The direct answer: do not start with a demand-forecast model. First build an eligibility layer between the ERP or OMS, catalog, paid media, and CRM. For each SKU, variant, location, and market, it should answer whether the item can be promoted now, under which delivery promise, price, and contribution floor, and which fallback applies when the decision is stale. Rules handle hard constraints; forecasting and optimization come later. Measure spend and contacts on eligible items, cancellations, and contribution after service and fulfillment—with reach and conversion as counter-metrics.
The problem is rarely a binary stockout. An item may be physically on hand but committed to orders, reserved as safety stock, unavailable in one region, unsupported by installation capacity, or unprofitable after delivery. If each channel receives a different state, platforms keep buying clicks and lifecycle programs keep sending offers the operation cannot fulfill.
Write the Demand Eligibility Contract
- Unit — SKU, variant, location, market, and channel covered by the decision.
- Sellable state — available, committed, reserved, unavailable, incoming, backorder, and safety stock.
- Promise — delivery, pickup, installation, service capacity, and eligible geography.
- Economics — price, discount, contribution floor, and variable fulfillment cost.
- Freshness — update timestamp, maximum acceptable age, and expected refresh.
- Action — promote, reduce, redirect, substitute, wait, or suppress.
- Fallback — category, substitute, back-in-stock capture, or safe brand message.
- Outcome — eligible spend, sale, cancellation, delay, substitution, contribution, and experience.
- Owner — accountable people for source, rule, channel, incident, and scale.
The Contract prevents an opaque score from replacing an operational rule. A model may forecast replenishment or probability of sale, but it should not declare availability after the order system reserved the last units. Version each decision so the business can explain which state reached a channel and when it expired.
Connect five sources before optimizing
- Identity — consistent SKU and variant across PIM, catalog, OMS, storefront, CRM, and platforms.
- Inventory — available by location after commitments, reservations, and safety stock.
- Promise — delivery and capacity by ZIP code, region, service, or store.
- Economics — price, discount, media cost, shipping, and contribution floor.
- Demand — impression, click, message, cart, order, cancellation, substitution, and delay.
Shopify distinguishes available, committed, unavailable, on-hand, and incoming inventory. Google Merchant Center requires submitted availability to match the landing page and checkout. Together they expose a common mistake: publishing physical quantity as commercial availability. Promotional state must reflect reservations, location, promise, and policy—not only an inventory column.
Choose technology proportional to the decision
- Rule and feed — for one channel, few markets, and explicit constraints.
- Native update — useful as protection against price or availability mismatches, not as the source of truth.
- Integration layer — required when paid media, CRM, and storefront need the same decision.
- Forecast — useful when replenishment horizon and uncertainty change how much demand to create.
- Optimizer — only after products, markets, and budgets compete for the same capacity.
Do not confuse synchronization with intelligence. A fast feed can distribute the wrong decision; a sophisticated model can operate on stale data. Prove identity, latency, fallback, and reconciliation first. Then compare rules, native capabilities, and forecasts in shadow mode.
An eight-week implementation pilot
- Weeks 1–2 — choose one category, one paid channel, and one CRM journey; baseline ineligible spend, cancellations, delays, and contribution.
- Week 3 — reconcile SKU, variant, location, sellable inventory, promise, and economics.
- Week 4 — publish shadow decisions and measure divergence, data age, and coverage.
- Week 5 — activate hard rules, fallbacks, and logs in one channel without forecasting.
- Weeks 6–7 — release limited budget or contacts while preserving a market, product, or time comparator.
- Week 8 — decide whether integration and rules are enough or forecasting and optimization justify their complexity.
Measure demand the operation can fulfill
- Eligibility — share of spend, impressions, and contacts allocated to valid items.
- Freshness — state age, propagation delay, and channel-to-checkout mismatch.
- Sales — conversion, revenue, substitution, backorder, and cancellation.
- Economics — contribution after discount, media, delivery, service, and operational exceptions.
- Experience — missed promise, complaint, back-in-stock capture, and satisfaction.
- Operations — fallback, override, incident, recovery time, and manual effort.
- Counter-metrics — reach, new-product discovery, inventory velocity, and valid demand suppressed by conservative rules.
Fewer out-of-stock impressions do not prove incremental revenue. The campaign may lose reach or learning; inventory can change after the click; new products can be excluded for lacking history. Preserve a comparator and evaluate contribution, cancellation, and experience together.
Cost, accountability, and limits
Cost rises with SKUs, locations, countries, channels, refresh frequency, reservations, currencies, delivery promises, and continuous operations. Commerce owns catalog and policy; operations defines sellable inventory and capacity; marketing owns objective and fallback; CRM governs contact; data reconciles events; technology integrates ERP, OMS, PIM, and APIs; finance defines contribution; service returns failure causes. NIST's Core emphasizes governing, mapping, measuring, and managing risk throughout the lifecycle.
Bring a real feed to the diagnostic
Connect the project to product recommendations at https://makinai.co/insights/en/ai-product-recommendations-ecommerce-conversion-margin, return prevention at https://makinai.co/insights/en/ai-reduce-ecommerce-returns-protect-margin, paid media beyond platform automation at https://makinai.co/insights/en/paid-media-ai-beyond-google-meta-automation, and real-time versus batch at https://makinai.co/insights/en/real-time-ai-crm-when-batch-is-enough. MAKINAI implements CRM and commerce at https://makinai.co/services/en/crm-ecommerce-commerce-transformation. Bring one category, the current feed, and four weeks of orders, cancellations, and media data; we can define the smallest useful pilot.
The pilot does not guarantee revenue or contribution. Partner inventory can remain incomplete, platforms may cache feeds, and conservative rules can suppress valid demand. The sources support availability consistency, inventory states, and continuous controls; the Contract and pilot design are MAKINAI editorial recommendations to validate in the real operation.