Assisted Support for Human-Led Conversations
An AI agent designed to scale real-time support without losing human touch
Empower Work is a U.S.-based non-profit organization providing confidential, real-time chat support to individuals facing challenges at work. Their support model is built around trained human volunteers who deliver empathetic, one-on-one guidance in sensitive situations where trust, tone, and human judgment are critical.
Visit website ↗Scalable support without compromising empathy.
GreenM designed and implemented an AI-assisted support workflow embedded directly into Empower Work's existing live chat environment. The solution provides real-time guidance during conversations, helping volunteers respond with confidence and consistency.
Demand doubled without room to scale
In 2024, demand for Empower Work's text-based support doubled and continued to rise into 2025. As volume grew, operational constraints began to limit capacity, consistency, and volunteer confidence during real-time conversations. Scaling through traditional means would have required lowering quality or significantly increasing operational overhead — both incompatible with the organization's mission.
- Capacity ceiling — volunteers could manage ~2 simultaneous chats
- Long, high-focus sessions — average conversation ~50 minutes
- Strict time pressure — expected reply time under 3 minutes
- "Stuck" moments — uncertainty in responses reduced confidence
- Manual workflows — summaries, handoffs, and resource lookup were manual
- Limited scalability — growing demand translated into higher workload
GreenM embedded contextual intelligence directly into the live chat environment. During conversations, the system provides contextual, editable suggestions to support responses, while maintaining full human control and enabling fast access to relevant resources and internal knowledge.
- Contextual, editable guidance in real time
- Human-first response support
- Fast access to internal knowledge
- Privacy-aware processing
- Seamless workflow integration
The AI-assisted workflow delivered measurable improvements across volunteer efficiency, support quality, and operational scalability.
Increased Capacity
Volunteers managing 3+ simultaneous conversations doubled (27% → 53%).
Higher Confidence and Quality
93% rated the assistant as helpful, 70% as very to extremely helpful.
Faster Resource Access
41% reduction in time spent sharing resources during conversations.
Reduced Admin Effort
Post-session summaries completed up to 60% faster.
Strong Adoption
Actively used by ~65% of volunteers during live shifts.
"It's like having an extra set of eyes watching the road, yet I am still driving." — Cyn · Peer Counselor, Empower Work
DELIVERY
The solution was developed in close collaboration between GreenM and Empower Work to bring a production-ready AI-assisted workflow into live operations quickly, while enabling continuous improvement based on real usage.
Phase 1
Human-Centered Workflow Design
Design of a human-in-the-loop AI workflow defining where AI assists and where human judgment remains final in sensitive, live conversations.
Phase 2
Retrieval-Augmented Implementation
Embedded AI assistance built using retrieval-augmented generation (RAG) to surface relevant context from vetted internal resources and prior conversations, without model fine-tuning.
Phase 3
Human-in-the-Loop Production Release
Production-ready AI-assisted workflow deployed for real-time conversations, with all suggestions optional, editable, and fully controlled by volunteers.
Phase 4
Feedback-Driven Continuous Improvement
Continuous refinement of prompts, retrieval logic, and user experience based on volunteer feedback and real production usage.
- LangChain
- LangSmith
- AWS Lambda
- Node.js
- Amazon S3
- Amazon RDS (PostgreSQL)
- Amazon CloudWatch
- PagerDuty
- PostHog
AI assistance was introduced in a way that preserves confidentiality, human oversight, and responsible use in sensitive, real-time support environments.
Privacy-First Processing
Conversation data is anonymized and handled in a privacy-aware manner.
Human-Governed Interactions
AI suggestions remain optional and editable, with full human control at every step.
Controlled Usage
The solution avoids unnecessary data duplication or exposure beyond live support operations.
Designed for Sensitive Contexts
Built for complex conversations where trust, discretion, and accountability are critical.
