Customer Understanding Needs Business Context.

Define value before choosing an action

Revenue, gross margin and contribution answer different questions. For a bounded campaign decision, a useful outcome may be contribution after the variable costs affected by the action: product cost, fulfilment, payment processing, incentives, returns and relevant service effort. Finance should agree which costs belong in that definition and which are already included elsewhere.

The definition also needs a horizon. Immediate contribution can improve while later returns or service costs erode it. Conversely, useful product guidance may create little immediate revenue but improve use and reduce avoidable enquiries. Keep those outcomes visible without converting every possible future benefit into a confident lifetime-value number.

For Maya, the retailer's current objective can be to resolve the delayed order and preserve the relationship within an agreed service commitment. The objective of a later promotion may be additional contribution. Keeping those decisions separate prevents a short-term sales target from displacing an unresolved obligation.

Bring the missing business evidence into view

A customer's interest identifies a possible opportunity. Product availability determines whether it can be fulfilled. Margin affects the room for an incentive. Delivery capacity affects what can be promised. Payment and return status affect the value already retained. Those facts often belong to different systems and change on different schedules.

Attach the evidence to the right level. Inventory can vary by product variant, location and fulfilment route; order-level margin may differ from a category average. A customer-level total cannot safely stand in for a missing line-item cost. A useful model of the business preserves those relationships instead of flattening everything into the customer profile.

A missing margin is not a zero margin, and an old stock count is not a promise that a product is available. Define freshness and fallback by decision. A promotion may require a current availability check before release; strategic planning can use a slower reconciled financial view. More frequent updates are useful only when the decision can use them.

More orders can still mean less contribution

Consider two illustrative groups of 100 eligible customers over the same observation window. One receives a discount offer; the other does not. For simplicity, each retained order has a value of 100 monetary units and contributes 25 before the campaign incentive and contact cost. The offer subtracts 10 for every treated order, including purchases that might have occurred without it.

The following arithmetic assumes equal group sizes, comparable retained orders and all other relevant variable costs already included in the 25-unit contribution. Discounts are subtracted once. In actual reporting, where sales are recorded net of discounts, they must not be subtracted again.

The offer group has four more retained orders and 400 more units of pre-discount order value, but 50 fewer units of contribution. That does not prove every discount is harmful. It shows why an order-count target and a contribution target can select different actions. These small invented counts do not establish statistical significance or justify a rollout.

The important cost is not only the incentive on the additional orders. The retailer also funds incentives on other treated purchases. A purchase-propensity score cannot identify which orders the incentive caused. Estimating that difference requires an appropriate comparison, rather than assigning every observed purchase to the campaign.

Choose among feasible actions

An economic objective needs boundaries. First establish what the company must do, what it is able to fulfil and which contacts are permitted. Only then compare discretionary alternatives. Required service, an approved refund or a contractual commitment should not depend on a customer's predicted profitability.

For Maya, feasible alternatives might be a necessary delivery update, service follow-up, or no additional promotional contact. After resolution, the set may expand to product guidance, a relevant offer without an incentive, a bounded incentive, or continued silence. Each option has prerequisites and costs. An unavailable product or an unresolved permission cannot be compensated for by a high response score.

An expected-value comparison can organize assumptions about the outcome under each action, its costs and the time horizon. It is not an oracle. Where evidence about treatment effects is weak, use a bounded test or a conservative policy and expose the uncertainty. The best prediction of who buys is not automatically the best estimate of whom an intervention helps.

Give each form of intelligence the right job

A maintained purchase count describes qualifying history. A churn estimate helps prioritize investigation. A channel recommendation helps compare available ways to reach someone. An inventory check constrains the feasible products. A decision policy combines these inputs under a business objective. An experiment evaluates whether a permissible intervention adds value. No one of these replaces the others.

Binoban's profiles, Computed Traits, audiences, Intelligence capabilities and Journey conditions offer building blocks for this relationship between evidence and action. Using financial or fulfilment context requires the relevant data connection and an agreed definition. It should not be presented as a universal profit optimizer merely because a journey can branch on a value.

Retail media adds another decision point. Relevance and campaign eligibility determine which candidates may participate; configured ranking, bids and performance evidence help compare them. Retailer revenue, advertiser outcomes and shopper usefulness remain different objectives. A stronger serving score does not on its own establish more incremental profit or a better long-term customer experience.

Measure effects after the order has had time to mature

Define the eligible population, assignment, outcome and observation window before interpreting a test. Where appropriate, random assignment can compare an offer with a permissible alternative or a holdout. Use the assigned groups for the primary comparison; restricting analysis to people who opened a message or purchased introduces selection that can distort the result.

Returns and cancellations need time to appear. Report an early estimate separately from a reconciled result, with a defined cutoff and a treatment for late updates. Track complaints, service demand and opt-outs alongside contribution. If capacity is shared, an intervention may also affect delivery for people outside the target group; the evaluation should account for that possibility.

Decide in advance what evidence would support expansion, revision or stopping. Include uncertainty around the estimated difference and the operational cost of applying the policy. A promising average may conceal a poor experience for a particular situation, such as customers already dealing with a delivery failure. Investigate those situations without treating every small subgroup fluctuation as a discovery.

Build a shared decision brief

Before automating the next offer, write down the decision, its owner, the eligible population, feasible alternatives, current evidence, cost definition and evaluation method. Name the unavailable data and the fallback. Specify what must be rechecked before execution and what event triggers reconsideration. Marketing, service, operations and finance should be able to read the same brief and recognize their responsibilities.

For Maya, that brief starts with resolving the order. Later it can support a promotion whose relevance, availability and economics have all been considered. Customer intelligence then becomes infrastructure for a business decision: evidence is connected to an objective, a choice is bounded by real constraints, and the result can challenge the original assumptions.