Do not use AI to find customers most likely to buy and give that group a discount. High-propensity customers often would have purchased without it. The more valuable starting point is incentive necessity: who changes because of an action, and whether that change pays for the benefit, channel, and operating cost. To protect contribution margin, make waiting and suppression compete with discounts, content, service recovery, and human outreach inside the same decision.
The commercial question is not how many people converted after receiving an offer. It is how many purchases, units of contribution, or retained periods occurred because of it. Click-attributed revenue mixes persuaded customers, customers who would buy anyway, and customers who merely moved a purchase forward or switched channels. A campaign can show strong conversion while destroying value by subsidizing natural behavior.
Define the Incentive Necessity Contract
- Outcome — purchase, renewal, upgrade, repurchase, or retention inside a declared window.
- Eligibility — customers who may receive the action without conflict, contact pressure, or commercial restriction.
- Natural baseline — expected probability and contribution without the new intervention.
- Competing actions — wait, suppress, inform, fix service, offer a non-price benefit, discount, or human contact.
- Margin floor — minimum contribution after product, fulfillment, benefit, media, service, and returns.
- Comparator — an eligible group that keeps the current process or receives no new incentive.
- Identity and window — decision unit, frequency, deduplication, and time to observe the result.
- Operations — owner, approval, capacity, exceptions, monitoring, and rollback.
The contract prevents a model from optimizing an isolated metric. If order-level contribution is unavailable, begin with conservative bands and rules. If identity is unreliable, operate on a smaller high-confidence cohort. If an exposed benefit cannot be withdrawn, limit the experiment before making the model sophisticated. Predictive accuracy cannot rescue an economically undefined decision.
Propensity, uplift, and incrementality answer different questions
Propensity estimates who is likely to buy. Uplift estimates who buys because of treatment. Incrementality compares the observed result with a counterfactual: what would have occurred without the action. A high-propensity, low-uplift customer should be suppressed from discounting. A moderate-risk customer with an addressable issue may respond better to guidance or service. A price-sensitive customer should receive an incentive only when expected incremental contribution remains positive.
Causal machine-learning methods can estimate different treatment effects across customers, but they require treatment history, outcomes, pre-decision context, and credible variation. Without an experiment or comparable prior policy, start with rules and randomized groups. Do not train only on past discount recipients; the model will reproduce the former policy rather than learn behavior without an incentive.
Create actions that do not depend only on price
- Wait — no message or benefit during the decision window.
- Inform — content about use, availability, evidence, comparison, or a real deadline.
- Resolve — fix fulfillment, onboarding, support, inventory, billing, or experience.
- Facilitate — reminder, saved checkout, replenishment, service, or preferred channel.
- Non-price benefit — priority, service, content, access, or convenience.
- Financial incentive — value capped by margin, frequency, expiration, and eligibility.
- Human outreach — for higher-value, complex, or sensitive cases.
AI should choose among actions only after lifecycle, finance, service, and marketing approve the library and limits. Many losses labeled price sensitivity are failures of service, understanding, or timing. Discounting masks the cause and trains customers to wait for the next offer. The best decision may be to send nothing.
An eight-week pilot
- Week 1 — select one journey; define outcome, window, contribution, and baseline; bound the population.
- Week 2 — reconcile identity, purchases, returns, benefits, contacts, and costs from before the decision.
- Week 3 — map current actions, exclusions, pressure, inventory, and rules; create a simple reference policy.
- Week 4 — reserve a stable comparator and randomize eligible actions; check cross-channel contamination.
- Week 5 — run in a controlled mode, record actual exposure, and block unauthorized combinations.
- Week 6 — reconcile purchase, contribution, returns, and contact; do not stop on an early spike.
- Week 7 — estimate average and prespecified segment effects; review false gains and adverse effects.
- Week 8 — retain the rule, expand the cohort, test uplift, or stop; document limits and operations.
Braze documents how a global control group compares users who receive messaging with withheld users and calculates uplift; it also warns that small or short tests weaken inference and that early stopping introduces bias. Klaviyo supports a related holdout mechanism. These functions help, but they do not reconcile contribution costs or remove overlap with paid media, sales, and other channels.
Measure incremental contribution, not attributed revenue
- Primary outcome — incremental contribution per eligible customer and action.
- Conversion — absolute difference, not only relative change.
- Unnecessary subsidy — incentives sent to customers who convert in the comparator.
- Pull-forward and cannibalization — accelerated purchases, product switching, or channel displacement.
- Experience — opt-out, complaint, promotional dependence, and contact frequency.
- Operations — correct exposure, collision, no-action cases, decision cost, and review.
- Reliability — size, balance, uncertainty, missing data, and market stability.
If treatment converts more but the difference represents only a few additional purchases while the incentive is paid on every treated purchase, attributed revenue rises while contribution falls. Include returns, fulfillment, service, and operating cost. Report results per eligible customer, not just per converter, so the cost of approaching people who did not change remains visible.
Costs, trade-offs, and ownership
Cost increases with markets, catalogs, variable margins, integrations, real time, identity, channels, actions, experiments, review, and monitoring. Rules are transparent and fit low volume. Native features reduce integration when the decision fits one platform. Custom uplift becomes defensible with volume, treatment variation, multiple actions, and capacity to operate the result. Lifecycle owns the decision; data and analytics own design and evaluation; finance owns contribution; engineering owns integration; service and product own nonpromotional causes; legal and privacy assess permitted use.
Adapt the program to the United States
U.S. programs may span ecommerce, stores, marketplaces, franchises, loyalty, and channel partners. Reconcile customer and order identity and account for returns, fulfillment, credits, and partner-funded promotions. Have counsel review state, federal, contractual, and sector requirements, especially when decisions affect price or eligibility. A score does not authorize discriminatory pricing. Keep explicit limits, review, explanation, and non-price alternatives.
Start with one journey and one margin floor
Compare post-lapse reactivation at https://makinai.co/insights/en/ai-customer-reactivation-crm-without-discounts, next-best action at https://makinai.co/insights/en/ai-next-best-action-crm-personalization, the experiment operating system at https://makinai.co/insights/en/ai-marketing-capacity-valid-experiment-operating-system, and preventable churn at https://makinai.co/insights/en/ai-preventable-churn-early-warning-retention. MAKINAI can design and implement the pilot in the current CRM: https://makinai.co/services/en/crm-ecommerce-commerce-transformation. Bring one journey, the current incentive, and the minimum contribution; the conversation starts by proving necessity, not buying another offer engine.
No model guarantees incremental contribution. Small samples, treatment contamination, seasonality, overlapping promotions, inventory changes, and incomplete identity limit inference. The sources document control groups and causal methods; the Incentive Necessity Contract and pilot are MAKINAI editorial recommendations to validate in the real operation.