Cutting broker processing time with wefox's first AI feature

At a glance

The bottleneck nobody wanted to talk about

It was switch season at wefox, the stretch of the year when brokers earn their keep. Every broker I shadowed had the same setup: two monitors, a stack of old insurer documents, a half-finished coffee. Flip a page, type a field. Flip, type. For hours.

Each contract started the same way. The customer handed over their previous insurer's document, and the broker copied the fields into our switch form by hand. Name. Address. Date of birth. Policy details. Again and again. It was the highest-volume task in their workflow, and the slowest. The more brokers we hired, the more we paid for the same repetitive work.

Could AI fix this?

What I actually heard in interviews

Before I sketched a single screen, I sat down with brokers across regions. I expected complaints about typing. I got something more interesting:

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

That line, in different forms, came up in nearly every interview. Brokers didn't just want speed. They wanted certainty, and they trusted their own fingers more than they trusted anyone, or anything, else.

Three findings shaped everything that followed:

Brokers re-checked every field they didn't type themselves. So automation only saves time if review is faster than typing.

Document quality varied wildly. Clean PDFs, blurry phone photos, smudged scans, all in the same week.

The brokers most resistant to OCR were the top performers. They had muscle memory for the old form. To them, any change was friction.

That last finding rewrote the brief. The risk wasn't that OCR would fail technically. It was that we'd ship a working feature and nobody would use it.

The decision: phase trust, don't force it

I pitched a three-phase rollout that separated the moment we launched the technology from the moment we changed broker behavior.

Phase 1, dual entry. Brokers chose between OCR and manual filling. Every correction they made quietly became training data.

Phase 2, AI training. OCR became the recommended path while manual stayed available. We measured confidence per field and let brokers correct errors.

Phase 3, full adoption. OCR became the default. Not because we forced it, but because brokers had already chosen it.

This was the most important decision of the project. It traded short-term adoption metrics for long-term trust, and it gave the ML team a steady stream of broker-validated data to retrain on.

When the first prototype flopped

We built a prototype, and I sat down with three in-house brokers to test it. It went badly.

The OCR took over three minutes to populate the form. Long enough that one broker shrugged, opened the form in another tab, and started typing manually while we waited. And when the form finally filled in, review felt heavier than typing. Brokers scanned every field top to bottom looking for errors, and by the time they finished, they could've keyed the form from scratch.

The speed problem went to engineering, where I worked with the ML team to bring processing time down to a threshold brokers would actually wait for. The review problem was mine.

Redesigning the review experience

I rebuilt the review screen around three principles:

Confidence-weighted review. Fields the model was sure about got visually quieter. Uncertain fields got attention. Brokers stopped scanning everything and started scanning what mattered.

Inline correction. Any field could be edited in place. No modals, no leaving the review screen.

Source document side by side. The original document stayed visible next to the extracted fields, so brokers could verify without switching context.

The next test went differently. The same brokers who'd given up on the first prototype were finishing review in under a minute.

Outcomes

• Field validation went from over 3 minutes to under 1 minute between the first and second prototype, hitting the success metric I defined upfront.
• Switch-form completion dropped from 3 to 4 minutes to 1 to 2 minutes per contract. Manually, brokers went back and forth between documents for every field. With OCR, they checked the filled form once and saved, even while the model was still being trained to read documents from different insurers.
• OCR became the default form-filling method once accuracy crossed broker-approved thresholds. Brokers chose it. Nobody forced them.
• The three-phase rollout became wefox's template for shipping AI features, reused by other teams after this project ended.

What I took from it

Trust is a design material. The dual-entry phase looked like a compromise on paper: two flows to maintain, slower headline adoption. In practice, it was the reason adoption stuck. Brokers chose OCR because they'd been allowed to verify it on their own terms.

Performance is a UX feature. A slow correct answer loses to a fast manual one, every time. Latency wasn't an engineering afterthought. It was a design constraint, and I treated it like one.

The interface is just the surface. Brokers weren't worried about AI being wrong. They were worried about being held accountable for AI's mistakes. The real design work happened underneath, in how we structured trust, control, and accountability.