Industries
AI for e-commerce and retail
Order and returns automation, catalogue enrichment and product Q&A connected to the OMS, PIM and helpdesk. Refunds above a set amount and bulk catalogue changes go to a person first.
The sector
What's different here
Retail support is dominated by a few repetitive questions: where is my order, how do I return this, can I swap the size. The data to answer them is structured and already sits in the OMS and carrier feeds, which makes this one of the easier places to put an agent to work.
The catalogue is the other big job. Supplier feeds arrive with missing attributes, inconsistent titles and no sizing data, and wrong attributes turn into returns and compliance claims. Models handle this normalisation well at volume, as long as a merchandiser approves before anything is published.
Shoppers are also arriving from AI assistants in growing numbers, though still a small share next to paid search and email. Complete, consistent product data serves your own assistant and the external ones equally.
1,300%
year-on-year rise in traffic to US retail sites from generative-AI sources over the 2024 holiday season (1 Nov to 31 Dec). Adobe notes this traffic is still modest next to paid search or email.[1]
693.4%
year-on-year rise in AI-referred traffic over the 2025 holiday season, as reported by Adobe.[2]
250 million
shoppers used Amazon's Rufus assistant in 2025, according to Andy Jassy, who said it is on track for more than $10B in annualised incremental sales. Company statements from an earnings call, via press coverage.[3]
Workflows
Where AI does useful work
Each of these runs inside systems the sector already uses. Steps that carry real risk wait for a person.
Order status and returns
An agent answers order questions from live OMS and carrier data and runs returns inside the written policy.
- 1Authenticate the customer by order number and email
- 2Look up the order in the OMS and the carrier's tracking
- 3Apply the returns policy to the item, date and condition
- 4Create the RMA or return label
- 5Escalate high-value refunds, suspected fraud and policy edge cases with a summary
A support lead approves refunds above the set threshold and any exception to the returns policy.
Commerce platform (Shopify Plus) · OMS · Carrier tracking APIs · Helpdesk (Gorgias, Zendesk)
Catalogue enrichment
Supplier feeds become complete, consistent product records that a merchandiser approves in bulk.
- 1Ingest supplier feeds and product images
- 2Generate or normalise titles, attributes and descriptions
- 3Validate against the taxonomy and the required attributes per category
- 4Flag claims that need evidence, such as materials or allergens
- 5A merchandiser approves in batches
- 6Publish to the PIM and storefront
A merchandiser approves each batch before it reaches the PIM, and safety and compliance attributes have a named owner.
PIM (Akeneo, Salsify) · Commerce platform (Adobe Commerce, commercetools) · Supplier feeds or EDI · DAM
Product Q&A on the storefront
A shopping assistant answers product questions from catalogue, review and Q&A data and links to the products it mentions.
- 1The shopper asks a question on a product or search page
- 2Retrieve catalogue attributes, reviews and existing Q&A
- 3Answer only from that data and link to the products
- 4Say so when the attribute is missing instead of guessing
- 5Log unanswered questions as catalogue gaps
No per-answer approval. The merchandising team reviews a weekly sample of answers and the gap log.
Commerce platform (Salesforce Commerce Cloud) · PIM · Reviews platform · Search index
Supplier and marketplace onboarding
Seller documents are read and checked, and the supplier record is prepared for the ERP.
- 1Collect registration, tax and compliance documents
- 2Extract the fields and check them against onboarding rules
- 3Verify bank details by call-back, never from email alone
- 4Prepare the supplier record for the ERP
- 5A buyer or finance user approves creation
Finance approves every new supplier and every bank-detail change.
ERP (NetSuite, SAP) · Supplier portal · Document capture · Sanctions screening
Demand planning explanations
Planners get the forecast with a plain-language account of what moved it.
- 1Run the existing ML forecast
- 2Pull drivers such as promotions, seasonality and stockouts
- 3Draft an explanation of the main changes per category
- 4The planner reviews and adjusts
- 5Purchase orders are raised only after planner approval
A planner approves every purchase order.
ERP · WMS · Forecasting model · Data warehouse (Snowflake, BigQuery)
Systems and rules
What the design has to respect
The systems we expect to connect to, and the regulation that decides what an AI system may do with the data in them.
- Commerce platform (Shopify Plus, Adobe Commerce, Salesforce Commerce Cloud, commercetools, BigCommerce)
- OMS
- WMS
- PIM (Akeneo, Salsify)
- ERP
- Helpdesk (Zendesk, Gorgias, Freshdesk)
- CDP and marketing automation
- Payments gateway
- Carrier tracking APIs
- PCI DSS v4.0.1
- The future-dated requirements, including 6.4.3 and 11.6.1 on payment-page scripts, became mandatory after 31 Mar 2025. We keep AI chat out of cardholder-data flows entirely.
- EU AI Act, Art. 50
- A chatbot must tell people they are talking to an AI. Under the Digital Omnibus schedule these transparency obligations apply from 2 Aug 2026.
- Consumer protection
- AI-generated product claims are still your claims. Misleading descriptions and personalised pricing fall under existing consumer law.
- GDPR and CCPA
- Personalisation and shopper profiles need a lawful basis, and customer messages sent to a model provider need a DPA.
- European Accessibility Act
- Applies from June 2025, and covers chat interfaces on the storefront as well as the pages around them.
Risks
What goes wrong, and how we design for it
- Wrong product attributes
- Required attributes are validated against the taxonomy, claims about materials, allergens and sizing need a source, and nothing publishes without merchandiser approval.
- The bot promising refunds outside policy
- Refunds go through an API with policy checks and amount limits, so the model can't commit to something the policy rules reject.
- Prompt injection through reviews or customer messages
- Retrieved text is treated as data, actions are allow-listed and every parameter is validated server-side.
- Peak-season load and cost
- Load-test at expected peak volume, cache common answers and set spend limits with a fallback to the standard helpdesk flow.
- Brand voice drift
- A style guide and a fixed example set are part of the evaluation, and the CX team reviews a weekly sample.
Where to start
Order status and returns, connected to the OMS and helpdesk
These enquiries are high-volume and repetitive, the data is structured, and deflection and CSAT are easy to measure. For retailers with large SKU counts and poor attribute data, catalogue enrichment is an equally good place to start.
- Connect the OMS, carrier tracking and the helpdesk, read-only at first
- Handle order status, return eligibility and return labels for one region or brand
- Encode the returns policy as rules the agent calls, not as prompt text
- Route refunds above the threshold, suspected fraud and policy exceptions to a support lead
- Measure containment alongside repeat contact within 7 days and CSAT for AI versus human
- Disclose the assistant as AI and keep a visible route to a person
In the sector
Publicly reported examples
What others in the industry have announced. These are not our clients.
Walmart catalogue data[4]
Walmart said it used multiple LLMs to create or improve more than 850 million pieces of catalogue data, work it said would otherwise have needed nearly 100 times the current headcount. Stated on an earnings call and reported by trade press.
Zalando shopping assistant[5]
About 6 million users in 2025. The company says customers add 13% more items to basket or wishlist from AI suggestions, and that size-related returns fell 8%. Company-attributed figures.
Use cases
Related use cases
- Customer support automationAn agent answers routine questions from your own policies and account data, takes a small set of allowed actions, and hands everything else to a person with the case already summarised.
- Document processingInvoices, forms and contracts are read, checked against your own records and posted to the right system, with low-confidence fields and failed checks sent to a person.
- Back-office workflowsReconciliations, approvals and onboarding are prepared by agents with the evidence attached, then routed to the right approver. Nothing is posted or paid without a person signing off.
- Questions over company dataManagers ask questions about sales, margin or operations in plain language and get figures from governed metric definitions, with the query shown so an analyst can check it.
Sources
- 1.Adobe, Traffic to US retail websites from generative AI sources jumps 1,200 percent (Mar 2025)
- 2.Adobe, Holiday shopping season results (Jan 2026)
- 3.Fortune, Amazon Rufus AI shopping assistant on track for $10 billion in sales (Nov 2025)
- 4.Retail Dive, Walmart generative AI product data points (Aug 2024)
- 5.FashionUnited, Zalando fully embraces AI: 90 percent of content now created by machines (Mar 2026)
Figures are as reported by each source on its publication date. Examples are public industry cases, not AI Sewa clients.
Next step