How a fintech startup used on-device LLMs to cut onboarding time by 40%

AI · 5 min read

The startup, Vaultly (fictional), faced high drop-off during identity and income-verification steps. Their original flow relied on server-side OCR and rule engines that added a 6–10 second round-trip per document, plus a confusing set of form fields for verification. The product team hypothesized that reducing perceived friction and offering immediate, context-aware help would increase completion rates.

They prototyped an on-device lightweight LLM for form summarization and guided corrections. Instead of returning binary validation results, the model provided a concise, human-readable hint for each failed capture (e.g., “rotate the document slightly and increase lighting”). This moved much of the feedback loop out of the seven-second server wait and into sub-second local responses.

After an A/B test with 12,000 users over four weeks, onboarding completion rose 40%, average time-to-complete dropped from 7.2 minutes to 4.3 minutes, and help-center contact rates fell 28%. The team documented trade-offs: increased APK size, the need for differential update plumbing, and a stricter vetting process to ensure the model didn’t hallucinate compliance advice. The case shows how careful on-device AI integration can shrink friction points without sacrificing privacy or auditability.