Data Academy · Tutorial 7 of 10 · Retail & consumer goods

Governed AI in retail & consumer goods

Retailers and brands are moving AI from merchandising dashboards into the decisions themselves: when to mark down, which returns to question, which retailer deductions to accept. Whether those agents scale depends on whether every price change, refund and write-off can be traced to an agreed rule, checked against consumer law and approved by the person who owns the margin.

01 What is changing

Where AI in retail & consumer goods is heading

From copilots that summarise sales to agents that propose and execute markdowns, transfers and claims inside the merchandising and finance systems.
From blanket seasonal discounts to the minimum markdown per product and location that clears stock by the end of the season, inside an approved price ladder.
From returns as a customer-service cost to returns as a risk domain, as fabricated damage photos and serial returning become harder to spot by eye.
From isolated pilots in one category to a few connected domains where pricing, inventory and finance read the same product, store and promotion data.

02 Use cases

Three use cases on the governed path

Each use case runs the same path: a business question, governed context, a deterministic rule, specialist agents, a policy check, an action and a record. How the path works →

Illustrative: names and figures are invented to show the flow.

Use case 1

End-of-season markdown within the price ladder

Seasonal lines left at full price too long end as deep clearance or write-off; marking down too early gives away margin. The lever is the smallest markdown that clears stock by season end, applied consistently and lawfully.

“Which summer lines are on track to be left over at season end, and what is the smallest markdown that clears them?”Asked by a merchandise planner

Context: Merchandising system: linen shirt range (style LS-204), full price £35, 3,600 units across 140 stores and the e-commerce DC
POS and web orders: selling about 300 units a week over the last four weeks
Season calendar: 6 weeks left before the autumn range lands, so cover is about 12 weeks against 6 remaining
Price history: the line has sat at £35 for the whole of the last 30 days; margin floor for first markdown is £26
Rule: If weeks of cover exceed weeks left in the season by more than 4, apply the first step of the approved markdown ladder (£35 to £28), provided the new price is above the margin floor and the reference price shown is the lowest price of the previous 30 days.
Decision: Apply first markdown step Severity: Medium
Agents: Demand agent Projects sell-through at £28 by store cluster and channel so the planner sees whether one step is likely to be enough.
Price-compliance agent Checks the reference price against the price history so the was-price shown on shelf and online meets the price indication rules.
Explanation agent Writes the plain-language rationale for the planner and store teams: cover, weeks left, margin floor and the step applied.
Policy: Planners may apply the first ladder step without further sign-off; any deeper markdown, or one below the margin floor, needs the category director. Reference prices must follow the EU Price Indication Directive (as amended by the Omnibus Directive) or the UK pricing rules for each market.
Action: Price change for LS-204 released automatically within policy to the pricing engine, POS and web catalogue, effective at next store open, with shelf-edge label job raised for the 140 stores.
Data: Merchandising and product hierarchyPOS and e-commerce salesInventory by store and DCPrice history and markdown ladderSeason and range calendar

Use case 2

High-value return claim review

Generated damage photos and serial returning make it harder to separate genuine claims from abuse, while wrongly refusing a genuine customer damages trust. The lever is routing only the risky claims to a person, with the evidence assembled.

“Should we refund this damaged-on-arrival claim straight away, or does it need a closer look?”Asked by a customer service team lead

Context: OMS: order 7781-QX, espresso machine at £640, delivered on 12 May by carrier with signature
Returns management system: damage claim opened on 14 May with three photos, refund requested to original card
Customer ledger: 9 returns in the last 6 months, totalling £2,900, 4 of them marked damaged on arrival
Evidence check: image metadata on two photos shows a creation date of 9 May, before the delivery date
Rule: If the claim value is above £250 and either the customer has 6 or more returns in 6 months or any evidence check fails, the refund is not issued automatically and the case goes to a returns specialist.
Decision: Refer claim to returns specialist Severity: High
Agents: Evidence agent Reads the photo metadata, carrier proof of delivery and warehouse pack record and lists what agrees and what does not.
History agent Summarises the customer's past orders, returns and outcomes without labelling the customer as fraudulent.
Explanation agent Drafts the case note for the specialist and a neutral holding message to the customer.
Policy: Only a returns specialist may refuse or partially refund a claim; no refusal is made solely by automated means (GDPR Article 22). Statutory rights under the Consumer Rights Act 2015 or the EU Consumer Rights Directive are applied before any store policy.
Action: Return case in the returns management system set to referred to a human specialist, refund not issued, holding message sent to the customer; no change to the customer account.
Data: Order management (OMS)Returns managementCustomer and payment ledgerCarrier proof of deliveryWarehouse pack and dispatch records

Use case 3

Retailer deduction validation for a consumer goods brand

Retailers short-pay invoices for promotions, shortages and penalties, and unchecked deductions become silent margin loss. The lever is matching every deduction to the agreed trade terms and proof of performance before it is accepted or disputed.

“Is this promotional deduction from our largest grocery customer valid, or should we dispute it?”Asked by a deductions analyst

Context: ERP accounts receivable: invoice 90044218 for €61,500 paid as €43,100, a deduction of €18,400 coded as feature and display
Trade promotion management (TPM): agreed fund for that feature and display, weeks 18 to 23, is €12,000
Proof of performance: display photos and retailer scan data present for 4 of the 6 promoted weeks
Remittance advice from the retailer Hollin Grocers references promotion ID P-2291 and no other agreement
Rule: If a deduction exceeds the agreed fund in TPM for the referenced promotion, or proof of performance is missing for any promoted week, the deduction is not cleared and is blocked pending a proof-of-performance check.
Decision: Block deduction pending proof of performance Severity: High
Agents: Matching agent Matches the remittance line to the promotion, fund, invoice and any other open agreements with the retailer.
Evidence agent Lists which weeks have display photos and scan data and which do not.
Explanation agent Drafts the dispute summary for the key account manager: the €6,400 above the fund and the two weeks without evidence.
Policy: Analysts may clear deductions that match TPM and proof of performance; write-offs and disputes sent to a retailer need the key account manager's approval. Supplier-retailer terms follow the Groceries Supply Code of Practice in the UK or the EU Unfair Trading Practices Directive where they apply.
Action: Deduction item in ERP set to blocked pending proof-of-performance check; draft dispute held for key account manager approval before anything is sent to the retailer.
Data: ERP accounts receivable and remittancesTrade promotion managementRetailer scan and display evidenceCustomer contracts and trade termsInvoice and delivery records

03 The foundation

What the agents need to understand

Core entities in the ontology

ProductSKUStoreChannelCustomerOrderReturnPricePromotionSupplier

Systems they come from

Merchandising system: product hierarchy, ranges, seasons, open-to-buy and markdown ladders
POS and e-commerce platform: transactions by store, channel, basket and time
Order management (OMS): orders, fulfilment routes, returns and refunds
Warehouse management (WMS): stock by DC and location, pick, pack and dispatch records
Trade promotion management (TPM): promotion calendars, funds, accruals and claims with retailers
ERP finance: invoices, receivables, deductions, payables and margin

04 Guardrails

The controls that let it scale

Lawful reference prices

Any was-price shown on a markdown is checked against price history to meet the Omnibus Directive or local pricing rules before release.

Margin floor

No automated price change goes below the agreed margin floor for the line; deeper cuts need the category director.

Human decision on refusals

Refunds can be issued automatically, but refusals and account restrictions are made by a person, in line with GDPR Article 22.

Customer data minimisation

Agents see the return history needed for the case, not full profiles, and no fraud labels are stored without review.

Trade terms first

Deductions and supplier claims are checked against signed terms and applicable codes of practice before money moves.

05 Rollout

From the first use case to many

1
Pick one margin decision
Start with seasonal markdown in one category where the ladder and margin floor are already agreed.
2
Model the business once
Define product, store, price, promotion and return in the ontology so pricing, returns and finance read the same entities.
3
Run in recommend mode
Let agents propose markdowns and referrals for a season while planners approve each one, and compare outcomes.
4
Release within policy
Allow the first ladder step and low-value refunds to release automatically; keep deeper cuts, refusals and write-offs with people.
5
Extend to connected domains
Add returns risk and deduction validation on the same model, with one audit trail across them.

06 What to measure

Outcomes, not activity

Full-price sell-throughEnd-of-season residual stockMarkdown margin given awayRefund cycle time for genuine claimsDeduction recoveryDecisions overturned on review

07 Pitfalls

What usually goes wrong

Automating a broken calendar. If season dates and ladders differ by team, an agent applies the wrong rule faster; agree them once in the business model first.
Pricing without compliance. Markdowns released without a price-history check create unlawful was-prices; make the reference-price test part of the rule.
Treating customers as suspects. Scoring every return alienates honest customers; route only claims that meet a clear rule and keep the decision with a person.
Fragmented product data. Different SKU codes in POS, OMS and TPM break matching; resolve product identity before scaling any agent.

08 Diagnostics

Questions to ask your team

1
Which price, refund and deduction decisions are we comfortable releasing automatically, and who signs off the rest?
2
Do pricing, returns and finance use the same product, store and customer identifiers?
3
Can we show, for any markdown, the rule, the data and the reference price behind it?
4
Where would a wrongly refused refund or a wrongly accepted deduction be caught today?

09 Keep going

Frequently asked

Will agents set prices without merchants?

Only inside the ladder and margin floor that merchants have approved. Anything outside it is proposed and waits for the category director, and every change is recorded with its reason.

Can AI decide that a return is fraudulent?

It can assemble the evidence and route the case, but the refusal is a human decision. Statutory consumer rights are applied first, and the customer is not labelled without review.

Does this suit consumer goods brands as well as retailers?

Yes. Brands use the same governed path for retailer deductions, promotion claims and order allocation, reading trade terms from TPM and ERP rather than from email.