Ten years ago, my team and I tackled our first major retention projects. The timing was impeccable. Nick Mehta had just dropped “Customer Success,” essentially writing the holy text for the new era of retention management. Suddenly, the tools caught up with the theory: Google rolled out Universal Analytics, Salesforce swallowed a rising star called Tableau, Python stopped being just a scripting language, and “Data Scientist” became the sexiest job title in tech.
I loved that era. I called what we were doing “inverted performance marketing” – applying the hardcore, data-driven ruthlessness of user acquisition to the people we already had. It was a paradigm shift. And the funny thing is, the core tenets of that 2016 shift haven’t really changed:
- Proactive > Reactive: Stop waiting for the support ticket. Anticipate the friction.
- Retention = Growth: As Fader and Hardie proved mathematically with their Buy ‘Till You Die models, optimizing Customer Lifetime Value (CLTV) will always beat aggressively burning cash on acquisition.
- Quantified Health: Gut feelings and lazy NPS scores are dead. Product usage and login frequency are the new vital signs.
- No Silos: If Product builds a useless feature, the Success team can’t save the user. It’s a company-wide game.
It’s all very obvious now. But I bring it up to highlight a slightly depressing reality.
Last week, as I started hunting for a full-time role to settle down here in the Netherlands, I came across a Head of Retention gig at a major fintech company. Now, fintech – alongside gaming and healthcare – is one of the absolute holy grails for pure, unadulterated client data flow. It’s juicy. It got me thinking about everything we built back in the 2010s, and it hit me:
The market is still stuck in 2016.
Sure, the dashboards are prettier, and user attention spans have flatlined, but the workflow? It’s identical. We run a cohort analysis. We spot a drop-off. We deploy a localized friction-fix or a targeted email sequence. We go home, kiss our partners, and wait a month to see what happened. Then we check the numbers, calculate the new retention rate, and pick a new cohort.
It feels like work. But fundamentally, we are just counting corpses.
We separate the terminally ill clients from the treatable ones, calculate the triage costs, and try to patch the bleeding. We are reacting to the inevitable. Nick Mehta’s dream was preventing churn, but we settled for diagnosing it.
Is it really that hard to predict churn? We know the actual churn rate of a cohort only after they leave. The cost of that data is the lost revenue. So, do retention marketers actually have prediction tools? Not really.
But you know who does? The Fraud Detection department.
Those guys have a 30-year head start on us. They don’t wait for a stolen credit card to clear; they calculate the probability of fraud in milliseconds using algorithms like Random Forests. And the crazy part is, that tech is now democratized. Give Claude or Cursor some sanitized test data and a few tokens, and it will spin up a Random Forest model for you in minutes. (Implementing Neural Networks or Gradient Boosting (XGBoost) could cost you extra like a meal in fine dining in token costs).
The only output you actually need from it is RP3m (Retention Probability in 3 months). Just one continuously updating float variable attached to every single user.
If you have that, three fundamental things change overnight:
- You stop ignoring the healthy. I want to see the RP3m fluctuate for users in the “safe green zone,” not just the red zone. That’s the only way to catch macro market shifts before they hit your bottom line.
- You kill the waiting period. I don’t want to wait a month to see the results of an A/B test (bleeding users the whole time). I want to push a UX fix and immediately watch the prediction model recalculate the RP3m.
- You remove human bias from cohorts. With modern data volume, picking cohorts based on “manager intuition” is negligent. Cohort selection should be a systemic, automated morning insight delivered by an AI agent.
Now, a CFO might look at this and say: “Ivan, integrating predictive AI infrastructure costs a tech startup 100 grand. Are we spending that just to shave a few days off A/B testing and find a micro-cohort of users who like My Little Pony? The ROI takes years.”
Yes. And no.
I chose the RP3m marker because, mathematically speaking, the lifetime churn probability of every single client is 1. Valar Morghulis. All men must die, and all clients must churn.
There are only two exceptions to this rule. First is a scam, where the business conveniently dies before the client gets a chance to leave.
The second exception is what I call the Padel Effect. Padel is that trendy racquet sport sweeping Europe. You literally cannot play it alone. You need four people. If one guy moves out of the country, the remaining three don’t quit the sport – they immediately recruit a replacement to keep the game alive.
When your product becomes a network, brand advocacy makes the client immortal.
And that is the real ROI of predictive modeling. Once you build a model to predict the probability of an event, you don’t just point it at churn. You invert it. You point it at Advocacy Probability.
Forget diagnosing failure. Use the model to predict the exact minute of true, unadulterated user success – the exact peak moment when they have the highest potential to transition from a power user to a brand advocate. I don’t mean the onboarding phase. I mean the precise moment they are mathematically most likely to share their win, their story with a colleague.
A decade ago, catching that organic moment was impossible, so marketers faked it with engineered “viral loops.” Today, predictive AI can hand us that moment on a silver platter.
That is the actual retention revolution. Or maybe it isn’t, and I’m just getting ahead of myself. I’m still in talks with the hiring manager. I’ll let you know how it goes.