The discount that cannot repair the reason for leaving

A telecom customer appears at the top of a churn-risk list after three failed service visits and a disputed bill. The retention engine offers ten per cent off a new twelve-month contract. The customer declines, repeats the complaint and receives another message two days later. The model was right that departure was likely. The programme was wrong about what could change it.

A churn model estimates who may leave under current conditions. A retention decision asks whether the organisation should act, which action is likely to change the outcome, whether the effect is incremental, whether value reaches the customer and firm, and whether contact is fair. Prediction and prevention are different quantities governed by different evidence.

High-risk customers may have decided to leave, be impossible to persuade, face an external change or require a repair that an offer cannot deliver. Low-risk customers may stay naturally and collect an unnecessary incentive. Some can be annoyed into departure by pressure. A highly accurate risk rank can therefore produce an expensive, unfair and negative retention policy.

The Causal Retention Operating System developed here has twelve connected parts: Exit and Relationship State Contract; Churn Evidence and Cause Graph; Eligibility and Customer Choice Gate; Intervention Library and Authority Map; Prediction, Causation and Value Ledger; Heterogeneous Treatment Effect Engine; Service-Cause Resolution Gate; Bounded Retention Experiment Registry; Contact Burden and Fairness Control; Treatment Delivery and Receipt Layer; Incremental Outcome and Value Reconciliation; and Retention Assurance and Learning Docket.

Define departure before scoring it

The Exit and Relationship State Contract identifies customer, account, product and horizon; voluntary cancellation, non-renewal, inactivity, competitor switch or involuntary closure; full exit, downgrade, partial defection and share-of-wallet decline; grace period, contract state, use and value realisation; complaints, affordability, prior save attempts, later reactivation, contact preference and authoritative receipt.

Thirty days of inactivity is not churn merely because it is easy to label. The customer may be seasonal, between purchases, unable to access the product, using another product or trapped in a failed cancellation. A relationship that has reached its intended end is not necessarily a failure. Clean exit can protect autonomy and reveal a product-market mismatch that discounts would conceal.

Keep risk, effect and value in separate ledgers

Four estimates are often collapsed into one number. Churn risk is the probability of exit under current conditions. Treatment effect is the difference in exit probability with versus without a specific action. Customer outcome value asks whether that action improves fit, service, access or realised benefit. Firm value is incremental margin after delivery, incentive, service, risk and capacity cost.

Multiplying churn probability by lifetime value does not produce preventable value. The calculation assumes that treatment changes the outcome, that the customer would not stay anyway, that retained use and margin will materialise, that delivery is available and that contact creates no harm. Each assumption needs evidence and an uncertainty range.

Useful ranges include baseline survival, treatment take-up, incremental response, retained duration, margin, discount and remedy cost, future support, cannibalisation, strategic behaviour, complaint burden, correlated campaigns, drift and population shift. A narrow point estimate can make weak identification look like operational certainty. The ledger preserves where randomisation exists and where observational assumptions carry the result.

The denominator is a population of decisions

Retention reporting needs the customers who were eligible for a decision, not merely the people selected by a score. Record who was excluded because of contact preference, open complaint, unavailable remedy, jurisdiction, capacity or uncertain relationship state. Otherwise a programme can show impressive response among the easiest cases while silently abandoning customers for whom the treatment could not be delivered.

Timing belongs in the denominator too. A score produced after a cancellation request is leakage when the decision was supposed to prevent departure before intent hardened. A renewal observed before a delayed cancellation is not durable retention. The state contract therefore fixes prediction time, intervention window, grace period and outcome horizon, then rejects features or labels that arrive after the legitimate decision point.

Policy value depends on opportunity cost. A specialist assigned to a natural stayer cannot repair another customer's preventable failure. A discount given to someone who would renew anyway consumes margin and may create a loyalty penalty when comparable quiet customers receive nothing. Capacity is allocated against expected incremental customer and firm value, with uncertainty and fairness constraints, rather than against risk or lifetime value alone.

This also changes forecasting. Finance sees a range for customers helped, natural stayers subsidised, lost causes contacted and sleeping dogs harmed. Service leaders see upstream defects whose removal could outperform any campaign. The board sees earned retention separately from delayed exit, discount dependency and relationships kept only by friction.

Four counterfactual types cannot be observed directly

Retention programmes often use four useful conceptual groups. Sure things stay whether contacted or not. Persuadables stay because of treatment. Lost causes leave in either condition. Sleeping dogs would stay without treatment but leave or suffer harm because of it. A fifth operational population belongs beside them: customers whose service failure must be repaired before a commercial intervention is even considered.

No individual arrives with one of these labels. Only one potential outcome is observed: the customer receives a treatment or does not. Heterogeneous treatment-effect methods estimate differences across populations under assumptions and uncertainty. They do not reveal an individual's missing counterfactual. Calling a person “persuadable” therefore describes a targeting hypothesis, not an established personal fact.

Thought experiment 1: a perfect churn model

Imagine a model predicts every departure correctly. It ranks one million customers without error but contains no treatment data. A campaign offers the highest-risk ten per cent a costly discount. Many have already chosen a competitor, some face unresolved service failure and others will leave despite any price. Meanwhile a moderate-risk group would respond to simple product guidance. Does perfect prediction create a good retention policy?

It does not. The model has learned the untreated outcome distribution without learning the difference made by each intervention. Ascarza's field-experimental work is central precisely because it challenges high-risk targeting as a default. Predictive accuracy, treatment sensitivity and programme profitability must be tested separately. A perfect risk model can remain causally uninformative.

Heterogeneous Treatment Effect Engines begin with randomisation where safe and feasible, preserve credible untreated or alternative groups, pre-specify outcomes and important subgroups, separate discovery from estimation where appropriate, and test policy value out of sample. Causal trees, forests and uplift methods are tools for estimation, not permission for individual persuasion.

Find the cause before choosing the treatment

The Churn Evidence and Cause Graph types customer statements, cancellation action, usage, service reliability, complaint, billing, price change, product withdrawal, delivery issue, market event, contact, offer, risk score, inferred cause, investigation, remedy and later recurrence. Each node keeps source, time, owner, permitted use, sensitivity, dispute and downstream decision.

The graph distinguishes an addressable service cause, commercial mismatch, adoption gap, completed or external need, unknown cause and harmful retention context. A decline in usage may follow a broken login, a decision already made or a naturally seasonal pattern. Risk does not settle which explanation is true.

The service-cause gate comes before the offer

Before any commercial save, the system asks whether a complaint, billing error, outage, access failure or broken promise is open; whether the customer is exercising a right; whether the organisation created the risk; whether repair can occur in time; and whether the present team holds authority. A discount cannot quietly replace a remedy owed for a service failure.

Applied example 1: subscription software adoption

A software customer uses one small feature and ignores the workflow they originally bought. The risk model rises. Before discounting, the team inspects implementation history and finds that data import failed. The treatment is a bounded specialist repair and onboarding path, not a cheaper renewal for an unusable product.

A randomised or credible phased test compares verified adoption, task completion, support burden, later renewal and realised margin. Customers for whom the product remains a poor fit receive a clear downgrade or exit path. Discount acceptance alone would misstate both customer value and prevention.

Experiment on incremental outcomes, not campaign response

A Bounded Retention Experiment Registry names the decision and causal hypothesis, eligible population, exclusions, churn definition, horizon, intervention, comparator, delivery channel, identification design, baseline balance, customer and business outcomes, cost, capacity, contact limits, fairness slices, spillover, stop owner, remedy and promotion rule before exposure.

Renewal, click or offer acceptance cannot stand alone. The experiment measures realised use, subsequent churn, complaint, discount dependency, customer value and clean exit. It verifies assignment, exposure and operational completion. A customer who accepts a credit that is never applied did not receive the treatment. A customer who remains because cancellation failed was not retained.

Contact burden accumulates across service, marketing, billing, product and loyalty channels. The control counts messages, calls, prompts, save questions, effort to decline, earlier complaints, channel preference, urgency framing, human access, portability and termination friction. Refusal is durable. A different model or team cannot re-enter the customer without a legitimate new basis.

Release a policy, not a model rank

Production release specifies which treatment may be offered to which evidenced state, through which channel, at what frequency and by which owner. It includes the minimum treatment-effect estimate, acceptable uncertainty, delivery capacity, cost ceiling and customer-value condition. The risk score can be one input to investigation, but it does not define the eligible population by itself.

Monitoring compares the deployed policy with risk-only targeting, broad treatment and no contact. It looks for drift in baseline departure and response, changes in product or service causes, deteriorating delivery, subgroup burden, repeated refusal and new channel overlap. A model update is a policy change when it alters who receives attention, even if its headline accuracy improves.

The stop mechanism must work across systems. Suppression propagates to marketing lists, service prompts, billing journeys, outbound calls and product messages. When a harmful segment effect appears, the team can suspend that treatment without disabling an unrelated service repair. When a common service defect drives risk, the system redirects capacity upstream rather than generating more individual save offers.

Independent review checks that an attractive effect has not been purchased with cancellation friction, selective delivery or missing complaints. It samples the source evidence behind inferred cause, traces the actual treatment and confirms that the promised benefit was available to comparable customers. A supplier score without assignment provenance, treatment history or outcome linkage remains unsuitable for consequential targeting.

Count who was helped, unaffected and harmed

Incremental Outcome and Value Reconciliation starts with the eligible decision population, preserves the counterfactual, verifies delivery, measures churn and customer outcomes over the agreed horizon, estimates average and heterogeneous effects, subtracts treatment and support cost, accounts for cannibalisation and harm, verifies margin and compares the deployed policy with risk targeting and no action.

Report customers targeted, delivered, helped, unaffected and harmed. Identify service causes that should be removed upstream. A positive average can conceal a harmed subgroup; a positive segment estimate can remain too uncertain, costly or capacity-constrained to operate. Reauthorise, redesign or retire on the realised policy result.

Eight cases test the operating boundary

Telecom contract. High risk alone triggers cause investigation. A plan-fit intervention proceeds only when evidence suggests positive incremental response and the same option is fairly available. Subscription software. Repair failed import and onboarding before testing price. Later renewal counts only with product use and customer value.

Retail loyalty. Test a published benefit against a credible comparator without making quiet loyal customers pay more than threatened leavers. Banking complaint. Correct the fee or service error under complaint authority before any commercial offer. Preserve the complaint record and remedy.

Insurance renewal. Use domain-specific regulatory, actuarial and conduct authority; provide clear terms and an unobstructed exit. Streaming service. Compare pause, downgrade, better discovery and clean cancellation rather than assuming a discount treats every cause.

B2B account. Distinguish full relationship exit, product contraction and a normal procurement cycle. Route evidence to product, service or commercial owners. Essential service. Protect continuity, hardship and vulnerability without using dependence as leverage or adding repeated persuasion.

Applied example 2: streaming pause versus discount

A streaming subscriber reduces use after a life change. A risk-only programme offers a discounted annual term. A causal experiment compares a one-click pause, a lower tier, a customer-selected discovery reset and clean exit. It measures resumed use, satisfaction, later cancellation, complaint and margin after delivery cost.

The pause may outperform the discount for customers whose need is temporary. Others may rationally leave. The programme creates value by matching treatment to state and by not spending incentives on natural stayers. It also learns when no contact preserves more trust than a save journey.

Negative controls include highest risk with zero effect, high lifetime value with an unresolved complaint, an expensive incentive for a natural stayer, negative treatment effect, prior refusal, unequal treatment availability, observational uplift with severe selection and short renewal followed by non-use and complaint. The console must refuse these attractive but unsupported targets.

The assurance docket follows the whole relationship

The Retention Assurance and Learning Docket joins the exit contract, cause graph, eligibility decision, treatment version, experimental design, contact history, delivery receipt, subgroup result, cost, complaint and reauthorisation decision. It records exceptions and the owner who accepted them. A campaign dashboard can summarise this evidence, but it cannot replace the underlying records.

Learning distinguishes model defects from product, service, price, process and capacity defects. If risk rises because a login release failed, better churn scoring is not the remedy. If a treatment works only while a large discount continues, the organisation has learned about price or product fit rather than durable loyalty. If customers leave after repeated offers, contact design is part of the causal system.

Closure is equally specific. The docket records verified repair, adopted treatment, sustained value, voluntary exit, portability, final bill and later reactivation as different outcomes. It does not relabel every surviving account as saved or every departure as a failure. This preserves the evidence needed to invest in better relationships rather than better-looking retention rates.

A board agenda for earned retention

Governance begins by asking how much reported retention is incremental. The board should see the eligible population, untreated survival, verified delivery, estimated treatment effect and realised margin on one page. It should also see customers who stayed without help, customers who left despite treatment and customers whose outcome worsened. A single save-rate percentage hides the operating choices that management can change.

The next question is where preventable departure originates. Rank causes by customer harm and removable volume, not just by predicted lifetime value. A recurring billing error, inaccessible product journey or failed service promise may justify investment before a new targeting model. Cause removal helps future customers without asking each person to signal departure first, and it avoids rewarding the threat to leave.

Capacity and authority should be reviewed together. A verified treatment that cannot be delivered is not an option. A customer offered a specialist repair, credit or downgrade should receive the promised action within the relevant time. If capacity is scarce, allocate it using expected incremental customer value, consequence, fairness and uncertainty. Do not let nominal account value displace an owed remedy or essential-service protection.

Finally, management should approve a retirement rule before scale. Stop when response depends on repeated contact, the untreated comparator is lost, subgroup harm breaches its ceiling, incremental margin disappears after full cost, or the product and service state no longer matches the experiment. Retirement is evidence that the organisation is learning. A campaign kept alive to protect its historical success metric is retention theatre.

These controls also change incentives for frontline teams. Staff should receive credit for confirming that no treatment is useful, enabling a clean exit and escalating a recurring service cause. Penalising those decisions pushes weak offers back into the channel and corrupts the evidence used for later models. Reward verified customer progress, honest causal learning and early retirement of wasteful interventions, not the number of accounts placed in a save queue.

Causal Retention Decision Console

The console routes a current retention decision to exactly five outcomes. It does not infer an individual's unobserved causal type. It shows the decisive reason, missing evidence, expected customer impact and next learning action. Choose a named scenario, then change individual fields to test the boundary.

The strongest rival is timely predictive intervention

The objection is operationally important. Risk can prioritise investigation, trigger service review and define bounded experiments. The organisation may act under uncertainty when consequence and burden are low, eligibility is fair, exit remains easy and the action produces useful evidence. Causal governance is not a demand for paralysis or individual certainty.

The boundary is correlation converted into persuasion authority. High risk does not authorise a discount, repeated contact or cancellation friction. Intervention must have a plausible mechanism, treatment-specific evidence, capacity and a customer benefit. Learning should compare the decision policy with risk-only targeting and no action.

Thought experiment 2: retention by broken exit

Imagine a subscription service reports its best retention quarter. The prediction model accurately identified likely cancellers, and renewal rose after a new save journey. Later evidence shows that the cancellation button failed on mobile, customers repeated the process and many remained billed without using the product. Did the programme prevent churn?

It prevented an event from being recorded, not a customer from losing value. Incremental outcome reconciliation would count blocked exit, non-use, complaint, refund and recovery. It would remove the apparent margin, stop targeting and open a service incident. A retention metric that ignores customer intent can reward captivity.

An observation that would weaken this architecture is sustained improvement in incremental customer value, durable retention, fairness and realised margin across changed populations without precise exit definitions, cause evidence, treatment controls, counterfactuals, delivery receipts or burden limits. Higher AUC, gross renewal, offer acceptance, predicted lifetime value, positive average uplift or a closed cancellation case would not establish that result.

Research sources remain bounded. Prediction studies evaluate models and stability in particular data and settings. Field experiments estimate treatment effects for their populations and interventions. Causal methods depend on design and assumptions. Valuation models are sensitive to retained duration, margin and cost. FTC and CMA materials address obstructive digital choice; GDPR, the Digital Services Act and EU AI Act apply within their scopes. NIST frameworks are voluntary risk-management resources. Accountable owners and counsel determine application to the actual programme.

Retain by improving the relationship

Begin with a real exit event and the customer's state. Preserve cause uncertainty. Repair service before offering an incentive. Separate risk, treatment effect, customer value and firm value. Test bounded interventions against credible alternatives, account for sleeping-dog harm and enforce durable refusal.

Verify what was delivered, then measure who was helped, unaffected and harmed over the proper horizon. Subtract every intervention, remedy and support cost. Let clean exit remain an authorised outcome. A retention system earns trust when it can explain not just who stayed, but whose outcome changed because of a treatment worth delivering.