Overview

Bdeo · AI mobile flow

If the user doesn’t trust it, the AI has nothing to analyse

Bdeo’s mobile flow asks policyholders for photos of their car so the AI can assess the damage. Many dropped out halfway and the photos that arrived weren’t always usable. We redesigned it by listening to 100 real drivers.

In 30 seconds

In 30 seconds

Problem

One in three policyholders declined the GPS permission, without which fraud can’t be detected, and almost a third of the photos arrived cropped. Every photo that doesn’t work means another request to the policyholder or an adjuster visit that can take more than ten days.

My role

I co-designed the flow with another designer: research, concept proposals, final design and defining what to measure in each phase.

Result

−49% cropped photos and −33% GPS denials, measured in production. Cases with at least one cropped photo fell from 65% to 35.5%.

Role

Senior Product Designer

Duration

3 weeks of research and design, plus phased measurement in production

Team

1 PM · 2 Product Designers

Context

Mobile flow where the policyholder photographs their vehicle after a claim. Bdeo’s AI analyses the images and the adjuster gets the result in their dashboard. The MVP had been built by an external team with no client research.

Impact

• −33% GPS permissions denied — From 33.73% to 22.46% of users declining GPS.
• −49% Cropped photos — From 30.83% to 15.82%. A cropped photo is one the AI can’t process.
• 65% → 35.5% Cases with at least one cropped photo — Before, more than half of the cases arrived with at least one photo the AI couldn’t use. Now, one in three.

The problem

Without Bdeo, a car claim needs an adjuster to go and see the vehicle. The appointment depends on their schedule and the policyholder’s, and it can take more than ten days. With Bdeo, the policyholder gets a link, takes a few guided photos in about five minutes, and the AI tells them whether the damage is covered.

The product already worked and had clients, but it had been built by an external team without talking to users. It showed in two numbers. One in three users declined the GPS permission, and without location the insurer can’t detect fraud. And almost a third of the photos arrived cropped.

Research

We asked drivers who had never used Bdeo to complete the existing flow. We recorded their screens, counted clicks, timed them and, at the end, asked about the experience. We looked for a mix of ages, genders, backgrounds and professions.

• 40 drivers tested the existing flow
• 60 new drivers tested the redesign
• 3 measurements per session: screen recording, clicks and time

1. Mistrust made people drop out. The flow asked for camera and location permissions halfway through and without explaining what for. Many people declined them.
2. They didn’t know how to take the photos. Nothing explained how far away to stand or from which angle.
3. They cropped the photos. There was no guide mark in the camera to frame the car.
4. Nobody read the help messages in the camera.
5. They needed examples. Seeing a good photo explained more than any text.

Decisions

Three changes, each tied to a research finding.

Ask for every permission up front, and explain it

The flow asked for camera and location permissions halfway through, when the user was already several screens in. We grouped them into a single moment at the start, with screens explaining what each one was for and a clear illustration. The friction was no longer spread across the flow; it sat at a single point the user understood.

Result: Users declining GPS dropped from 33.73% to 22.46%.

A car silhouette in the camera

Users didn’t know how far away or in what position to take the photo. We added templates with the vehicle’s silhouette over the camera so they could fit the car inside it. It was the first round of improvements and the one that moved the numbers most.

Result: Cropped photos: from 30.83% to 17.50%. Cases with at least one cropped photo: from 65% to 42%.

Help right where they hesitated

The general help messages went unread. In the recordings we saw which screens people stalled on and put short help there, with an example photo: don’t get too close, avoid reflections, make sure the damaged part is visible.

Result: Cropped photos: from 17.50% to 15.82%. Cases with at least one cropped photo: from 42% to 35.5%.

My product role

I co-designed the flow with another designer, and we worked together on every phase: the research, the concept proposals and the final design. With product and the AI team we decided to release the changes in phases, and I defined what to measure in each one. That’s why we know which change moved each number, instead of having a single before and after.

Validation

With the changes in a prototype, we repeated the test with sixty new drivers, with the same mix of profiles, and grouped the results. We then measured cases created versus completed in production, before and after the redesign. Completed cases also went up: grouping the permissions at the start didn’t make more people drop out.

1. Test of the existing flow with 40 drivers
2. Three changes tied to findings
3. Test of the redesign with 60 new drivers
4. Production measurement of cases created versus completed

Designing for AI

1. In a flow that asks for photos and permissions, trust comes before usability. If users don’t understand why you’re asking for something, they never reach the screen you designed.
2. A visual example solves what text can’t. The car silhouette did more than all the help messages put together.
3. The AI is only as good as the images it receives. Improving the capture improved the model without touching the model.

What I’d do differently

I’d test next to the car. We validated the redesign with 60 drivers by recording their screens, but not in the real situation. Today I’d also run tests with policyholders next to their car, in poor light, in a hurry and with the phone in one hand, which is where a photo fails.

And I wouldn’t consider the flow finished. After the in-context help, almost 16% of the photos still came out cropped. I’d run a second round of research with just those cases to understand what they’re missing.

Closing

For the AI to work, it first has to be understood by whoever takes the photo.