Redesigning identity verification as a trust system

Context & constraints

For years, Remitly’s Tier Limit system used fixed send thresholds to trigger enhanced due diligence. Customers who crossed a threshold were stopped after starting a transfer, asked to submit identity and source-of-funds information, and placed in a review queue that could take 48 hours. Fewer than one in five blocked customers went on to complete a send.

Regulations still required enhanced due diligence. The opportunity was to direct that scrutiny according to risk, rather than applying the same friction to everyone. The Customer Risk Rating model ran on a 24- to 48-hour batch cycle, and Australia was the first market in a staged rollout.

Problem

The legacy experience created too much friction: customers dropped off and never finished the document collection flow, so they never sent money. The end-to-end success rate was 17% (35% opened, 70% completed, 70% approved). Fewer than one in five blocked customers went on to complete a send.

Solution

The redesigned flow moved document collection before the final send step, replaced free-text occupation answers with structured categories, and gave customers a clearer path through verification. The result was a 48-hour to 2-hour average review time and a higher completion rate for customers flagged for enhanced due diligence.

My role

I led the customer-facing design, partnering with a product manager, four engineers, Compliance, and Customer Success. My scope covered the send-flow collection experience and a deeper issue in the challenge questions: more than half of customers were answering the occupation question with free text that could not be classified reliably.

Key decisions & trade-offs

I moved information collection before the final send step, where the transfer details could explain why the information was needed. I designed separate optional and mandatory states: an early request that could be deferred, and a clear completion gate when the transaction depended on it.

I also worked with Compliance to replace free-text occupation responses with structured categories aligned to the risk model. That narrowed the input, but made it possible to evaluate ordinary answers consistently and reserve manual review for cases that warranted it. The design worked within the accepted batch-processing delay rather than implying that risk decisions happened in real time.

Outcomes