A health score that lives in a quarterly spreadsheet isn't a health score — it's a retrospective. The agent computes health scores continuously from six signal categories: product engagement (login frequency, feature depth, session length), value realization (are they using the capabilities tied to their stated goals?), support activity (open tickets, CSAT, escalation history), commercial signals (overdue invoices, downgrade requests, contract change history), relationship signals (CSM touchpoint recency, executive engagement), and external signals (leadership changes, company contraction). Each category is weighted based on its predictive correlation with churn for that customer segment. The score updates daily and triggers alerts at defined thresholds.

Health scores measure current state. Early warning detection looks for patterns that predict future decline — before the score itself drops. The agent monitors for a specific set of behavioral patterns correlated with churn in the 60–90 days prior: declining login frequency over a 3-week trend (not just a single week), support ticket themes shifting from 'how do I' to 'this doesn't work,' champion engagement dropping off while power user engagement holds steady (signals internal champion may be disengaging), and contract data showing no expansion in 12+ months in a high-growth account cohort. When a pattern matches, the early warning fires independently of the health score — so the CSM knows something is changing before the number drops.

When a health score drops below threshold or an early warning fires, the agent generates a specific intervention playbook for the CSM — not a generic 'reach out and check in' prompt, but a structured plan tailored to the risk pattern. A champion disengagement playbook looks different from a product adoption failure playbook, which looks different from a competitive threat playbook. Each includes: the specific signals that triggered it, a recommended outreach message calibrated to the customer's communication history, a suggested agenda for the conversation, a list of internal resources (relevant training, product features, success stories) to address the underlying issue, and an escalation trigger if the first touchpoint doesn't get a response in 5 days.

Individual account interventions are only half the picture. The agent maintains a portfolio-level churn risk view that shows the distribution of account health across the book, the total ARR at risk in each tier, and the CSM workload required to address it. When a significant portion of ARR is concentrated in the at-risk tier, leadership can make resourcing decisions before accounts start canceling. The view also tracks intervention outcomes: which playbooks are working, which CSMs have the best save rates, and which risk patterns are hardest to reverse. This feeds back into the scoring model to improve early detection over time.