Questions d'entretien : CRM Manager

A lifecycle state model turns customer behavior into explicit states with controlled transitions between them.

• Each state needs observable entry and exit rules, such as activation within 14 days or lapse after two expected purchase cycles.
• Transition events should move the profile once and preserve transition time so journeys can react to both state and tenure.
• The model simplifies arbitration because treatments can target the current state instead of reimplementing conflicting stage logic in every campaign.

A segment-specific journey changes the treatment logic, not just the words shown to each customer.

• Segments may receive different goals, timing, channels, offers, and exits because their needs or expected value differ.
• The segmentation variable must predict a meaningful response difference, such as onboarding maturity or purchase cadence, rather than merely describe the profile.
• Shared journey components reduce duplication, while segment-level measurement shows whether added complexity produces incremental value.

Next-best-action selects the most valuable eligible action for a customer at a particular decision point.

• Inputs can combine lifecycle state, recent behavior, predicted response, customer value, channel availability, and real-time context.
• Candidate actions are filtered by consent, product eligibility, contact policy, inventory, and active journey commitments before scoring.
• The selected action should optimize a defined outcome with an exploration or control mechanism, because the highest historical response is not necessarily incremental.

A decision engine evaluates eligible treatments and returns a consistent action, channel, or content choice at decision time.

• It combines rules, model scores, priorities, and context behind a stable interface that journey tools or product surfaces can call.
• Central decisioning prevents each channel from applying different consent, eligibility, or prioritization logic to the same customer.
• Its outputs need reason codes, model and rule versions, and latency limits so decisions remain auditable and usable in real-time surfaces.

A contact policy defines which communications may reach a customer, how often, and under which exceptions.

• It separates mandatory service messages from marketing and sets channel-specific quiet hours, recency rules, and rolling frequency caps.
• It includes legal suppressions, customer preferences, fatigue controls, and protections for sensitive lifecycle moments such as complaints or cancellations.
• Enforcement should happen at final decision time across all sending systems, because campaign-level checks alone miss concurrent contacts.

Treatment arbitration chooses one compatible treatment when several campaigns or journeys want to contact the same customer.

• Hard constraints remove unlawful or ineligible options before soft priorities compare urgency, expected incremental value, and customer need.
• The arbitration unit can be a message, offer, or journey slot, but every candidate must use a comparable score or explicit precedence.
• A losing treatment may be delayed, skipped, or reconsidered later, and that disposition must be logged to avoid hidden selection bias in reporting.

Lifecycle transitions describe durable customer state, while campaign entry rules describe temporary treatment eligibility.

• A customer can become activated once but qualify for several activation-related treatments under different channel and timing conditions.
• Keeping state logic upstream gives analytics and every activation tool the same definition instead of duplicating it across journey canvases.
• Campaign rules can then change without rewriting customer history, while transition versions preserve comparability when the business definition evolves.

A predictive churn score ranks customers by estimated loss risk within a stated future window.

• Features can include usage decline, failed payments, tenure, support interactions, and purchase recency, all calculated before the prediction cutoff.
• Calibration and lift by score band matter more than raw accuracy when churn is rare and intervention capacity is limited.
• The score identifies risk, not persuadability, so a randomized test is still needed to learn whether a retention treatment changes outcomes.

A propensity model predicts an outcome, while an uplift model predicts how treatment changes that outcome.

• Purchase propensity often ranks customers who would buy anyway, which can raise observed conversion without creating incremental sales.
• Uplift estimation needs treated and untreated examples to distinguish persuadable customers from sure things and customers harmed by contact.
• Uplift is harder to train because treatment assignment and sample size must support causal comparison, so simpler randomized rules may outperform a weak model.

RFM can be extended with dimensions that capture margin, cadence, channel behavior, or product relationship beyond past spend.

• Expected interpurchase time makes recency relative to each category or customer instead of applying one lapse threshold to all buyers.
• Margin, returns, discount dependence, and category breadth can separate profitable loyalty from frequent but costly purchasing.
• Extra dimensions increase sparsity and maintenance cost, so each addition should change a treatment decision and be validated against future behavior.

Feature freshness determines whether a segment or score reflects the customer at the moment a treatment is chosen.

• Fast-decaying signals such as cart activity may need minute-level updates, while tenure or historical value can tolerate daily refreshes.
• Every feature needs an event-time cutoff, last-updated timestamp, and stale-value behavior so old intent is not mistaken for current intent.
• Fresher pipelines cost more and can be noisier, so update frequency should follow decision latency and the cost of a wrong contact.

Cohort LTV estimates cumulative or expected customer value for groups sharing a meaningful start period or acquisition context.

• Revenue should be adjusted for margin, refunds, servicing costs, and a consistent observation horizon when the decision concerns economic value.
• Mature cohorts provide more observed value, while young cohorts require survival or retention assumptions and should carry wider uncertainty.
• Comparing cohorts is useful only after controlling for customer age, acquisition mix, product changes, and discounting conventions.

Predictive segments should be validated for ranking quality, stability, actionability, and treatment impact.

• Backtests must preserve time order and compare score bands on future outcomes to avoid training on information unavailable at decision time.
• Segment size, score drift, missing-feature rates, and performance across key customer groups reveal whether deployment remains reliable.
• A randomized treatment test confirms business value because a stable high-risk or high-propensity segment may still be unresponsive to CRM contact.

A CDP can collect customer events, resolve identities, build profiles and audiences, and distribute data to activation tools.

• Collection normalizes inputs from product, commerce, service, and marketing systems into governed schemas.
• Profile and audience layers make linked history usable, while destinations send traits, events, or segment membership downstream.
• A CDP is not automatically the source of truth for orders, consent, or account status, because authoritative ownership can remain in operational systems.

A source of truth should be assigned per data domain, not declared for the entire customer profile.

• The commerce platform may own order status, the account service may own plan state, and a consent ledger may own marketing permissions.
• Downstream copies need ownership, update direction, conflict rules, and acceptable latency so a convenient field does not silently become authoritative.
• Derived fields such as lifecycle state should document their inputs and calculation version rather than overwrite raw source facts.

Deterministic matching uses verified shared identifiers, while probabilistic matching estimates that records belong to the same person.

• Account ID, verified email, or a trusted login link can support deterministic rules when identifier reuse and household sharing are handled.
• Probabilistic methods combine weaker signals such as device, location, and behavior into a confidence score rather than a certain link.
• Higher match rates increase personalization coverage but also raise false-merge risk, so sensitive activation should require stricter evidence than aggregate analysis.

Profile unification requires field-level conflict rules, history preservation, and consent-safe activation in addition to identity links.

• Each attribute needs a winning source or recency rule, because the latest received value is not always the latest business truth.
• Events should retain original identifiers, source, and timestamps so merged history remains traceable and can be separated after a bad link.
• Consent and regional restrictions should not be broadened during a merge, and activation must use the applicable permission for the chosen identity and channel.

A durable event schema captures who acted, what happened, when it happened, and the business context needed downstream.

• It needs a stable event name, event ID, subject identifiers, event-time timestamp, source, schema version, and typed properties.
• Names should represent completed facts such as order_completed rather than UI interactions whose meaning changes with implementation.
• Required fields, allowed values, ownership, and compatibility rules belong in a data contract so producers cannot silently break journeys and metrics.

Event time says when customer behavior occurred, while processing time says when the platform received or handled it.

• Journey windows, sequence logic, and attribution should usually use event time so network or batch delays do not rewrite customer history.
• Processing time is essential for monitoring latency, scheduling safe waits, and deciding whether an event is too late to trigger a message.
• Pipelines need a lateness policy because unlimited backfill can start obsolete journeys, while rejecting all late events corrupts history and analysis.

CRM activation depends on completeness, validity, accuracy, consistency, uniqueness, and timeliness at the decision boundary.

• Completeness asks whether required identifiers and properties exist, while validity checks formats, ranges, and allowed values.
• Accuracy and consistency compare business truth across sources, and uniqueness controls duplicate profiles or events.
• Timeliness must be measured against the use case because a daily product feed may suit reporting but fail a real-time abandonment journey.