Cutting Churn With AI in Subscription Portcos

Table of Contents

1. Why Churn Kills Subscription Portcos
2. The Churn Economics: What You’re Actually Losing
3. Early-Warning Models: Detecting Flight Risk Before It Happens
4. Building Your Churn Prediction Engine
5. Agentic Retention Plays: The Operating Model
6. Practical Implementation: From Data to Action
7. Measuring What Works: Retention Metrics That Matter
8. Common Pitfalls and How to Avoid Them
9. Next Steps: Building Your Retention Infrastructure

---

Why Churn Kills Subscription Portcos

Subscription businesses live or die on churn. A 5% monthly churn rate doesn’t sound catastrophic until you map it across a $10M ARR portfolio company: you’re replacing your entire customer base every 20 months. That’s not growth. That’s a treadmill.

For private equity operating partners, churn is the hidden tax on every acquisition. It erodes EBITDA, kills unit economics, and forces you to spend more on customer acquisition to stand still. In a 3–5 year hold, that compounds into tens of millions of value destruction.

Yet most portfolio companies treat churn as inevitable. They track it, they report it, but they don’t engineer it. They don’t build systems to predict which customers will leave, and they certainly don’t deploy intelligent, scalable interventions before the customer hits the cancel button.

AI changes that. Not through magic, but through precision. Early-warning models let you identify flight-risk customers 30–90 days before they churn. Agentic retention systems let you run thousands of personalised retention plays simultaneously—the right offer, to the right customer, at the right moment, without burning your team’s bandwidth.

This is how you turn churn from a leak into a controlled valve.

---

The Churn Economics: What You’re Actually Losing

The Math That Matters

Churn compounds. If you have $10M ARR and 5% monthly churn, you’re losing $500K in recurring revenue every month. Over a year, that’s $6M in gross churn. Your sales team has to replace it just to stay flat.

But the real cost is deeper. Research on customer retention from McKinsey shows that reducing churn by just 5% can increase customer lifetime value by 25–95%, depending on the business model. That’s not a rounding error. That’s value creation.

For a $50M ARR subscription business with 40% gross margins, a 1% improvement in annual churn retention could be worth $2M–$5M in incremental EBITDA by year three. That’s the operating leverage you’re after.

Why Churn Accelerates

Churn isn’t random. It clusters around moments: after a price increase, after a major feature breaks, after a customer hits a usage ceiling, after they’ve been ignored for six months. These are signals. They’re detectable. But only if you’re watching.

Most subscription teams have churn data buried in a CRM or a data warehouse. They see it monthly, in a cohort table, after the fact. By then, the customers are gone.

The companies that win—and the ones you’ll want to acquire or improve—are the ones that catch churn signals in real time. They know which customer is at 80% of their usage limit. They know which account hasn’t logged in for 45 days. They know which customer just got a new CFO who’s likely to audit all their SaaS spend.

That’s where AI comes in.

---

Early-Warning Models: Detecting Flight Risk Before It Happens

What a Churn Prediction Model Actually Does

A churn prediction model is a machine-learning system that ingests customer behavioural data—login frequency, feature usage, support tickets, payment failures, contract milestones—and outputs a probability: “This customer has a 73% chance of churning in the next 90 days.”

That’s not a guess. It’s a pattern learned from thousands of historical customer journeys. When you feed it current data, it identifies which customers today match the patterns of customers who left yesterday.

Academic research on churn prediction shows that well-tuned machine-learning models outperform rule-based churn detection by 20–40% in precision and recall. That means fewer false alarms (you’re not wasting retention resources on customers who weren’t going to churn) and fewer missed signals (you’re catching the ones who really are at risk).

The best models combine behavioural signals (usage, engagement), transactional signals (payment health, contract stage), and contextual signals (industry trends, competitive activity, the customer’s own growth trajectory).

The Signals That Predict Churn

Not all signals are equal. Some are noise. Others are early warnings.

Usage collapse is the loudest signal. If a customer drops from 40 logins per week to 2, they’re either solving their problem elsewhere or they’ve deprioritised the problem. Either way, churn is coming. A well-built model will flag this within 7–14 days of the trend starting, not after two months of silence.

Feature adoption failure is another strong signal. If a customer paid for a premium tier but never uses the premium features, they’re paying for something they don’t value. Upgrade churn is often as predictable as downgrade churn.

Support ticket velocity matters. An increase in support tickets often precedes churn—the customer is struggling, they’re asking for help, and if you don’t solve it, they leave. But a sudden drop in support tickets can be worse: it means they’ve given up asking.

Payment friction is immediate. Failed payments, declined cards, payment plan changes—these are moments when customers reconsider whether they still want the product. If you don’t intervene within 48 hours, the churn rate doubles.

Contract milestones are predictable churn windows. A customer’s renewal date, their annual review cycle, a price increase announcement—these are moments when customers actively evaluate alternatives. Your model should flag these 90 days out, not on the day the contract expires.

Competitive activity can be inferred from behaviour. If your customer just attended a competitor’s webinar, or if their industry peer just signed a competitive contract, your churn risk goes up. If you have access to firmographic data or market intelligence, fold it in.

Harvard Business Review research on churn measurement emphasises that understanding why customers churn is as important as predicting that they will. Your model should surface not just risk scores, but the specific factors driving that risk for each customer.

---

Building Your Churn Prediction Engine

The Data Foundation

You can’t build a churn model without data. But you don’t need perfect data.

Start with what you have: your product database (logins, feature usage, session duration), your billing system (subscription dates, payment history, plan tier), your CRM (support tickets, customer interactions, account health notes), and your email platform (open rates, click rates, engagement). That’s your baseline.

If you have it, add contextual data: company size, industry, geography, how they found you, what they were promised in sales. If you can licence it, add third-party signals: funding events, leadership changes, headcount growth, public news.

The goal is a single customer dataset where each row is a customer and each column is a signal. You’ll have hundreds of potential signals. Most will be noise. Your job is to find the 20–30 that actually predict churn.

Data quality matters, but perfection is the enemy of done. A model built on 80% complete data that ships in four weeks beats a perfect model that ships in six months. You’ll iterate. You’ll add signals. You’ll improve. But you need to start.

Choosing Your Modelling Approach

You have three main options:

Logistic regression is the simplest. It’s interpretable, fast to train, and often performs as well as fancier methods on structured data. If you’re starting from zero, this is where you begin. You can build it in Python in a day, deploy it in a week, and understand exactly which signals matter.

Gradient boosting (XGBoost, LightGBM, CatBoost) is the workhorse. It handles non-linear relationships, feature interactions, and messy data better than logistic regression. It’s slightly harder to interpret, but only slightly. Most winning churn models at scale use gradient boosting.

Deep learning (neural networks) is overkill for most subscription churn problems. It requires more data, more compute, and more tuning. Unless you have millions of customers and thousands of features, it’s not worth the complexity.

For a portfolio company with $10M–$100M ARR, gradient boosting is your sweet spot. It’s powerful enough to capture real patterns, simple enough to deploy in weeks, and interpretable enough that your customer success team can understand why a customer is flagged as high-risk.

Google Cloud’s Vertex AI, Microsoft Azure Machine Learning, and Amazon SageMaker all provide managed environments for building, training, and deploying these models. You don’t need to build your own ML infrastructure. Use theirs.

Training and Validation

Here’s the critical bit: your model is only as good as your training data.

You need historical data on customers who churned and customers who didn’t. Ideally, you want 12–24 months of history. For each customer, you need to know: what were their signals 90 days before they churned (or before your observation window ended, if they didn’t churn)?

Then you split your data: 70% for training, 15% for validation, 15% for testing. You train your model on the training set, tune hyperparameters on the validation set, and evaluate performance on the test set. That test set is sacred—you never touch it until the end.

Your success metric is not accuracy. Accuracy is a trap. If 95% of your customers don’t churn, a model that predicts “no churn” for everyone is 95% accurate and completely useless.

Instead, optimise for precision and recall:

• Precision: Of the customers you flag as high-risk, what percentage actually churn? You want this high—ideally 60%+—so you’re not wasting retention resources on false alarms.
• Recall: Of the customers who actually churn, what percentage did you catch? You want this high too—ideally 70%+—so you’re not missing the ones who really are at risk.

There’s a tradeoff. You can increase recall by lowering your threshold (flag more customers as at-risk), but you’ll tank precision. You can increase precision by raising your threshold, but you’ll miss churners. Your job is to find the sweet spot for your business.

For a $50M ARR company with 5% monthly churn, flagging 15% of customers as high-risk with 70% precision means you’ll identify 70% of churners and run 15 retention plays for every 10 that actually churn. That’s acceptable. If you can get to 75% precision, you’re elite.

---

Agentic Retention Plays: The Operating Model

What “Agentic” Means in Retention

Agentic systems are AI systems that can observe state, decide on action, and execute autonomously, within guardrails. In retention, that means: the system sees a customer is at high churn risk, decides what intervention to deploy, and executes it—without a human in the loop.

That’s the unlock. You can’t manually run 500 retention plays a month. You can run 5,000 if they’re agentic.

The agent doesn’t make arbitrary decisions. It operates within a decision tree you’ve built:

• If churn risk is >80% and the customer is in their renewal window, offer a 15% discount for annual commitment.
• If churn risk is 60–80% and they haven’t used feature X, assign a customer success manager for a onboarding call.
• If churn risk is 40–60% and they’re a high-LTV customer, send a personalised product roadmap aligned to their use case.
• If churn risk is <40%, do nothing (preserve your team’s bandwidth).

Each branch can have sub-branches. Each action can be personalised: the discount varies by customer LTV, the CSM assignment is routed to the right specialist, the roadmap is tailored to their industry.

The system executes these plays at scale. It sends emails, it creates tasks in your CRM, it triggers Slack alerts, it updates customer health scores. It does the work that would take your team 200 hours a month to do manually.

Designing the Decision Tree

Your decision tree should reflect your business model and your team’s capacity.

Start simple. You have three levers:

1. Discounts and incentives: Price reductions, free months, extended trials, feature unlocks. These work, but they train customers to expect discounts. Use sparingly.

2. Engagement and education: Onboarding calls, product demos, webinars, personalised tutorials. These work if the customer’s problem is lack of knowledge. They don’t work if the customer’s problem is the product doesn’t fit.

3. Product and feature: Implementing features they’ve requested, fixing bugs they’ve reported, building integrations they need. This is the most expensive lever, but it’s the most sustainable. If you can ship a feature that moves a customer from 40% churn risk to 10%, you’ve won.

For a portfolio company, your decision tree might look like:

IF churn_risk > 0.80 AND customer_ltv > $50K AND days_to_renewal < 90:
  THEN offer 20% annual discount + executive business review
ELSE IF churn_risk > 0.80 AND customer_ltv < $50K:
  THEN send educational content + assign CSM call
ELSE IF churn_risk > 0.60 AND feature_adoption < 0.3:
  THEN trigger onboarding flow + send feature tutorial
ELSE IF churn_risk > 0.60 AND payment_health == failed:
  THEN send payment recovery email + offer payment plan
ELSE IF churn_risk > 0.40 AND days_since_login > 30:
  THEN send re-engagement email + free training offer
ELSE:
  THEN do nothing

This tree is deterministic. It’s not magic. But it’s systematic, and it scales.

Personalisation at Scale

The magic isn’t in the tree. It’s in the personalisation.

When you send a retention email to a high-risk customer, the email shouldn’t be generic. It should reference their specific use case, their specific pain point, their specific contract terms. If they’re a fintech and they haven’t adopted your compliance reporting features, mention compliance. If they’re a retailer and they’re struggling with inventory forecasting, mention that.

This is where agentic systems shine. The agent can:

• Pull the customer’s industry, use case, and feature adoption from your data warehouse.
• Generate a personalised email subject line and body using a language model.
• Insert the right discount code (different for each customer based on their LTV and churn risk).
• Schedule the send at the optimal time (when they’re most likely to open it).
• Track opens, clicks, and responses, and trigger follow-up actions if they don’t engage.

All of this happens automatically. All of this scales to thousands of customers.

Research on personalisation from McKinsey shows that personalised retention offers have 2–3x higher conversion rates than generic ones. If your baseline retention offer has a 20% acceptance rate, personalisation can push it to 40–60%.

That’s the lever that moves the needle.

---

Practical Implementation: From Data to Action

The 90-Day Build Plan

You don’t need to boil the ocean. You can have a working churn prediction system and agentic retention plays running in 90 days. Here’s how:

Weeks 1–2: Data Assembly

Pull your historical customer data. You need: customer ID, signup date, churn date (or last activity date), subscription tier, monthly spend, and behavioural signals (logins, feature usage, support tickets). Aim for 12 months minimum.

If you don’t have all of this in one place, build a SQL query that joins your product database, billing system, and CRM. This is tedious. Do it anyway. This is your foundation.

Weeks 3–4: Exploratory Analysis

Build a cohort analysis. What’s the churn rate by customer segment (by industry, by company size, by product tier)? What’s the churn rate by cohort (customers acquired in month 1 vs. month 12)? What’s the correlation between feature adoption and churn? Between support ticket volume and churn?

You’re not building a model yet. You’re understanding your data. You’re finding the signals that obviously predict churn. These become your first features.

Weeks 5–8: Model Development

Build your first churn model. Start with logistic regression or gradient boosting. Use your top 20–30 signals. Train on 12 months of data, validate on the next month, test on the month after that.

Don’t obsess over performance. A model with 70% precision and 65% recall that ships in week 6 is better than a model with 78% precision that ships in week 10. You’ll improve it iteratively.

Deploy your model to a data warehouse or a cloud ML platform. Set it up to score all your current customers weekly.

Weeks 9–10: Decision Tree Design

With your team, design your retention decision tree. What interventions will you deploy? What’s the budget? What’s the success metric?

Start conservative. You’re not trying to save every customer. You’re trying to save the ones where the economics make sense. A $5K annual contract isn’t worth a $2K retention discount. A $100K contract is.

Weeks 11–12: Agentic Execution

Build the automation. Connect your model output to your CRM, email platform, and task management system. Write the logic to:

1. Identify high-risk customers.
2. Determine which intervention to deploy.
3. Execute the intervention (send email, create task, trigger Slack alert).
4. Log the action and track the outcome.

You don’t need a fancy platform. A Python script running on a cron job, a Zapier workflow, or a basic workflow engine will do. The goal is to execute your decision tree automatically.

Technology Stack

You don’t need bleeding-edge tech. You need reliable tech that integrates with your existing systems.

Data warehouse: Snowflake, BigQuery, or Redshift. These let you centralise your data, write SQL queries, and run analytics. They integrate with most ML platforms.

ML platform: Vertex AI, SageMaker, or Azure ML. These provide managed environments for training and deploying models. You don’t manage infrastructure. You upload data, define your model, and get predictions back.

Workflow automation: Zapier, n8n, or a custom Python script. This orchestrates your decision tree—it takes model scores, applies your logic, and executes actions.

CRM and email: Salesforce, HubSpot, or Pipedrive for CRM. SendGrid, Mailchimp, or your CRM’s native email for sending. These are where your retention plays land.

Total cost for a portfolio company: $5K–$15K per month in cloud and SaaS tools. For a $50M ARR business, that’s 0.1–0.3% of revenue. The payoff is 10–20x that in recovered churn.

Governance and Guardrails

Agentic systems need guardrails. You can’t let an AI system offer unlimited discounts or make promises about product features.

Set hard limits:

• Discount cap: No offer greater than 30% of annual contract value.
• Frequency cap: No more than 2 retention offers per customer per quarter.
• Feature promise cap: Retention emails can only mention features already on your roadmap, not vaporware.
• LTV filter: Only deploy expensive interventions (executive calls, custom features) for customers with $50K+ annual contract value.

Have a human review high-stakes decisions. If the system wants to offer a $100K discount to a customer, it should flag that for approval. If it wants to promise a custom feature, it should route to your product team.

Monitor for drift. As your product, customer base, and market change, your model’s performance will degrade. Retrain it quarterly. If precision drops below 60%, pause the system and investigate.

---

Measuring What Works: Retention Metrics That Matter

The Right Metrics

Track three things:

1. Churn rate by risk segment: What percentage of customers flagged as high-risk actually churn? What about medium-risk and low-risk? This tells you if your model is calibrated.

2. Retention lift by intervention: Of customers who received a discount offer, what percentage were retained? What about customers who received an onboarding call? What about a control group (customers you didn’t intervene on)? This tells you which interventions work.

3. Return on retention spend: How much did you spend on retention offers and interventions? How much revenue did you save? What’s your payback period?

Don’t obsess over aggregate churn rate. It’s a lagging indicator. It tells you what happened, not what’s happening. Your leading indicators are:

• Customers at high churn risk: How many customers are flagged as >70% churn probability? Is this number growing or shrinking?
• Usage trends: What percentage of customers are showing usage decline? Are new cohorts adopting features faster or slower than previous cohorts?
• Intervention velocity: How many retention plays did you execute this month? What was the acceptance rate?

These tell you if your system is working now, not six months from now.

Benchmarking

Your numbers are only meaningful relative to your baseline and your peers.

For SaaS:

• B2B SaaS: 3–7% monthly churn is typical. 1–3% is excellent. >10% is a crisis.
• B2C SaaS: 5–10% monthly churn is typical. 2–5% is excellent. >15% is a crisis.
• Enterprise SaaS: 1–3% monthly churn is typical. <1% is excellent. >5% is a crisis.

Your portfolio companies probably sit in the 5–10% range. A 2-percentage-point improvement in monthly churn is a $2M–$5M value creation event for a $50M ARR business.

That’s your north star. Track it monthly. Report it to your board.

---

Common Pitfalls and How to Avoid Them

Pitfall 1: Building a Model Nobody Uses

You build a beautiful churn model. Precision: 75%. Recall: 72%. Your data science team is proud. Your customer success team ignores it.

Why? Because the model outputs a number (“73% churn probability”) but doesn’t tell them what to do. They don’t have time to interpret a model. They have time to act on a decision.

Fix: Don’t deploy a model. Deploy a decision tree. Don’t say “this customer has 73% churn risk.” Say “this customer is flagged for a retention call this week.” Make the action obvious.

Pitfall 2: Overselling Discounts

Your first instinct is to discount. Customer at risk? Offer 20% off. Still at risk? Offer 30% off. Still at risk? Offer 50% off.

You retain the customer. You also train them that they can always negotiate down. Next renewal, they expect a discount. You’ve traded margin for retention.

Fix: Use discounts as a last resort, not a first move. Start with engagement and education. Only discount if the customer’s problem is actually price sensitivity. For most churn, it’s not.

Pitfall 3: Ignoring Cohort Effects

You build a model on all your historical data. But customers acquired in 2020 behave differently from customers acquired in 2023. Your product has evolved. Your market has evolved. Your model learns patterns from 2020 customers that don’t apply to 2023 customers.

Fix: Retrain your model quarterly on recent data only (last 12 months). If you’re predicting churn for customers acquired in the last 6 months, train on customers acquired 12–18 months ago (so you have churn outcomes). Don’t mix old and new cohorts.

Pitfall 4: Confusing Correlation With Causation

Your model finds that customers who open support tickets have higher churn. You conclude: “Support tickets cause churn.” You tell your support team to close tickets faster.

What’s actually happening: customers who are struggling open support tickets. If you don’t solve their problem, they churn. The ticket is a signal, not a cause.

Fix: Use your model to identify risk, not to diagnose problems. When the model flags a customer as high-risk, investigate why. Did they open a support ticket because the product doesn’t fit, or because they need help? These require different interventions.

Pitfall 5: Deploying Too Fast

You build a model. It looks good. You deploy it to production and automate all your retention plays.

Then you realise the model is wrong. It’s flagging customers who aren’t actually at risk. You’ve wasted budget on false alarms. You’ve annoyed customers with unsolicited offers.

Fix: Pilot on a segment first. Deploy to your top 20% of customers by LTV. Run it for a month. Measure precision and recall. If it’s >65% precision, expand to the next 20%. Don’t go full-scale until you’re confident.

---

Next Steps: Building Your Retention Infrastructure

For Operating Partners at PE Firms

Churn is your biggest lever on EBITDA. It’s not sexy. It doesn’t make for good board presentations. But it’s where the money is.

Here’s what to do:

1. Audit your current churn. Get the data. Understand the cohort analysis. What’s your monthly churn? What’s driving it? Is it price-sensitive customers? Feature-fit issues? Competitive displacement? You can’t fix what you don’t understand.

2. Hire or contract for churn prediction expertise. You don’t need a full data science team. You need one person (internal or external) who can build a churn model in 6–8 weeks. If you don’t have that person on staff, PADISO can help you build it. We’ve built churn models for 15+ portfolio companies. We know the playbook.

3. Allocate budget for retention infrastructure. You need $50K–$100K to build the model, integrate it with your systems, and operationalise it. That’s a 6-month payback on a $50M ARR business with 5% monthly churn.

4. Set a churn reduction target. Pick a number: 1-percentage-point reduction in monthly churn over 12 months. That’s aggressive but achievable. That’s worth $2M–$5M in EBITDA.

5. Measure and iterate. Track churn monthly. Track intervention performance weekly. Adjust your decision tree. Improve your model. This is not a set-and-forget system. It’s an operating rhythm.

For Founders and CEOs

If you’re running a subscription business and you don’t have a churn prediction system, you’re leaving money on the table.

Start today:

1. Export your data. Pull your last 12 months of customer data: signups, churn, usage, support tickets, payments. Put it in a spreadsheet or a database.

2. Analyse it yourself. Don’t wait for a data scientist. Ask: which customers churned? What did they have in common? When did they start showing warning signs? You’ll find patterns just by looking.

3. Build a simple rule-based system. You don’t need ML. You can build a churn risk score with simple rules: if usage dropped >50%, if payment failed twice, if it’s been >60 days since login, flag as high-risk. Score your customers.

4. Manually intervene on the top 20 at-risk customers this month. Call them. Ask why. Offer to help. You’ll learn more in those 20 conversations than in any data analysis.

5. Then build the ML model. Once you understand your churn drivers, you can train a model to automate what you’ve learned.

If you need help, PADISO’s AI Advisory Services can guide you through this. We’ve done it for 50+ startups. We know what works.

Building the Right Team

You don’t need a huge data science team. You need:

1. One ML engineer (full-time or fractional). This person builds and maintains the model. They understand the business, the data, and the ML. They’re not a PhD. They’re a pragmatist.

2. One data engineer (full-time or fractional). This person builds the data pipelines that feed the model. They own data quality. They’re obsessive about accuracy.

3. One product manager or operator (full-time). This person owns the decision tree and the retention playbook. They work with customer success to design interventions. They measure outcomes.

For a $50M ARR company, that’s 1.5–2 FTE. Cost: $200K–$300K per year. Payoff: $2M–$5M in recovered churn. You do the math.

If you don’t have these people in-house, PADISO can provide fractional CTO leadership to build and operate this system. We’ve done it. We know the path.

The Technology You’ll Need

You’ll integrate:

• Your product database (usage, feature adoption, session data)
• Your billing system (subscription dates, churn dates, payment history)
• Your CRM (customer interactions, account health notes)
• Your email platform (engagement history)
• A cloud ML platform (model training and serving)
• A workflow engine (decision tree execution)

If you’re a financial services business in Australia, you’ll need to ensure everything is APRA, ASIC, and AUSTRAC compliant. If you’re in insurance, you’ll need LIF compliance. We can help you navigate that.

If you need to pass SOC 2 or ISO 27001 compliance as part of your infrastructure build, we can help with that too. We’ve guided 30+ companies through Vanta for audit-readiness.

The 12-Month Roadmap

Month 1–2: Data assembly and exploratory analysis. Month 3: First churn model in production. Month 4–6: Pilot retention plays on top 20% of customers. Month 7–9: Scale to all customers. Refine decision tree based on results. Month 10–12: Iterate on model performance. Add new signals. Improve precision.

By month 12, you should have:

• A churn prediction model with >70% precision.
• Agentic retention plays running automatically.
• A 1–2 percentage-point reduction in monthly churn.
• A clear roadmap for the next 12 months.

That’s worth $2M–$10M in value, depending on your business size.

---

Conclusion: Churn as a Competitive Advantage

Most subscription businesses treat churn as inevitable. They budget for it. They forecast it. They accept it.

The best ones engineer it. They predict it. They prevent it. They turn churn from a leak into a controlled, measurable, optimisable system.

That’s where AI comes in. Not as magic. Not as hype. But as a tool to see further, act faster, and scale interventions that would otherwise be impossible.

A churn prediction model doesn’t make you special. Dozens of companies have them. But a churn prediction model plus agentic retention plays plus disciplined measurement plus a team that’s obsessed with retention? That makes you special.

That’s the operating model that wins. That’s the system that protects recurring revenue. That’s the lever that creates value.

If you’re running a subscription portco and you’re not doing this, start today. If you need help, reach out. We’ve built this playbook for 50+ companies. We know what works. We know what doesn’t. We can compress your learning curve from 18 months to 90 days.

Your churn is your destiny. Make it your competitive advantage.