Appearance
VoC Agent Architecture Patterns
How to structure a Voice of the Customer agent system — from data infrastructure to agent orchestration.
Data Infrastructure First
Before deploying any agents, you need a unified data layer:
- CDP Stack (Ramp's choice): Snowflake + dbt + Hightouch
- Snowflake: warehouse + Cortex AI for natural language queries
- dbt: transforms raw data into queryable models
- Hightouch: reverse ETL pushes insights back into tools (Salesforce, Slack, etc.)
- Unstructured data handling: Gong transcripts, support tickets, emails — all flowing into the warehouse
- PII controls: Exclude SSNs, partial identifiers — build privacy in from day one
- MCP Server: Snowflake now supports Model Context Protocol for direct LLM connectivity
Parallel Agent Architecture
Ramp's core pattern — launch multiple specialized agents simultaneously:
[Coordinator Agent]
|
├── [Gong Agent] ——> scans call transcripts
├── [Zendesk Agent] ——> scans support tickets
├── [Competitor Agent] ——> scrapes competitor sites
├── [Codebase Agent] ——> searches internal code
├── [Salesforce Agent] ——> pulls CRM data
└── [Usage Agent] ——> analyzes product analytics
|
v
[Synthesis Agent] ——> merges artifacts into insights- Each agent writes a markdown artifact with its findings
- A coordinator agent applies map-reduce synthesis — identifies patterns across all artifacts
- 6-10 agents per research task is Ramp's sweet spot
Agent Design Principles
Constrained Decision Spaces
- Learned from Ramp's Tour Guide agent: the most effective improvement is constraining what the agent can do
- Give agents narrow, well-defined tasks rather than broad mandates
- Single-step action generation — agent produces one action at a time based on current state
L0-L3 AI Adoption Framework
| Level | Description | Example |
|---|---|---|
| L0 | No AI | Manual research |
| L1 | AI assists human | Copilot suggests, human decides |
| L2 | AI acts, human reviews | Agent drafts spec, PM edits |
| L3 | Full autonomy | Agent handles end-to-end |
Ramp pushes every function toward L2-L3.
Two-Tier Model Strategy
| Tier | Data | Use Case |
|---|---|---|
| General | Aggregated across 15K+ customers | Benchmarks, trends, category insights |
| Sensitive | Opt-in, in-context only | Customer-specific analysis, no persistent training |
RAG Pipeline for Customer Data
For ingesting Gong transcripts and similar unstructured sources:
- Chunking: Fixed-size 512-1024 tokens (recommended for transcripts)
- Metadata enrichment: Call ID, participants, timestamps, topic labels, sentiment
- Speaker mapping: Map speaker IDs to actual names
- Namespace partitioning: Isolate by tenant/customer for security
- Vector indexing: Embed chunks with metadata for semantic search
- Sync strategy: Historical bulk load + webhook/polling for ongoing updates
Key Takeaways
- Data infrastructure is prerequisite #1 — agents without clean, unified data just hallucinate faster
- Parallel agent execution with map-reduce synthesis is the dominant pattern for multi-source VoC
- Constrain agent decision spaces aggressively — narrow tasks >> broad mandates
- Build two model tiers (general aggregated vs. sensitive opt-in) to handle data privacy
- RAG with 512-1024 token chunks, rich metadata, and namespace isolation is the standard approach for transcript data