Most of your retention budget goes to people who were never going to leave, and some of your outreach causes the churn it was meant to prevent. Presage prices both against your own history, read‑only, in fifteen minutes.
| Member | Uplift | Recommendation |
|---|
Margin you give each year to customers who would have stayed anyway
How this is computed. Offers sent, times the share landing on customers who stay regardless, times average redeemed offer cost. Two priors carry the estimate: 70% of retention offers reach customers with no meaningful uplift, and average redeemed offer cost runs 4.8% of annual customer value. Both are PROVISIONAL, pending the design-partner benchmark, and your own history replaces them at tier T2. No forecast is involved. This is arithmetic on offers you have already sent.
Churn in the three months after a proactive retention campaign, measured against a randomised control group.
Ascarza, Iyengar & Schleicher, The Perils of Proactive Churn Prevention,
Journal of Marketing Research, 2016.
Read the paper
Every book splits four ways on one question: what does contact actually change? Hover a class to find it in a sample of 600 members.
600 members, none selected.
Stays whether you contact them or not. Every retention offer sent here is margin handed over for a decision already made.
Would have stayed, until you reminded them they could leave. The contact itself is what moves them out.
The only class where contact changes the outcome in your favour. Usually the smallest of the four.
Leaves regardless. Offers here buy a short delay at full discount cost, then the same outcome.
We separate them in public because the difference decides how much you should believe each one.
Retention offers that landed on Sure Things and Lost Causes. Margin given away for an outcome that was already decided.
Sleeping Dogs suppressed rather than contacted, valued at the remaining lifetime margin the contact would have cost you.
Incremental retained margin on Persuadables, measured against a randomised control arm. This is the line we prove at renewal, not at signature.
Uplift estimates the difference contact makes to each customer rather than their probability of leaving. A customer at 80% churn risk whom nothing can save is worth less attention than one at 30% whom a call moves. A lifecycle model places each customer on the latent path from engaged to gone, so that uplift figure is read against how much time is actually left.
The stack resolves both into one instruction per customer per period: suppress, observe, re-engage, offer, or expand. One class each period, so no contact can be credited twice. For most customers the instruction is NO ACTION.
| Tier | Label | Source | Where it appears |
|---|---|---|---|
| T1 | Benchmark | Cross-customer priors and published effect sizes | The free audit, before any of your data is modelled |
| T2 | Modelled | Your own history, applied counterfactually to past periods | Audit output and business case, quoted at the lower bound |
| T3 | Measured | Randomised holdout on your live book | The renewal. The only tier allowed to use the word caused |
Optimove's own documentation states that control groups have no impact on its decisioning algorithm. For them the control arm is a line in the report. For us it is the objective function the model trains against.
It will not contact a customer whose outreach causally raises their churn, even if you ask it to.
It will not authorise retention spend on a customer who was going to stay.
The method needs volume and history to say anything defensible. Below these lines it produces wide intervals and confident-sounding noise, which is the thing this product exists to avoid.
If the audit comes back small, we will tell you and end the conversation.
| The objection | The answer |
|---|---|
| “Braze and Optimove already do holdouts.” | They report the holdout. We optimise on it. Optimove's own documentation states that control groups do not feed the algorithm. |
| “Sending less means fewer conversions.” | Only for contacts with positive uplift. We raise spend on those and stop it on the other three classes, so what changes is where the budget lands. |
| “Another system for my team to learn.” | Read-only overlay. Output arrives as tasks in the CRM you already use. Nobody logs into a second tool. |
| “How do I know the waste number is real?” | It is arithmetic on offers you already sent. Every figure carries its tier, confidence interval and sample size. |
Ascarza, Iyengar and Schleicher. A field experiment in which the contacted group churned at 10% against 6% in the control arm.
SSRN Journal of Marketing Research, 2018Ascarza. The customers most likely to churn and the customers retention spending can actually move turn out to be largely different people.
SAGE Vendor documentation“Control groups have no impact on the AI Journey Decisioning Agent algorithm.” The basis for our claim about how holdouts are used.
Optimove Academy Vendor documentationThe incumbent claim we are measured against, in its own words. Worth reading before the competitive call rather than after.
braze.comFitness chains, telecoms, insurers, media. Tens of thousands of low-touch members, a save-offer budget nobody has measured against a control arm, and a churn number that has not moved in three years.
Customer success teams running QBRs and health scores on a book too large to touch individually. Who is at risk is the easy question. Which of them a human hour changes is the one that decides where the week goes.
Read-only, fifteen minutes. If the recoverable number is small, we will say so.