Retail AI is not one decision. It can rank products, predict demand, recognise an item, flag suspected theft, allocate a promotion or calculate a basket. Each action changes a different relationship between customer, colleague, stock record and price.
The useful question is therefore not whether a store is “AI powered”. It is whether a particular decision is accurate enough, fair enough and recoverable enough for the consequence it creates. A recommendation that can be ignored has a different risk from a camera match that triggers an intervention or a checkout system that charges a card.
This guide reflects UK regulator material available on 31 July 2026. Consumer and data-protection rules are UK-wide, while licensing, accessibility enforcement and some operational requirements can differ by nation and premises. Retailers should map the exact jurisdictions and services involved.
Map the Decision Before Selecting a Model
Describe each proposed use case as a chain from input to customer effect. This prevents a supplier demonstration from hiding the operational work.
| Use case | Input | AI output | Customer effect | Mandatory fallback |
|---|---|---|---|---|
| Product discovery | Query, catalogue and availability | Ranked products | Changes what is easy to find | Browse and search without personalisation |
| Personalised offer | Purchase and preference history | Offer eligibility or ranking | Changes the promotion shown | Published non-personalised offers and opt-out |
| Shelf monitoring | Images, planogram and stock events | Suspected gap or misplaced item | Creates a colleague task | Physical check before adjustment |
| Checkout-free basket | Camera, sensor and product events | Item and quantity attributed | Creates a payment charge | Itemised receipt, correction and human support |
| Loss-prevention alert | CCTV or biometric comparison | Risk or possible match | May prompt staff intervention | Trained review and non-confrontational escalation |
| Demand forecast | Sales, availability, events and lead time | Expected units by period | Changes order or labour plan | Approved planning baseline |
Name the person accountable for the output, who can override it and where the decision is recorded. Do not describe a customer-service agent as the reviewer if the interface gives them no evidence or practical authority to correct the result.
For higher-consequence decisions, apply the control pattern in human review for automated decisions.
Keep Product Discovery Commercially Honest
Recommendation systems should help a customer find a suitable, available product. They should not quietly optimise only for margin, supplier funding or stock the retailer wants to clear.
Define the ranking objective and its constraints. Separate relevance, availability, delivery promise, price, paid placement and business rules in the audit record. If sponsored results are included, label them where the customer sees the ranking. Test queries where no suitable product exists; the system should abstain or widen the search rather than fabricate a match.
The CMA’s March 2026 work on agentic AI and consumers makes clear that existing consumer law applies to algorithmic and agentic interfaces. Businesses remain responsible where an AI system misleads, pressures or steers customers unfairly. That is especially relevant when an agent can add an item, accept a substitute or transact rather than merely present information.
Evaluate discovery using more than click-through rate:
- successful search and zero-result rate;
- in-stock relevance at the customer’s location;
- returns attributed to unsuitable recommendations;
- paid-placement exposure and labelling;
- price and availability errors;
- customer corrections and complaints;
- performance for keyboard, screen-reader and low-vision journeys; and
- opt-out use and performance of the non-personalised route.
The related guide to AI product discovery in commerce covers retrieval and catalogue controls in more depth.
Make the Displayed Price the Payable Price
An AI offer engine does not create an exception to consumer law. The price and mandatory conditions presented at the decision point must remain clear when inventory, channel or customer context changes.
The CMA’s current price-transparency guidance, updated in January 2026, explains how mandatory fees, taxes and charges should be included and how drip or partitioned pricing can mislead. Build those requirements into the price service upstream of recommendation, signage and checkout; do not rely on generated text to reconstruct a lawful total later.
Maintain a price-integrity ledger containing:
- product and variant identifier;
- store, channel and effective time;
- base price and each mandatory component;
- promotion eligibility and evidence;
- displayed price at shelf, app and checkout;
- final charged amount;
- model or rule version involved; and
- correction, refund and complaint outcome.
Reconcile shelf-edge labels, electronic labels, online pages, loyalty offers and point of sale. A scale gate should require zero unresolved cases where the system charges more than the price the customer was led to expect. “The model changed” is not a customer remedy.
Do not infer vulnerability or urgency to increase a price. Test personalised offers for unjustified exclusion and proxy discrimination. Where the promotion is direct marketing, the ICO’s direct-marketing guidance requires transparent collection, an appropriate lawful basis and respect for objections. The right to object covers profiling related to direct marketing.
Design Checkout-Free as a Dispute-Ready Ledger
A checkout-free store combines identity or entry, product recognition, basket state, price and payment. The engineering target is not a cinematic walk-out. It is an itemised, contestable transaction.
Every basket event should retain the product, quantity, timestamp, confidence, sensor provenance and subsequent correction. Low-confidence or conflicting observations should enter review before charging, not be hidden in an unexplained total. Customers need a prompt receipt and a correction route that does not require proving the system wrong from memory.
An illustrative transaction flow is:
- entry credential creates a session without exposing unnecessary identity to the vision layer;
- shelf and product events update a provisional basket;
- contradictions remain marked as unresolved;
- the customer can inspect the basket before or immediately after exit;
- uncertain items are reviewed under a defined service time;
- payment is captured only under the approved confidence and dispute policy;
- corrections preserve the original event and reason; and
- repeated error patterns feed product, camera and layout changes.
Measure incorrect item, quantity and price rates separately. Track time to receipt, time to correction, first-contact resolution, refund completion and error rate by product form, store zone and accessibility route. A low average error can conceal repeated problems with loose produce, multipacks, mobility aids or customers shopping together.
Treat Personalisation as Optional Service Design
Purchase history can make a list more relevant without needing to infer a person’s mood, health, religion or financial position. Start with the least data that can deliver the stated benefit.
The ICO’s 2026-updated guidance on collecting and profiling for marketing says profiling must be fair, transparent, accurate and not excessive. It also highlights the additional risk of special-category data. A retailer should not infer pregnancy, illness, faith or hardship from a basket and then use that inference for promotion merely because a model can.
Separate these purposes in the data map:
- operating an account and fulfilling a purchase;
- remembering a customer-selected preference;
- recommending products within the current session;
- profiling across sessions;
- sending electronic marketing;
- measuring aggregate campaign performance; and
- training or evaluating the model.
Explain each active use in plain language and provide a functioning objection or opt-out. The generic route should remain usable, not deliberately degraded. Complete a Data Protection Impact Assessment where monitoring, biometric recognition, large-scale profiling or joined datasets create likely high risk. Our UK AI privacy guide provides the wider assessment pattern.
Put Accessibility Into the Store Workflow
Smart retail can remove barriers or create new ones. App-only entry can exclude a customer without a compatible device or data connection. A screen-only receipt can fail a blind customer. Gates, camera zones and narrow intervention points can make mobility harder.
Test the whole journey with disabled customers and staff:
- finding current product and price information;
- entering without a smartphone or biometric enrolment;
- requesting assistance without being profiled as suspicious;
- receiving an accessible itemised receipt;
- understanding and challenging a charge;
- using payment and loyalty without fine motor precision; and
- leaving safely during a system or power failure.
Keep a staffed route and reasonable-adjustment procedure. Accessibility is not demonstrated by a conformance statement for the app if the physical service still blocks the person.
Set a High Bar for Facial Recognition
Ordinary anonymous counting and facial recognition are not the same. Identification creates biometric data and a risk of wrongly associating a person with suspected wrongdoing.
In July 2026 the ICO published retail-crime guidance explaining that CCTV information can be used lawfully for crime prevention while emphasising the high bar for facial recognition in public places. A retailer considering it should document necessity, alternatives, watchlist provenance, match thresholds, human review, retention, notices, sharing and redress before deployment.
Never instruct colleagues that a model match proves theft or identity. Train for safe observation, de-escalation and escalation under store policy. Measure false matches, interventions without corroboration, demographic performance where lawfully testable, complaints, watchlist corrections and outcomes—not only the number of alerts.
Run a Controlled Store Pilot
Choose one decision and one operating environment. Capture the manual baseline before activating the tool. Then use shadow mode so the system produces outputs without changing a price, charge, offer or intervention.
Adjudicate errors against source records and physical checks. Test peak periods, product moves, promotions, groups shopping together, poor connectivity, camera obstruction and supplier outage. Run tabletop exercises for a disputed charge, false loss-prevention alert, inaccessible entrance and incorrect mandatory fee.
| Release gate | Evidence required |
|---|---|
| Purpose | Named owner, intended decision, prohibited uses and non-AI alternative |
| Data | Lawful basis, minimisation, DPIA where required, retention and supplier access |
| Shadow | Error types measured locally; no material group or product class below its floor |
| Limited live | Itemised evidence, trained reviewers, accessible support and rollback tested |
| Commercial | Displayed price reconciles to charge; sponsorship and offer logic are transparent |
| Scale | Complaints, corrections, opt-outs, alert load and customer outcomes remain within limits |
Before scaling, require all high-consequence outputs to be reviewable, every price discrepancy to have an owner and every automated charge to have a tested correction path. Stop the deployment if reviewers cannot keep up, evidence is unavailable, input conditions drift or the customer cannot use the fallback.
Retail AI earns trust when customers can understand the offer, receive the right item at the stated price and correct a mistake without fighting the machine. Convenience is the result of a well-controlled service, not a substitute for one.
