How to build an AI agent that monitors customer signals and scores churn risk so customer success can prioritize outreach to prevent subscription cancellations.
Rise and shine — it’s Monday morning.You’re a data analyst at Bluepine, a subscription-based learning platform with a few thousand customers. Most weeks you write SQL queries and build dashboards. When somebody asks a question, you run a query and return a table or a chart.If someone asks what plans customers are on, you run a simple query:
SELECT plan, COUNT(*)FROM customersGROUP BY plan;
Your job is to find the right data and make sense of it for the business.Then this email lands.It’s from Claire, who runs the customer success team. The subject line reads, “We lost Northgate.”Now read the email together.Northgate canceled Friday afternoon. Nobody on my team saw it coming. I have 12 renewals this quarter, and I can’t tell which one of them is going to be next. I need to know before they tell me.That last sentence is the actual requirement. Claire isn’t asking how many customers we lost already or what our churn rate was last quarter — she’s asking about the customers who haven’t left yet.Bluepine makes money one subscription at a time. Customers pay every month, and as long as they stay, revenue repeats. When Northgate canceled, that repeating stopped. This is churn.
Churn compounds. Losing 2% of customers every month sounds small, but 12 months later more than a fifth of the base can be gone. At that point, every new sale is just refilling a leaking bucket, not genuine growth.
Refilling is expensive: marketing spend, sales calls, demos, onboarding — often several times the cost of keeping an existing customer. Retention often comes down to simpler, cheaper interventions like a timely phone call from customer success.Bluepine’s customer success team exists to prevent churn. Their quarter is judged by one number: how many customers stayed.So your job is to stop cancellations before they happen. The catch is customers rarely say, “I’m leaving.” They fade away: logins drop, support tickets increase, engagement declines.
By the time the cancellation email arrives, there is often nothing left to save. Bluepine never missed an announcement and nobody ignored a warning — there simply wasn’t a clear warning surfaced to the team. The signals were in the data the whole time: weeks, months, or even years.For example, consider an account — Bluepine with customer ID C002. It has one login in the last 30 days and six support tickets. No one noticed because catching it requires cross-referencing login counts against ticket queues for every account, every day.So where do we start?You do the obvious thing first: open the dashboard. It isn’t helpful — it counts customers who already left, but Claire asked for the ones who haven’t left yet.
Answering Claire’s question means digging into the raw signals and labeling accounts by risk. Useful signals include:
Signal category
What to measure
Example
Recent usage
Login count, feature usage over window
usage_30d: 1
Tenure
How long they’ve been a customer
tenure_years: 2.1
Support activity
Open tickets, ticket volume and type
tickets_open: 6
Billing
Failed payments, disputes, invoicing delays
failed_payments_30d: 2
Seat / product changes
Drop in seat counts or decreased feature adoption
seats_active_change: -30%
A simplified per-account snapshot you might examine:
If an account looks high risk, someone on the customer success team needs to call — that requires human judgment to prioritize and explain why.This is exactly the kind of continuous monitoring and prioritization you can offload to an AI agent: one that watches signals, scores risk, and highlights accounts for outreach. Bluepine doesn’t have that agent yet.Building it breaks down into three parts:
Build the agent itself and understand what it’s made of.
And that last task — reaching the data — often contains a surprise: the data you need usually lives in a database, data warehouse, or a set of tools that require connectors, permissions, and careful data modeling before an agent can reliably act on it.
Start by defining the risk signals you care about and where each lives (auth logs, support system, billing, product telemetry). This makes it easier to map the data sources and prioritize integrations.
