← Back to all builds
Production demo

Renewal Risk Intelligence Engine

A three-agent n8n pipeline that ingests an at-risk customer account, retrieves context from a Pinecone vector database, and produces a fully styled HTML case file with risk assessment, intervention plan, and CSM-team recommendation — in under 30 seconds.

Tools
n8n, Claude Opus, Pinecone
Pattern
Multi-agent RAG pipeline
Status
Shipped end-to-end
Loom walkthrough coming soon.

In the meantime, the architecture and case-file output are documented below. To see the build in action live, book a 15-minute walkthrough with Sean.

CS teams spot risk too late.

By the time a CSM realizes a renewal is in trouble, the champion has often already left, the economic buyer has gone dark, and the paid-but-unused modules have quietly become the de facto reason the customer is shopping competitors. Risk signals are everywhere — in support tickets, usage data, billing records, sentiment, who's not showing up to QBRs — but they're scattered across systems and read by no one in aggregate. Most "early warning" tools surface lagging indicators (NPS drops, support ticket spikes) rather than the *contradictions* and *silences* that actually predict churn.

Three agents. One case file.

The pipeline runs in sequence. Each agent has a single job, a single output schema, and a single Pinecone namespace to pull from. The final node assembles every agent's output into a styled HTML case file the CS leader can open in any browser.

1
Webhook Trigger
Accepts {account_name, account_id} via POST.
2
Gather All Intelligence
Code node runs three parallel Pinecone searches across accounts, historical-data, and team namespaces. Returns concatenated text payloads for each downstream agent.
3
Agent 1 — Risk Analyst
Claude Opus
Analyzes signals, surfaces contradictions, flags "silence is a signal" patterns (the key contact who stopped logging in, stopped responding, stopped submitting tickets), assesses economic-buyer status, identifies paid-but-unused modules with ARR at risk. Returns structured JSON.
4
Agent 2 — Intervention Planner
Claude Opus
Receives Agent 1's risk profile plus historical data on similar accounts (saved and churned). Identifies the closest analog, extracts the key lesson, builds a priority-ranked intervention plan with owners and timing, and estimates save probability with vs. without intervention.
5
Agent 3 — Team Matching
Claude Opus
Key innovation: scans every CSM profile for prior relationships with the at-risk contacts. A CSM who previously worked at a company where these contacts were customers is worth more than any skill score. Then recommends primary lead, supporting role, and VP escalation requirements.
6
Build Case File
Code node assembles a styled HTML report — risk score, hidden-relationship alert, situation summary, top risks, contradictions, silence signals, paid-but-unused modules, historical pattern match, intervention plan, team recommendations. Returns to the webhook caller.

What the CS leader sees.

The case file is a single self-contained HTML report. Below: live screenshots of the executed pipeline output.

Screenshots coming soon.

The case file includes a risk score (1-10), urgency tier, hidden-relationship alerts, top risks with cited evidence, detected contradictions (e.g., positive NPS from IT while the economic buyer goes dark), silence signals, unadopted-module ARR at risk, historical pattern match with save-probability estimates, the intervention plan, and the recommended CSM team configuration.

Hidden relationships beat skill scores.

The single most important design decision in this build: Agent 3 is instructed to scan every CSM profile for prior relationships with the at-risk contacts before evaluating any skill or bandwidth score. A CSM who previously worked at a company where the now-unreachable economic buyer was a customer can re-open the door in a way no amount of "skills match" can.

This reflects how senior CS leaders actually triage red accounts in practice — they ask, "who do we know inside?" first, then optimize for skill and bandwidth second. Most automated CSM-matching tools invert this and produce mathematically correct but practically useless assignments.

Book 15 minutes.

The build is a working demo. The fastest way to evaluate it is to watch it execute end-to-end on a real account scenario. Pick a 30-minute Calendly slot — first 15 minutes is the demo, second 15 is whatever you want to discuss about it.