Separate the ambition from the first delivery
I mapped the wider email-to-case process and a bounded initial MVP.
The proposal made the intended destination explicit without treating every stage as delivered.
AI consulting · Process design · MVP delivery
Customer service AI automation
Turn an automation ambition into a reviewable first step.
A financial institution wanted to turn incoming correspondence into actionable cases. I defined the workflow and delivered an MVP.
- Contribution
- Architecture & MVP development
- Period
- April–May 2026
- Status
- MVP delivered
- Delivered
- MVP and solution architecture
- April–May 2026
- project period
- Pre-production
- end-to-end impact not established
Customer service AI automationMVP delivered
Incoming correspondence
AI-assisted processing
Case preparation & human review
The challenge
The target workflow connected incoming credit-business correspondence to topic identification, extraction, sender matching and case preparation.
Automating individual steps would not be enough. Case handlers needed a coherent process, with clear human handoffs and traceable decisions.
My role
I developed the solution architecture and stakeholder materials, then built and delivered an MVP using LangGraph, Make and Supabase.
The decisions
Design the human handoffs
Verification, human involvement and auditability were requirements of the target workflow.
These requirements shaped the proposed handoffs between automation and case handlers; the case study does not claim all stages were implemented in the MVP.
The result
I delivered the MVP and solution architecture for review. The wider email-to-case workflow remained the target; this was a pre-production delivery.
What I’d measure next
- Measure
- Correctly prepared cases, handling time and human corrections.
- How
- Run a de-identified sample of correspondence through the existing process and MVP. Compare required-field accuracy, median handling time, correction rate and unresolved cases. Include missing attachments and ambiguous sender details.
- Check alongside it
- Review critical errors separately and verify where a person must intervene before assessing automation coverage.