Project guide / Ecommerce
Choose the store decision before the AI project.
A store can produce more descriptions, replies and recommendations without making shopping or operations better. Start with a specific decision: which product fact to publish, answer to send, item to show, stock exception to review, or message to prepare. Then follow that decision through corrections, customer experience and operating cost.
Start with one decision
Make the proposed change small enough to inspect.
Write a change ticket for the trial: exact product or customer context, source records, proposed output, owner, and acceptance measure. A useful ticket lets the reviewer compare the suggestion with the facts and see what would happen if it were released.
- Product facts precede wording
- A convincing description cannot establish material, dimensions, compatibility or a benefit. Missing attributes remain questions for the catalog owner.
- Eligibility precedes ranking
- An item can score well while being unavailable, incompatible or excluded from the channel. Check those conditions before ranking eligible choices.
- Accepted work precedes value claims
- A published change or sent reply is only an intermediate result. Include corrections, unresolved contacts, cancellations and returns in the trial.
Five project candidates
Match the task to the records you can trust.
These are bounded trial designs. Start with a reviewed sample and an unchanged comparison process; widen use only when the evidence supports it.
Prepare catalog changes for exact variants
01Store decision: Which wording can be published for this item and channel?
- Starting situation
- Verified product attributes exist, but writers repeatedly turn supplier material into storefront copy.
- Required records
- Exact SKU and variant, approved attributes, rights-cleared source material, market, locale and channel rules.
- AI contribution
- Draft descriptions and identify unsupported statements. Show each factual addition against its source rather than silently completing missing fields.
- Release authority
- The catalog owner resolves facts and approves the exact change. Channel publishing rules remain explicit, including required treatment of AI-generated content.
- Review artifact
- Before-and-after copy, variant identity, attribute references, unresolved questions and publication target.
- Failure to test
- A family-level feature is assigned to the wrong variant, or a generated benefit becomes an unsupported product claim.
- Trial measure
- Measure accepted descriptions, factual corrections, reviewer effort, feed rejection and later product-information complaints.
Draft support replies from current records
02Store decision: What can staff accurately tell this customer about this request?
- Starting situation
- Agents repeatedly locate order events and policy passages before composing an answer.
- Required records
- Authenticated customer context, exact order, fulfillment events, current policy version and permitted support records.
- AI contribution
- Retrieve relevant evidence and propose a reply with unresolved details marked. Separate recorded events from estimates and proposed remedies.
- Release authority
- Staff approve the response. Refunds, cancellations, address changes and exceptional promises use their own authorized workflows.
- Review artifact
- Source-linked draft, policy version, order event time and the action that remains for staff.
- Failure to test
- Another customer's order appears, an old shipment event is treated as current, or a draft promises a refund that was never authorized.
- Trial measure
- Compare incorrect promises, repeat contacts, unresolved cases, review time and customer-reported resolution.
Suggest an eligible merchandising shortlist
03Store decision: Which products deserve review for this query or collection?
- Starting situation
- Merchandisers can define a customer task and product constraints but spend time assembling candidates.
- Required records
- Catalog identities, query or collection purpose, compatibility evidence, current offers, channel eligibility and declared ranking rules.
- AI contribution
- Generate and rank candidates within the eligible set. Show the material reasons and any commercial placement rule that changes the order.
- Release authority
- Merchandising owners approve the objective, exclusions and presentation. Exact price, availability and compatibility remain source-controlled.
- Review artifact
- Included and excluded candidates, rule versions, displayed order and recorded exposure.
- Failure to test
- An attractive alternative does not fit the requested item, unavailable stock is promoted, or paid placement looks like an independent relevance judgment.
- Trial measure
- Measure constraint violations, useful selections, returns and complaints alongside commercial results. Use a suitable comparison before attributing uplift.
Explain inventory exceptions for a buyer
04Store decision: Which stock situation needs investigation before a purchasing decision?
- Starting situation
- Buyers already have a planning process but must reconcile sales, stock and supplier exceptions manually.
- Required records
- SKU-location history, available and reserved quantities, incoming supply, cancellations, returns, promotions and supplier lead-time records.
- AI contribution
- Summarize anomalies and possible explanations, with source references and uncertainty. Compare any forecast with the existing planning baseline.
- Release authority
- The buyer sets replenishment policy and approves orders. The model cannot create sellable stock, revise a received quantity or promise a delivery date.
- Review artifact
- An exception packet with the observation window, inventory states, assumptions and buyer disposition.
- Failure to test
- Stockout periods look like low demand, returned units count as sellable before inspection, or incoming supply is treated as already available.
- Trial measure
- Track false alarms, missed exceptions, buyer effort and forecast error by item group; review purchasing outcomes over the relevant lead-time window.
Prepare lifecycle messages for an eligible audience
05Store decision: Which approved message is appropriate for this defined segment?
- Starting situation
- Marketing teams have maintained permission records and approved offers but repeatedly prepare campaign variants.
- Required records
- Purpose-specific subscriber state, suppression records, permitted segment criteria, current offer terms and relevant order status.
- AI contribution
- Draft message variants from the approved brief. Preview audience inclusions and exclusions separately from the creative output.
- Release authority
- Marketing owners approve the audience, message and schedule; the sending system rechecks permission and suppression before release.
- Review artifact
- Segment definition, eligibility snapshot, approved creative, offer version and send receipt.
- Failure to test
- An opt-out after preview is ignored, a cancelled purchase triggers an inappropriate message, or personalization invents a sensitive personal inference.
- Trial measure
- Measure suppression failures, complaints, unsubscribes, creative correction and relevant downstream behavior rather than opens alone.
Before anything changes
Keep the source of authority visible at release.
A model output is a proposal. Each store action needs the records and owner appropriate to its consequences.
| Proposed change | Source to inspect | Cannot establish | Release owner | Hold the change when |
|---|---|---|---|---|
| Product wording | Exact variant attributes and approved claims | A missing feature or benefit | Catalog owner | A factual statement has no matching source |
| Customer answer | Authorized order events and current policy | A new promise or completed remedy | Support owner | Identity, event state or policy is unresolved |
| Displayed shortlist | Eligible items and explicit ranking rules | Compatibility or independent endorsement | Merchandising owner | A candidate violates a hard constraint |
| Replenishment proposal | Inventory states and planning assumptions | Received stock or purchasing approval | Buyer | Quantities or demand history cannot be reconciled |
| Lifecycle message | Current permission, suppression and offer | Consent or permission to send | Marketing owner | Eligibility changed after preparation |
| Commercial result | Exposure, costs and downstream outcomes | Causal uplift from correlation | Experiment and business owners | The comparison or observation window is inadequate |
Build a useful comparison
Follow one work item past the output.
Select a task with enough reliable examples to reveal mistakes and enough ownership to repair them.
- 01
Choose a narrow operating slice
Name a category, support topic, collection, SKU group or campaign type. Define the current process, permitted records and responsible owner.
- 02
Assemble representative cases
Include missing attributes, conflicting policies, empty inventory, new products, returns and changed permissions. Keep personal data limited to an approved purpose.
- 03
Evaluate the proposed work
Have reviewers check facts, constraints and omissions against source records. Record rejected outputs and correction effort, not just the best examples.
- 04
Test release and recovery
Exercise stale records, duplicate events, failed publishing and withdrawn approval. Verify that staff can identify the exact released version and correct its downstream uses.
- 05
Compare the downstream result
Define the observation window and comparison before the trial. Include operating cost, customer harms and delayed outcomes; separate workflow improvement from causal revenue claims.
Controls across the five trials
The store must remain correct when the suggestion is wrong.
These are engineering controls to assess against the actual platform and market. They do not establish legal compliance or commercial performance.
- Preserve identity and freshness
- Keep product, variant, offer, location, order and customer identifiers distinct. Recheck changing price, availability, policy and permission at the action boundary.
- Limit data and tool access
- Use the minimum authorized records. Treat customer messages and supplier text as untrusted content that cannot grant access or approve a store action.
- Make approvals specific
- Bind approval to the exact output, audience and destination. Separate drafting from publishing, sending, purchasing and refunding; prevent retries from duplicating actions.
- Keep correction paths open
- Record where a released change went. Support withdrawal, source correction and reprocessing, with a named owner for affected feeds, messages or customer cases.
Questions before choosing
Resolve the assumptions behind the demo.
The most suitable starting point depends on your records, review capacity and ability to observe outcomes.
- Which ecommerce AI project should come first?
- Choose the task with reliable input records, a clear reviewer and a measurable recurring problem. Catalog drafting may suit one store; support preparation may suit another. A neglected catalog or unreliable order history may need repair before either trial.
- Can generated product copy go straight into a feed?
- Review its factual content and the destination's rules first. Google Merchant Center requires its structured-title attribute and generative source value for AI-generated titles. Keep channel formatting and source metadata in the publishing process rather than relying on a writer to remember them.
- Does more inventory data make a forecast reliable?
- Not automatically. Stockouts, promotions, changing assortments and returns can distort the observations. Keep inventory states separate, compare with the current baseline and inspect error by relevant item groups before changing purchasing decisions.
- Is a customer segment permission to send marketing?
- Treat segment membership and send eligibility separately. Shopify documents subscriber filtering and suppression behavior, but the project must verify its actual sending path and applicable obligations. Recheck current eligibility after drafting and before sending.
- How do we know the project improved the business?
- First establish that accepted work improved after review and correction. For commercial claims, use a suitable comparison, exposure records and an observation window that includes relevant cancellations and returns. Report uncertainty rather than attributing every change to AI.
Source basis
Sources behind the control model.
- 01
Shopify
Automatically generating product descriptionsDocuments merchant responsibility for accuracy and the possibility of unsupported benefits in generated descriptions.
- 02
Google Merchant Center
Title and structured titleSpecifies structured-title submission and generative source labeling for AI-generated titles. These are channel-specific requirements.
- 03
Shopify
AI-generated suggested replies in Shopify InboxDescribes staff composing assistance and the need to review generated answers for accuracy before sending.
- 04
Shopify
Understanding inventory statesDistinguishes available, committed, unavailable, on-hand and incoming inventory. Informs the inventory-state boundary, not a forecast-performance claim.
- 05
Shopify
Collecting customer contact informationExplains marketing opt-in configuration and merchant responsibility for applicable requirements; settings do not replace legal assessment.
- 06
Shopify
Email subscriber list managementDocuments subscriber filtering, unsubscribe handling and suppression. Actual integrations and sending paths still need testing.
Start with one real workflow
A Systems Audit is the usual starting point. If the opportunity is already clear, we can move directly into a focused build.
Show Us the WorkflowStart with the free automation readiness checklistOBSERVEQUANTIFYDECIDEBUILD