Multi-Agent IT Helpdesk
A production multi-agent system on AWS Bedrock serving 1,500 employees. Deflects ~15% of the L1 queue and answers in seconds instead of hours.
Role
Product Owner & AI Lead
Timeline
Live since June 2026 · 8 months to production
The challenge
IT was buried in first-line volume, hardware requests, VPN access, software provisioning, access and permissions, while issues needing real expertise sat behind a week-long average resolution time.
The decision
Chose Human-in-the-loop over Full auto-close
A wrong auto-close costs more trust than the minutes it saves.
How I approached it
- 01Mined ticket history to find the repeatable intents worth automating; wrote the business case and PRD with deflection and resolution time as north-star metrics.
- 02Designed a multi-agent architecture: a parent agent routing via Azure AD to an employee self-service bot and an IT-support bot, each with its own knowledge and permissions.
- 03Built RAG on AWS Bedrock grounded on IT knowledge and Matrix42 ticket data, with PII redaction in the ingestion pipeline.
- 04Treated governance as part of the build rather than a hurdle after it: GDPR, EU AI Act classification and works council alignment ran alongside development, and launch included human handover for unresolved cases and feedback loops to keep improving answer quality.
- 05Ran the build inside the governance checks that later became standard for every AI project in the company: risk classification, works council alignment, go-live checklist.
Impact
- ✦~15% of L1 tickets deflected since go-live in June 2026.
- ✦Tickets that reach IT arrive pre-filled with conversation context.
Stack
- AWS Bedrock
- Claude
- S3 / Lambda
- Strands Agents
- Azure AD
- Python