Account Churn Prediction
“Which accounts will not renew?”
Book a demo and get a free trial of the full platform: research agent, fine-tune capabilities, and forward-deployed engineer support.
By submitting, you accept the Terms and Privacy Policy.
Loved by data scientists, ML engineers & CXOs at
A real-world example
Which accounts will not renew?
B2B SaaS companies lose 5-15% of ARR annually to churn. For a $200M ARR company, each percentage point of churn reduction is worth $2M. CSM teams manage 50-80 accounts each and cannot manually monitor usage patterns across all of them. The churn signal is rarely in a single metric: it is in the combination of declining user engagement, increasing ticket volume, champion departure, and contract timing that unfolds over 60-90 days before renewal.
Quick answer
The most accurate way to predict B2B SaaS churn is to connect account usage patterns, support ticket escalations, champion activity, and contract data in a relational model. The earliest signal is typically champion departure: when your primary power user goes inactive, churn probability increases 8x. On the RelBench benchmark, relational models score 76.71 vs 62.44 for single-table baselines.
Approaches compared
4 ways to solve this problem
1. Health score spreadsheets
CSMs assign manual health scores based on qualitative assessments of relationship strength, product usage, and support sentiment.
Best for
Captures qualitative signals (executive sponsor tone, competitive mentions in calls) that quantitative models miss.
Watch out for
Does not scale. CSMs managing 50-80 accounts cannot monitor usage patterns across all of them. Scores are updated monthly or quarterly, missing rapid deterioration.
2. Product usage thresholds
Flag accounts where DAU/MAU ratio, feature adoption, or login frequency drops below defined thresholds.
Best for
Catches the most obvious usage decline signals. Easy to automate and integrate into CSM workflows.
Watch out for
Usage thresholds miss context. An account with stable usage but an inactive champion and an approaching renewal with auto-renew disabled is at high risk. Thresholds do not see contract or people signals.
3. XGBoost on account-level features
Build a feature table with usage metrics, ticket counts, NPS scores, and contract details, then train a classifier.
Best for
Good baseline when you have a data team that can engineer features from product telemetry and CRM data.
Watch out for
Treats each account independently. Misses cross-account competitive signals: when multiple accounts in the same vertical start reducing usage simultaneously, it signals a competitive threat that no single account's data reveals.
4. KumoRFM (relational graph ML)
Connect accounts, users, feature usage, support tickets, and contracts into a relational graph. The GNN learns champion departure patterns, cross-account competitive signals, and multi-signal deterioration sequences.
Best for
Highest accuracy with 90-day advance warning. Captures champion departure, usage decay trajectories, and cross-account vertical trends that flat models miss.
Watch out for
Requires user-level activity data linked to accounts. If your product only tracks account-level aggregates, the champion-departure signal is unavailable.
Key metric: RelBench benchmark: relational models score 76.71 vs 62.44 for single-table baselines on account churn prediction tasks.
Why relational data changes the answer
B2B churn unfolds across multiple tables over 60-90 days. The signals include user activity (the champion going inactive), feature usage (declining engagement with core features), support tickets (escalating priority and unresolved issues), and contract terms (auto-renew disabled, approaching renewal date). A flat feature table with 'dashboard_views_30d = 45' hides that this number dropped 68% from last month, that the champion has not logged in for 35 days, and that two similar accounts in the same vertical already churned.
Relational models connect users to accounts to tickets to contracts and learn temporal sequences like 'champion went inactive, then dashboard usage dropped 68%, then a P1 ticket went unresolved for a week, then auto-renew was disabled.' On the RelBench benchmark, this multi-table approach scores 76.71 vs 62.44 for single-table baselines. For a $200M ARR company, that gap is the difference between saving an account 90 days before renewal (when intervention works) and discovering the risk when the non-renewal notice arrives.
Monitoring account health from a usage dashboard is like monitoring patient health from a heart rate monitor alone. The heart rate looks normal, but the patient stopped taking their medication (champion departure), started complaining of new symptoms (escalating tickets), and canceled their next three appointments (auto-renew disabled). The full medical chart tells a story the single vital sign cannot.
How KumoRFM solves this
Graph-learned product intelligence across your entire account base
Kumo connects accounts, users, feature usage, support tickets, and contracts into a relational graph. It learns that accounts where the primary champion has gone inactive, where 3+ users switched to a competitor's integration, and where ticket escalation rate doubled in the last 30 days churn at 8x the base rate. The model captures cross-account patterns: when accounts in the same industry vertical start reducing usage simultaneously, it signals a competitive threat that account-level models miss.
From data to predictions
See the full pipeline in action
Connect your tables, write a PQL query, and get predictions with built-in explainability — all in minutes, not months.
Your data
The relational tables Kumo learns from
ACCOUNTS
| account_id | name | arr | contract_end | plan_tier |
|---|---|---|---|---|
| ACC001 | Acme Corp | $120,000 | 2025-06-30 | Enterprise |
| ACC002 | TechStart Inc | $24,000 | 2025-04-15 | Growth |
| ACC003 | Global Mfg | $360,000 | 2025-09-01 | Enterprise |
USERS
| user_id | account_id | role | last_active | is_champion |
|---|---|---|---|---|
| U001 | ACC001 | Admin | 2025-03-02 | Y |
| U002 | ACC001 | Viewer | 2025-02-10 | N |
| U003 | ACC002 | Admin | 2025-01-28 | Y |
FEATURE_USAGE
| usage_id | account_id | feature | events_30d | trend |
|---|---|---|---|---|
| FU01 | ACC001 | Dashboard | 1240 | Stable |
| FU02 | ACC001 | API calls | 8500 | +12% |
| FU03 | ACC002 | Dashboard | 45 | -68% |
TICKETS
| ticket_id | account_id | priority | category | created_date |
|---|---|---|---|---|
| TK01 | ACC002 | P1 | Bug report | 2025-02-25 |
| TK02 | ACC002 | P2 | Feature request | 2025-02-28 |
| TK03 | ACC001 | P3 | How-to question | 2025-03-01 |
CONTRACTS
| contract_id | account_id | start_date | end_date | auto_renew |
|---|---|---|---|---|
| CON01 | ACC001 | 2024-07-01 | 2025-06-30 | Y |
| CON02 | ACC002 | 2024-04-15 | 2025-04-15 | N |
| CON03 | ACC003 | 2024-09-01 | 2025-09-01 | Y |
Write your PQL query
Describe what to predict in 2–3 lines — Kumo handles the rest
PREDICT BOOL(CONTRACTS.RENEWED = 'N', 0, 90, days) FOR EACH ACCOUNTS.ACCOUNT_ID WHERE CONTRACTS.END_DATE <= '2025-09-01'
Prediction output
Every entity gets a score, updated continuously
| ACCOUNT_ID | ARR | RENEWAL_DATE | CHURN_PROB |
|---|---|---|---|
| ACC001 | $120,000 | 2025-06-30 | 0.12 |
| ACC002 | $24,000 | 2025-04-15 | 0.84 |
| ACC003 | $360,000 | 2025-09-01 | 0.06 |
Understand why
Every prediction includes feature attributions — no black boxes
Account ACC002 -- TechStart Inc, $24K ARR
Predicted: 84% churn probability at renewal
Top contributing features
Champion last active
35 days ago
30% attribution
Dashboard usage trend (30d)
-68% decline
24% attribution
P1 tickets (last 30d)
1 unresolved
19% attribution
Auto-renew status
Disabled
15% attribution
Peer accounts in vertical churning
2 of 5 similar
12% attribution
Feature attributions are computed automatically for every prediction. No separate tooling required. Learn more about Kumo explainability
PQL Documentation
Learn the Predictive Query Language — SQL-like syntax for defining any prediction task in 2–3 lines.
Python SDK
Integrate Kumo predictions into your pipelines. Train, evaluate, and deploy models programmatically.
Explainability Docs
Understand feature attributions, model evaluation metrics, and how to build trust with stakeholders.
Frequently asked questions
Common questions about account churn prediction
What is the best ML model for SaaS churn prediction?
Graph neural networks on relational data (users, feature usage, tickets, contracts) outperform flat-table models. On the RelBench benchmark, relational models score 76.71 vs 62.44 for single-table baselines. The critical signal is the connected deterioration sequence across tables, not any single metric.
How early can you predict B2B SaaS churn?
With relational ML, reliable churn signals appear 90 days before renewal. The earliest indicator is champion departure: when the primary power user goes inactive. Traditional health scores based on CSM judgment detect risk 30 days out at best, often after the decision to leave has already been made.
What is a champion departure signal?
Champion departure means your primary power user or executive sponsor has stopped engaging with the product. This is the single strongest churn predictor in B2B SaaS. When the person who drove the purchase decision goes inactive (left the company, changed roles, or lost interest), churn probability increases 8x within the next renewal cycle.
How do you detect competitive threats across accounts?
Relational models detect cross-account patterns: when multiple accounts in the same industry vertical simultaneously reduce usage or adopt a competitor's integration, it signals a competitive threat. This vertical-level signal is invisible to models that evaluate each account independently.
What is the ROI of SaaS churn prediction?
A $200M ARR company with 10% annual churn that reduces churn by 30% through early intervention saves $8M annually. The model identifies at-risk accounts 90 days before renewal, giving CSMs enough time to re-engage champions, resolve escalated tickets, and offer retention packages before the decision is made.
Bottom line: A $200M ARR SaaS company that identifies at-risk accounts 90 days before renewal and intervenes effectively reduces net revenue churn by 30%, saving $8M annually. Kumo detects champion departure, usage decay, and cross-account competitive signals that health-score spreadsheets miss.
Related use cases
Explore more B2B SaaS use cases
Topics covered
One Platform. One Model. Infinite Predictions.
KumoRFM
Relational Foundation Model
Turn structured relational data into predictions in seconds. KumoRFM delivers zero-shot predictions that rival months of traditional data science. No training, feature engineering, or infrastructure required. Just connect your data and start predicting.
For critical use cases, fine-tune KumoRFM on your data using the Kumo platform and Research Agent for 30%+ higher accuracy than traditional models.
Book a demo and get a free trial of the full platform: research agent, fine-tune capabilities, and forward-deployed engineer support.
