wefox · B2B · 2021–2022

Getting brokers to trust the machine

During switch season, brokers retyped policy data by hand for hours a day. I designed the OCR-assisted flow that replaced the retyping, and the three-phase rollout that got brokers to actually adopt it. Task completion time dropped 22%, and other Wefox teams later reused the rollout for their own AI features.

Role

Product Designer

Scope

Research · personas · flows · UI/UX · prototyping

My ownership

Research · personas · review-flow design · rollout framework · prototyping

Team

ML team · engineering · product (confirm counts)

Domain

Insurance · broker tooling

Tools

Figma · Miro

Timeline

2021 – 2022

The problem

Every switch season, brokers moved client policies between insurers by hand. Two monitors, the old policy document on one side, the switch form on the other, and hours of retyping in between. In a regulated industry, a mistyped field isn't a typo, it's a liability.

OCR could read the documents. That part was solvable. The real problem was that brokers didn't trust machine-extracted data with their clients' policies, and honestly, they had a point: early accuracy wasn't perfect, and a wrong value that slips through quietly is worse than an empty field.

The speed problem went to engineering. The review problem was mine.

Designing for trust, not just speed

An OCR flow that brokers don't trust is slower than no OCR at all, because they re-check everything anyway. My first prototype proved it: reviewing the extracted data took over three minutes per form, worse than some brokers' manual muscle memory, because the UI treated every field as equally suspect, and so did the brokers.

Watching them scan every field top to bottom made the design question obvious: the review UI had to tell them where to look. So I built the review around the model's confidence. Fields the model was sure about receded. Fields it was unsure about stepped forward and asked for attention. Review turned into triage.

I'd set the target at the start: field review and validation in under a minute. The second iteration got there.

Confidence-weighted fields beside the source document: the review redesign that got brokers finishing in under a minute.

The rollout: earning the default

I pitched a rollout in three phases, built on one rule: brokers choose OCR, we don't impose it. It traded short-term adoption metrics for long-term trust, and gave the ML team a steady stream of broker-validated data to retrain on. That framework outlived the project: other teams at Wefox later reused it to launch their own AI features.

The three phases

Phase 1: OCR as a suggestionExtracted data appeared next to the manual form. Brokers could use it, correct it, or ignore it completely. Every correction went back to the ML team as validated training data, so even the skeptics improved the model with each fix.

Phase 2: OCR as the starting pointOnce accuracy crossed thresholds the brokers themselves had helped set, the extracted data pre-filled the form, with the confidence-weighted review guiding their attention.

Phase 3: OCR as the defaultManual entry stayed available. By this point, most brokers weren't using it.

Before and after: the user's journey

The broker's switch season, before and after this work.

Receive the document

Before

Customer hands over the old insurer's paperwork. Clean PDF, blurry phone photo, or smudged scan, all in the same week.

Routine, but a quiet dread: this is where the typing starts.

After

Broker uploads the document, in whatever quality it arrives.

One action, done.

Enter the data

Before

Two monitors. Flip a page, type a field. Name, address, date of birth, policy details. For hours.

Tedious but safe. "I trust my own fingers."

After

OCR extracts the data and prefills the form in seconds, fast enough that brokers wait instead of giving up.

Strange at first. Faster, but is it right?

Verify

Before

The broker re-checks what they typed as they go. Entry and verification are the same act, which is why it feels trustworthy.

Confident. Nothing on this form got there without their hands.

After

Confidence-weighted review: certain fields go visually quiet, uncertain ones ask for attention. Source document sits beside the form. Any field edits in place.

Focused. They scan what matters, and finish in under a minute.

Submit

Before

Form completed after a long manual slog. The highest-volume task in the workflow was also the slowest.

Relief, briefly. The next document is already waiting.

After

Form submitted after a sub-minute review. Every correction quietly improves the model.

Momentum. The stack shrinks noticeably faster.

Adopt (or not)

Before

New tools were viewed with suspicion, and top performers had muscle memory invested in the old way.

Skeptical by default. "I'd rather type it myself than check someone else's work."

After

Three-phase rollout: dual entry, then recommended, then default. Brokers moved when the accuracy earned it.

Ownership. OCR became the default because brokers chose it.

What actually changed

The work didn't just get faster, the trust model changed. Before, trust lived in the broker's own fingers. After, the interface earned that trust field by field, and the broker stayed in control of when to hand it over.

Outcomes

✓Cut switch-form completion time 22% against the manual baseline.

✓Field review and validation came down to under a minute, the success metric I set at the start, hit in the second design iteration.

✓OCR became the default because brokers switched on their own once the accuracy earned it. Nobody was forced.

✓Other teams at Wefox adopted the phased rollout to launch their own AI features, the outcome I didn't plan for and am most pleased about.

✓Broker corrections fed straight back to the ML team as validated training data, without a separate labeling effort.

Reflection

The design work here wasn't making the machine faster. It was making the humans confident enough to let the machine help. Give people control at the exact moments they doubt the system, and let them decide when to hand it over. They will, on their own schedule, and it sticks better that way.

More work

Redesigning METYCLE around its traders

Metycle · B2B ↗

Cutting heycar's search from five fields to three

heycar · B2C ↗

Let's build something ✷

Have a project in mind? I'd love to hear and connect with you. Let's get in touch and discuss how we can bring your ideas to life. Email me at nr9473@gmail.com

in · LinkedIn ↗Résumé ↓

← Back to the board