into dialogue revisions before model training.
What WOz tests that analytics can't
07

Voiceflow Architecture

Main flow structure
▶ Start
💬 Welcome
🎯 Intent capture
◇ Route by intent
💬 Intent response
🎯 Follow-up capture
⚠ Fallback (3rd attempt)
📞 Escalate to agent
■ End session
Block definitions
Block typeUsageKey settings
SpeakAll bot responsesLanguage: ES/EN via variable. Max 3 sentences per block.
IntentCapture user input after each speak12 intents. Confidence threshold: 0.7. Below → Clarification block.
ConditionRoute by intent, check entity presenceIf country=null AND intent=FEE_INQUIRY → Fee_Clarification flow
Variablesuser_language, detected_country, session_intent, fallback_countfallback_count increments on unrecognized input. At 3 → force escalation.
API BlockCall Claude API for dynamic responsesSystem prompt with Félix knowledge base. Max tokens: 300.
ButtonsSuggested replies (3 max)User-perspective options only. Never "Tell me more."
No MatchInput doesn't match any intent ≥0.71st miss: clarify. 2nd miss: offer categories. 3rd miss: escalate.
No ReplyUser idle >30 seconds"Still there? Take your time — or start fresh when you're ready."
System prompt for API block
// Injected into Voiceflow API block // Variables: {user_language}, {session_intent}, {detected_country} SYSTEM: "You are Félix, assistant for Félix Pago remittance service. Respond in: {user_language} Current intent: {session_intent} Detected country: {detected_country} Rules: - Max 3 sentences. No markdown. No asterisks. - Start with relevant emoji on its own line. - Never ask for account numbers or personal data. - If out of scope: redirect to wa.me/16693333549 - End with: SUGGESTIONS: option1 | option2 | option3"
Fallback escalation logic
IF intent.confidence < 0.70: SET fallback_count += 1 IF fallback_count == 1: → Clarification prompt ELIF fallback_count == 2: → Category menu ELIF fallback_count >= 3: → Force escalation to human agent → Speak: "Let me connect you with a Félix agent." → CTA: wa.me/16693333549 · END session
08

Reflection

This case demonstrates something I strongly believe: conversational UX is not a separate discipline from product design. The core skills are identical — understanding user goals, defining clear states and flows, handling errors gracefully, building trust through transparency.

What's different is the medium. Instead of screen states, you design turns. Instead of button labels, you write utterances. The experience designing fintech flows at Shell Box and Ewally mapped directly onto this conversational context.

The argument in one sentence

I designed this assistant end-to-end — from intent architecture to voice principles to a working deployed prototype — in under a week, using existing product design skills applied to a new channel. That's not a gap in experience. That's evidence of transferable craft.

What I'd do differently with more time