Adoption at scale
Rolled out beyond a single team to tens of thousands of active users across business units.
Insurance/5 screens
Insurance AI Platform
Scaled an internal claims and policy assistant to 20,000+ users with monitored, auditable workflows.
Screen walkthrough
A fuller gallery for the product story.
This gallery is meant to show progression, not just a single hero frame. Use it to talk through navigation depth, records, analytics, and workflow context during a call.
/
Chapter 01
Brand-aligned visual — client interfaces are anonymized for this launch story.
Overview
A tier-one insurance group needed production-grade AI on real claims and policy data — not a pilot chatbot. We embedded with underwriting and claims teams, shipped retrieval-first assistants with human escalation, and operated the stack with observability from day one.
Strongest story angle
Use this story when buyers ask whether AI can survive messy PDFs, legacy policy wording, and compliance review — with proof of adoption, not slide decks.
Observable modules
Document intakePolicy retrievalAdjuster routingAudit loggingHuman escalation
Client story
Problem, approach, and outcomes
Context
A large insurance group in Southern Africa needed AI that could run on real claims volumes, integrate with existing core systems, and pass internal risk review.
Problem
Claims and policy questions were trapped in email threads and manual lookups. Customers waited on hold while specialists searched PDFs. Pilots had failed because answers were not grounded or auditable.
Approach
We started with the three highest-volume inquiry types, built retrieval-first assistants on approved document corpora, and wired human escalation into every workflow. We shipped incrementally behind feature flags with full request logging.
Architecture
API gateway → workflow orchestration (LangGraph-style state machines) → vector + keyword retrieval on approved stores → LLM with strict citation prompts → event bus for audit and CRM updates. Deployed in the client’s environment with secrets in their vault.
Tech stack
PythonFastAPIPostgreSQLLangChainLangGraphOpenAI + AnthropicAWSDocker
Results
- 20,000+ internal and partner users on the shipped assistants
- Claims cycle time reduced from multi-week queues to roughly three days on targeted lines
- First-response latency for Tier-1 policy questions under ten seconds with source links
Timeline
Phase 1–2 discovery and safety review: 4 weeks. MVP to first business unit: 10 weeks. Multi-unit rollout: ongoing with weekly releases.
“Finally something our compliance team could trace end-to-end — not another black-box demo.”
Why this one works
Three angles worth carrying into the final write-up.
Measured cycle-time impact
Targeted the highest-friction claims paths first so leadership could see end-to-end time drop.
Compliance-first instrumentation
Every recommendation carries citations, model version metadata, and retention controls aligned to the insurer’s policies.
Motion outline
This sequence can still become a short teaser.
- 01
Open on the anonymized insurer context and production constraint.
- 02
Walk the intake → triage → adjuster path with citations visible.
- 03
Close on adoption and monitoring as the definition of shipped AI.
Next publishing pass
The structure is now cleaner: better screenshots, stronger conversion paths, and shared page chrome that behaves correctly. The next layer is adding repository-backed build notes and verified outcome data.
Still worth adding
- 1. Verified repository context for stack and architecture notes.
- 2. Approved proof points to replace generic performance language.
- 3. Short teaser renders once the repository evidence is in place.