CRM Manager interview questions

Retention should be modeled as a compounding driver of the active customer base, not as revenue credited to CRM sends.

• The growth identity should reconcile opening customers, acquisition, reactivation, expansion, contraction, and churn so retention changes can be traced to future base size and cash flow.
• Retention raises the value of acquisition by improving payback economics and creating more opportunities for repeat or expansion revenue, while poor acquisition quality can depress later retention.
• Investment decisions should use incremental contribution margin over a stated horizon, with uncertainty and capacity constraints, rather than attributed revenue or a single retention percentage.

I would decompose revenue into customer volume, retained activity, purchase or usage frequency, value per action, and expansion or contraction.

• Subscription revenue separates logos, seats or units, price, expansion, contraction, and churn, while commerce separates buyers, order frequency, basket value, returns, and margin.
• Each lifecycle program should map to a mechanism it can plausibly change, such as activation affecting early survival or replenishment affecting purchase timing, without claiming unrelated components.
• The decomposition must reconcile to financial totals and preserve cohort, tenure, product, and acquisition source so aggregate growth does not hide mix shifts.

Acquisition and retention are coupled because channel and offer choices change both who enters and how those customers behave later.

• A low CAC source can destroy value if it brings discount-dependent customers with weak margin or survival, while a higher CAC source may repay faster through durable behavior.
• Cohorts should be compared on contribution margin and retention at equal tenure, with acquisition cost and incentive treatment attached to the same customer definition.
• CRM can improve post-acquisition outcomes, but it should not be used to conceal structurally poor traffic by attributing natural survivor revenue to lifecycle contacts.

The next unit of budget should go to the option with the highest credible incremental contribution margin after risk and timing are considered.

• Acquisition evaluation includes marginal CAC, cohort quality, payback, and channel saturation, while retention includes reachable eligible volume, incremental lift, treatment cost, and future margin preserved.
• Average historical returns are insufficient because both channels have response curves, so allocation should use experiments or quasi-experiments near the current spend level.
• Cash constraints, inventory, service capacity, and strategic growth floors can override a pure return ranking, but those constraints should be explicit in the decision.

A defensible case links a specific retention mechanism to incremental margin, required capabilities, and a measurable counterfactual.

• Size the eligible population and baseline behavior, then model conservative lift ranges, margin timing, treatment cost, platform cost, and implementation cost rather than presenting influenced revenue.
• Separate reusable capability value, such as identity and decisioning, from the value of individual journeys so shared infrastructure is not counted repeatedly.
• Define the holdout design, maturity horizon, downside case, and stopping rule before funding, because the largest forecast is usually driven by the least certain long-term assumption.

Returns diminish as the program reaches less persuadable customers and adds contacts that overlap with existing treatments.

• Early journeys often fix obvious leakage, while later treatments compete for the same conversions, consume contact pressure, or subsidize behavior that would occur anyway.
• Product limits, customer intent, channel reach, consent, inventory, and service capacity create a ceiling that better messaging cannot remove.
• The portfolio should track marginal lift by added treatment and pressure level, then redirect investment toward product or acquisition when incremental margin falls below the alternative use of capital.

The portfolio should be organized around customer state transitions and economic jobs, not channel calendars or isolated campaigns.

• Map acquisition, activation, adoption, repeat, renewal, expansion, risk, lapse, and recovery states to owned outcomes, eligible populations, and valid treatment families.
• Distinguish always-on journeys, episodic interventions, and experiments so capacity, controls, and measurement horizons are visible at portfolio level.
• Review coverage gaps, treatment overlap, incremental margin, and complexity cost together, retiring journeys whose unique contribution no longer justifies another path.

A common state model should standardize durable business meaning while allowing product-specific evidence for each transition.

• Global states need mutually coherent semantics and stable transition contracts, while a purchase product and a usage product may use different qualifying events to prove activation or lapse.
• State ownership, effective time, version, and transition reason must be stored centrally so reporting and decisioning reproduce the state seen at the time.
• Local extensions are justified only when they change a material decision, because unrestricted local states recreate incompatible lifecycle definitions under a shared label.

Central decisioning should own comparable treatment selection and shared constraints, while source systems retain business truth and channels retain rendering mechanics.

• It should receive current state, eligible candidates, policy results, model outputs, cost, and context, then return a treatment ID, channel or surface, expiry, and reason code.
• It should not infer entitlement, consent, order status, or account ownership when authoritative services already own those facts.
• Channel tools may format and deliver the selected treatment, but they should not silently re-rank competing actions or implement separate global pressure rules.

A treatment taxonomy makes actions comparable, governable, and measurable across journeys, channels, brands, and vendors.

• Each treatment should carry a stable ID, objective, message class, offer economics, intrusiveness, eligibility dependencies, expiry, and expected customer action.
• Content variants can inherit one treatment identity when they express the same causal action, while materially different offers or channel sequences need separate identities.
• The taxonomy enables arbitration, exposure logging, cost allocation, and portfolio lift analysis without relying on vendor campaign names that change during migration.

Eligibility should be evaluated as ordered hard gates before any model or priority score can choose a treatment.

• Identity confidence, consent, purpose, region, entitlement, product availability, and safety suppressions form hard gates that expected value cannot override.
• Contact budget, active commitments, freshness, timing, and treatment expiry determine whether an otherwise lawful action is usable now or can be reconsidered later.
• Every gate needs a source, policy version, reason code, and safe behavior under missing data so destinations can prove why a customer was included or blocked.

Arbitration should optimize expected incremental customer value net of treatment cost, subject to hard customer and business constraints.

• Candidate values need a common horizon and unit, such as incremental contribution margin, instead of mixing click probability, churn risk, and attributed revenue.
• Urgent service needs and contractual commitments require explicit precedence because a short-term value score should not displace a payment warning or promised onboarding step.
• Losing candidates must be logged as expired, delayed, or reconsiderable so opportunity cost and selection bias remain visible in portfolio reporting.

Exploration should reserve controlled traffic for learning where decision uncertainty has economic value, without relaxing safety or consent.

• Randomization can occur among near-equal eligible treatments, under-served segments, or new actions, with probability and candidate set recorded in the decision log.
• The exploration budget should reflect downside risk, sample needs, and how quickly customer state changes, rather than apply one fixed percentage everywhere.
• Exploration data must remain separable from exploit traffic so uplift models can learn counterfactual response instead of reinforcing the policy's previous choices.

Global policy should define non-negotiable customer protections and shared semantics, while local policy may narrow or specialize valid actions.

• Consent enforcement, global opt-out, identity confidence, message classes, decision logging, and enterprise pressure ceilings belong in the common control layer.
• Regions and brands may add legal restrictions, quiet hours, channel availability, offer rules, or lower pressure limits, but should not broaden a global prohibition.
• Exceptions need scope, owner, legal or business basis, effective dates, and expiry so temporary local needs do not become permanent policy forks.

Scaled orchestration needs separate layers for customer truth, derived state, policy, decisioning, journey timing, channel delivery, and measurement.

• Operational systems own facts, the warehouse or profile layer derives history and features, and policy services determine whether candidate actions are allowed.
• Decisioning selects among candidates, journey orchestration manages sequence and waits, and channel platforms render and deliver without redefining upstream truth.
• Shared treatment IDs, state versions, exposure events, and conversion keys connect the layers so failures can degrade safely and results survive vendor boundaries.

A CDP should make governed customer data usable for activation, not become an unquestioned master for every customer fact.

• It can collect and normalize events, resolve approved identities, compute or distribute audiences, and provide activation-ready profiles at the required latency.
• Orders, entitlements, consent evidence, account ownership, and financial records should remain authoritative in their domain systems even when copied into the CDP.
• The design should be judged by lineage, identity quality, latency, replayability, destination controls, and portability rather than by the number of bundled vendor features.

Authority should be assigned per business domain and field contract, never to a single universal golden record.

• Identity, account, order, product usage, loyalty balance, consent, and service case domains each need an owner, authoritative key, update semantics, and accepted latency.
• Replicas and derived traits must carry source time, transformation version, and conflict behavior so a convenient downstream copy cannot silently overrule the owner.
• Cross-domain decisions should compose current facts at decision time and fail conservatively when a required authority is unavailable or contradictory.

Identity architecture should model these as related entities with explicit edges rather than collapse them into one customer profile.

• A person may use several devices and addresses, a household may share a device or purchase context, and an account or organization may contain several authorized people.
• Consent, communication address, entitlement, value, and outcomes must attach to the correct entity because household behavior cannot grant personal permission or account authority.
• Decisioning and experiments should declare their unit, such as person, account, or household, to prevent duplicate contacts, outcome leakage, and invalid randomization.

Resolution thresholds should follow the cost of false merges and missed links for each intended use, not one enterprise-wide match rate.

• Deterministic authenticated links suit personal communication and rights handling, while probabilistic household links may be acceptable for aggregate analysis or low-risk content.
• The graph should preserve source identifiers, edge reason, confidence, effective dates, and reversibility so contaminated features and exposures can be recalculated after a split.
• Quality reporting must include precision, coverage, duplicate-contact rate, false-merge incidents, and performance by region or identifier source, not only profiles unified.

Zero-party data should capture explicit customer choices, while first-party behavior should supply observed context without pretending to be a stated preference.

• Declared goals, topics, frequency, and channel choices are valuable when the product can honor and refresh them, otherwise collection creates expectation debt.
• Product use, purchases, and service interactions can improve timing or eligibility, but inferred interests should carry freshness and confidence and should not overwrite explicit choices.
• The data strategy should compare incremental decision value with collection friction, governance cost, sensitivity, and retention period before adding either type.