DE-IDENTIFIED CASE
WHAT I CHANGED
The name and logo, now an invented brand
Typography
The color palette
Illustrations and images
Names, IDs and every piece of data
Internal terms in the copy
WHAT I KEPT
The structure and layout of every screen
The flows, step by step
Interactions and states
The research, the decisions and the reasoning
Measured results as reported; projections labeled as a model
Role
Senior Product Designer, owning the product’s design since I joined: research, UX/UI, conversation design and the design of its AI layer.
Team
Product Owner (client side), engineering, operations and business stakeholders, a fellow researcher
When
Nov 2025 – ongoing
Status
Leading the redesign. The proposal is now the product’s direction.
The problem
This advisor platform is an operational, guided-support tool that helps advisors resolve issues in real time during Banco Vecino’s digital loan application process. When I joined, its console did four things: reschedule a visit, edit the customer’s address, offer a rough guide for the call, and log the outcome against a short list of categories.
Everything else happened somewhere else. The embedded map had outdated streets, so advisors located homes in Google Maps, in another tab. What they needed about the customer and the field agent was spread across several screens. Calls ran long, and many ended without resolving the case.
It was used because the process required it, not because it helped anyone do their job.
4 actions
all the console could do when I took it over
What I found
With a fellow researcher, I mapped the process end to end, audited the data, shadowed advisors on live calls, and ran interviews, surveys, think-aloud sessions and field tests. There are three main findings that emerged from the research.
The tool was slower than the work. Most of a call went to looking things up: the home on a second map, the customer’s history on another screen, the field agent’s notes on a third. And confirming who was on the line had no method at all, in a regulated process.
Within this unit, the bank possessed the structure and human capacity to provide support throughout its lending processes. This potential was not being leveraged, and the digital product they were using could have served a much greater purpose than simply rescheduling a visit or locating a client’s address on a map.
One channel, one script, every case waiting for a person. The volume was the bottleneck, and the conversation itself had never been designed.
Each one became a decision.
Three decisions
01
Decided: expand what an advisor can do without leaving the call. An address editor with several map sources. Identity verification with knowledge-based questions that confirms who’s on the line without undue friction or exposing the customer’s data, on calls and on WhatsApp. And a customer-detail panel that brings the customer’s and the field agent’s information into one place.
Why: the screens weren’t what made calls slow; the trips out of the tool were. Each one cost minutes and broke the conversation, and every lookup brought inside gave that time back on every call. Calls got about 10% shorter, and locating a home went from 15 to 10 minutes.
Discarded
A visual refresh of the same four actions.
02
Decided: take the console into origination. Advisors reach the customer who got stuck mid-application, while they’re still in it, by call or WhatsApp, and guide them step by step with components built for it: an accompaniment card that walks through each step, and an assisted call that closes a pending verification without a second home visit.
Why: a generic reminder treats every abandonment as the same failure. The customer stuck on step three needs someone to walk them through step three. And replacing a second visit with an assisted call removes a whole logistics operation, which no follow-up message can do.
Discarded
A follow-up campaign at the end of the funnel, contacting everyone who had gone quiet.
03
Decided: a multichannel model where the system resolves the first level (identify, gather, confirm what can be confirmed) and a human advisor takes the second, where the decision carries consequences. Personalized by type of customer and of loan, on WhatsApp and on calls, and held together by a conversation framework that goes from the logic (states, paths, exits, when a person steps in) down to the dialogue in each channel. The latest flow built on it, reaches customers who were approved and paid out but never used their loan.
Why: in a regulated process, a wrong approval isn’t a bad experience, it’s a compliance failure. But routing everything to a person keeps the exact bottleneck the unit started with. Splitting by consequence lets the volume through without letting the decision through, and never takes the person out of the conversations that need one.
Discarded
Full automation, the cheaper and more fashionable answer.
BEFORE
Every case waits for an advisor
1 · IDENTIFY
2 · GATHER
3 · CONFIRM
4 · DECIDE
QUEUE
Every stalled case, simple or not, waits for a human
ADVISOR
Looks the customer up across other platforms
Asks again for data the customer already gave
Confirms what could have been confirmed automatically
Decides
CUSTOMER
Waits for a call
Repeats what they already gave
Waits again
Gets an answer, or gives up
BOTTLENECK
Advisor time: every case, however simple, needs a person
AFTER
Two levels: the line is drawn by consequence
1 · IDENTIFY
2 · GATHER
3 · CONFIRM
4 · DECIDE
ASSISTANT
Level 1: identify, gather and confirm what can be confirmed
ADVISOR
Not needed yet
Not needed yet: the assistant gathers what’s missing
Opens the case with everything already gathered
Level 2: validates the decision that carries consequences
CUSTOMER
Is identified by the assistant
Sends what’s missing, in the same channel
Confirms
Gets an answer
HUMAN TIME
Only decisions with consequences reach a person
THE FRAMEWORK
From the logic to the dialogue
1 · STATES & PATHS
2 · EXITS & HAND-OFF
3 · TURNS
4 · CHANNELS
LOGIC
One logic for every channel: the paths don’t change, only how they’re said
DEFINES
Where the customer is and what they told us. Each answer opens its own path
Every path ends in an exit. Exits that carry consequences go to a person
One purpose per turn: what’s said, the expected answers, a fallback and where it goes next
The same turn, written for each channel: tappable on WhatsApp, conversational on a call
UNUSED LOAN (Real case using this framework)
Approved and paid out, but unused. Their answer: “I don’t know how to use it”
The assistant walks them through it. If they’re still stuck, an advisor calls
“Which step are you on?” If it’s unclear: “Tell me what you see on your screen”
WhatsApp sends the next step. On the call, the advisor picks up from what was sent
WHEN AN AI, WHEN A HUMAN
A human steps in where a decision has consequences, never because the script ran out
The result
1 · Capacity
2 · One console
3 · Recovery
4 · Multichannel
From four actions to a platform that serves more of the business.
time per call
Measured by the client, last 12 months
minutes to find a customer’s address
Measured by the client
97%
of 51 advisors found the panel’s contact history useful
Usability test, two variants
loans a year from the rescue flows
A model with explicit assumptions, not a measurement
What I learned
That decision is a product decision, not a technical one. Put the human too early and you’ve automated nothing; too late and you’ve automated the one thing you couldn’t afford to get wrong.