2024 • At-Bay • Product Design Lead
AI-driven invoice automation for cybersecurity insurance claims
Role
Led product strategy and design for the invoice automation initiative — mapped the AI extraction pipeline with engineering, defined validation logic and escalation workflows with PM and claims stakeholders, and designed the adjuster experience from ingestion to payment.
Outcome
85% straight-through processing
80% reduction in processing time for reviewed invoices
$130M in payments processed in year one
Invoice volume scaled in step with 113% YoY claim growth — absorbed without adding headcount
Eliminated external payment processing vendor
Vendor payment time: 7+ days → 1–2 days
Team / Stakeholders
Product Manager
Engineering (Tel Aviv)
Claims Team (Operations, Adjusters, Managers)
An AI-powered system that automatically extracts, validates, and processes vendor invoices for claims payments — routing clean invoices to auto-payment and surfacing blocked invoices with clear resolution paths for adjusters.
The Challenge
Adjusters were manually reviewing, entering, and submitting thousands of invoices per day — ~12 minutes each.
When invoices couldn't be processed, they disappeared into inboxes for months with no tracking or accountability.
Key Decisions
Validation pipeline over blind automation
Rather than auto-paying everything the model extracted, we defined conditions with PM and claims stakeholders — coverage position, retention, reserves, threshold — that had to pass before payment. This kept financial safeguards intact while still automating 85% of volume.
Made the queue the system of action
Each blocked invoice shows the exact failure reason and links directly to the resolution page. Resolution speed improved not by notifying louder, but by putting the problem and the fix in the same place.
Phased trust calibration
Launched supervised with human review on every auto-payment, then moved to unsupervised after accuracy stabilized — raising the threshold from $5K to $10K. Hard constraints stayed regardless: unknown vendors, missing claim IDs, and high-dollar invoices always route to humans.
Adjuster’s view - one of three role-based views.
What informed my decisions
Conversations with adjusters revealed that the real problem wasn't processing speed — it was that blocked invoices had no resolution path, leaving financial exposure invisible to the business. I mapped the full system before designing any UI — extraction logic, validation conditions, failure states, human-in-the-loop checkpoints, and escalation paths — which is what surfaced the need for the validation pipeline and the blocked invoice resolution flow.
Results
85% straight-through processing
80% reduction in processing time for reviewed invoices
$130M in payments processed in year one
Invoice volume scaled in step with 113% YoY claim growth — absorbed without adding headcount
Eliminated external payment processing vendor
Vendor payment time: 7+ days → 1–2 days
View other projects
AI-powered transcription for broker meetings
Customer-facing death claims intake & profile experience
