The vendors and deployments behind shopping advice, store navigation and retail operations, with a closer look at what it takes to make them work.
Who builds the assistants, what they can do, and what retailers and shopping centres need before deploying them.
Walmart wants an AI-assisted shopper to finish an order. Westfield La Maquinista wants a visitor to find the right shop. Lowe’s gives its associates an assistant so they can answer questions on the sales floor.
The software may look similar from the customer side. Behind it sit different records, permissions and systems. A mall directory cannot promise that a shoe size is in stock; a product catalog cannot calculate an accessible route through a building.
The useful way to evaluate a retail AI agent is to follow the task it can complete, the data it can access and the result it can prove. This guide separates retailer-owned assistants, products available to other businesses, and the mapping, catalog and payment services behind them.
It reviews published implementations and product documentation, not hands-on benchmarks. A named deployment, a vendor’s product announcement and a planned feature are identified separately. Reported performance figures retain their original meaning.
The term covers several levels of capability. An assistant may answer from a product catalog. A more connected system can check an order, request directions, or prepare a cart. Another may submit an authorized purchase or carry out a staff workflow.
For this guide, the practical distinction is whether the system can move beyond a plausible answer to a verifiable action. The action need not be a payment. Retrieving the correct order status or generating a usable route can be valuable without giving software permission to spend money. A mall concierge that recommends a shoe retailer can’t necessarily check a particular size. A product assistant that finds that size online may know nothing about the store downstairs. Those are separate integrations, not small differences in wording.
| Task | Typical user | What the system needs | A useful completion event |
|---|---|---|---|
| Find and compare products | Shopper | Catalog, attributes, prices and merchant rules | A relevant, available shortlist |
| Find a shop or service | Mall visitor | Current tenants, hours, units and routable maps | Correct directions to an open destination |
| Complete or service an order | Shopper or support team | Account, cart, order and payment access | Confirmed order or resolved service request |
| Help on the sales floor | Store associate | Product guidance and store-specific data | A verified answer during the customer interaction |
| Support property operations | Mall or store manager | Sales, maintenance and operational records | A checked analysis or an approved action |
Sparky helps customers search Walmart’s assortment, compare options and understand product reviews. The assistant is part of Walmart’s shopping experience. It is not a general-purpose concierge that another shopping centre can install. Walmart also has an app inside ChatGPT. The published integration supports linked accounts, loyalty and Walmart payments. That distinction matters: a customer can discover products in an external assistant while the retailer supplies the account and transaction experience.
For another retailer, the lesson is architectural rather than a supplier recommendation. The conversational entry point and the checkout do not have to belong to the same company. What matters is whether the customer can move between them without losing the order or the terms offered.
Amazon’s current consumer-facing name is Alexa for Shopping, combining Rufus and Alexa+ capabilities. Older guides that list Rufus as a separate current product need that correction.
Amazon describes product comparisons, personalized guidance, price monitoring, cart-building, and scheduled shopping actions across its US shopping surfaces. Some functions prepare a purchase; others can execute eligible purchases under customer instructions. Eligibility and permissions matter more than the broad label “autonomous shopping.”
This is another retailer-owned environment, not neutral infrastructure for mall operators. It is useful to brands as a discovery channel and to product teams as an example of combining advice with account history and purchasing tools.
The supplier question is different from the visibility question. A brand may need to make its assortment understandable to Amazon’s assistant without being able to buy or control that assistant.
At Westfield La Maquinista in Barcelona, the assistant West helps visitors find shops, restaurants and services and generates routes through the centre. The published pilot uses QR codes as an entry point. Eleva AI, the developer, reports that approximately one in three conversations ends in a route to a specific store. That is a navigation output. It does not establish that the visitor followed the route, entered the shop or bought anything.
The case nevertheless gives a mall operator a concrete starting point. The assistant combines a visitor’s stated need with a destination in the property. The valuable input is not simply a collection of brand descriptions. It is the current tenant list, unit locations, and a map that can support directions.
Eleva also identifies Gavi at La Gavia as a deployment of its shopping-centre product. Gavi is promoted on La Gavia’s official Klepierre website. That provides a second named centre rather than a claim based only on a supplier’s logo wall.
These implementations support a case for testing tenant discovery and navigation. They do not yet provide a common, independently audited benchmark for additional tenant sales.
Google’s store-calling feature lets eligible US searches in Search or AI Mode trigger calls to nearby businesses about product availability. Google returns the findings to the shopper.
This is relevant to malls even when the operator has no assistant of its own. The recommendation may be made elsewhere, using information obtained from a tenant.
A call can resolve an availability question that a directory cannot. It is not a reservation, however, and a positive answer does not guarantee that the item remains on the shelf until arrival. The store’s identity, the time of the check, and the scope of the answer should travel with the recommendation.
Mappedin’s AI Navigator implementation at YYC Calgary International Airport combines conversational assistance with spatial information. Satisfi Labs supplies the conversational platform; Mappedin supplies the mapping component.
The named implementation is an airport, not a mall. Its relevance is the separation of jobs: understanding a question is different from calculating a route through a building.
That distinction is useful when evaluating mall proposals. A map image, a searchable tenant directory and a routable indoor map are not interchangeable. Accessible routes, floor changes and blocked passages require spatial data and rules. An AI model should not invent a path from the appearance of a floor plan.
Nor does a mall’s use of a mapping supplier establish that it has deployed that supplier’s AI product. The venue, module and operating status need to be checked separately.
Constructor’s AI Shopping Agent is designed for conversational product discovery on a retailer’s own digital properties. It works with the merchant’s assortment and merchandising environment rather than recommending from an unrestricted view of the market.
Its Agentic Checkout announcement with Stripe, dated October 5, 2026, extends that conversation toward payment on retailer-owned sites and apps.
This is a vendor-announced product capability. The announcement should not be treated as proof that every retailer using Constructor has enabled checkout or achieved an incremental sales gain.
The practical evaluation is whether a real customer’s constraints survive the entire journey: budget, size, compatibility, stock and delivery. A persuasive recommendation is of limited value if the cart cannot fulfill it.
Algolia’s shopping-assistant offering and Agent Studio provide tools for building AI-assisted search and product discovery. The relevant foundation is the merchant’s indexed product data, with controls over retrieval and merchandising. It is a build platform, not a finished mall concierge. A team still needs to define its data sources, allowed actions, interface and escalation path. Agent Studio’s product documentation and changelog describe controls such as tools, usage limits and guardrails; their presence does not remove the need to test behavior.
A retailer already using Algolia should investigate how much of its catalog and search setup can be reused. A mall starting with a tenant directory should not assume the same integration gives it stock information from every tenant.
Pandora’s Agentforce implementation makes the distinction between tasks unusually clear.
Clara handles customer-service questions such as order status and jewellery care. It draws on connected commerce and order systems, including IBM Sterling, and passes unsupported requests to a person. Salesforce reports 60% autonomous case deflection. That is a support outcome, not an increase in product sales.
Gemma is a separate personal-shopping agent, described as being tested in selected markets. It recommends jewellery around the occasion, recipient and budget. Its testing status should not be confused with Clara’s reported service performance.
Salesforce Professional Services and Publicis Sapient are named implementation partners. The useful lesson is that the model, commerce platform, operational data and implementation work are distinct parts of the deployment. “Uses Agentforce” alone does not tell another retailer how much of that work it would need to repeat.
Lowe’s rolled out Mylow Companion to associates across more than 1,700 stores in May 2025. Developed with OpenAI, the assistant runs on staff devices and supplies product information, project guidance and inventory information.
It shares a foundation with customer-facing Mylow, but the person using it is different. An associate remains in the conversation with the shopper and can interpret the answer in context.
This is an important alternative to putting another chatbot on a consumer website. Some physical-retail tasks are better served by helping the person already standing beside the customer. Evaluation should focus on answer accuracy, time saved and the quality of the completed interaction, not just usage.
HyperIn’s Manage AI Assistant is aimed at property teams. Its published use cases include asking about performance, preparing a property summary for a tenant meeting and checking operational information without assembling reports manually.
This is not the same product as a visitor-facing concierge. In the source reviewed, HyperIn described its AI Mall Concierge as still in development. The documented management assistant and the future shopper product should not appear as one fully deployed offering.
For a mall team, the first question is what underlying records the assistant can read. A fluent answer about tenant performance is only useful if the sales periods, occupancy changes and comparison rules are visible and correct.
Fujitsu and AEON Food Style announced a field trial in a physical store, applying agents to store strategy and shelf-allocation planning. Inputs included product information, head-office guidance and conditions at the store.
A separate Fujitsu announcement on October 6, 2026 introduced a trial environment for four retail agents covering sales analysis, loyalty analysis, merchandise planning and store-manager support. Trials with seven retail companies were planned in phases; commercial launch was targeted for June 2027.
Those are trial and roadmap statements, not evidence of seven completed rollouts. The AEON project also does not establish that software independently changes store layouts. Planning assistance and permission to execute a plan are different responsibilities.
Shopify’s Agentic Storefronts connect merchant catalogs with external AI shopping channels. Shopify also offers an Agentic plan for brands that want catalog distribution without moving their entire commerce business onto Shopify.
The purchase flow varies by channel. Shopify’s published ChatGPT flow uses an in-app browser on mobile and a merchant-site link on desktop. The merchant remains responsible for the sale and fulfillment.
This is commerce distribution infrastructure, not a mall assistant. It can help a tenant’s products appear in an external shopping journey, but it does not automatically locate the correct branch or map a route through the centre.
Stripe’s Shared Payment Tokens let payment permissions be scoped, for example to a seller, amount or time window, without handing raw payment credentials to an agent. Such controls address what a system may spend, not whether it recommended the best product.
A concrete retailer example comes from Tapestry. Coach and Kate Spade support single-item purchases in Gemini and Google Search’s AI Mode in the US through the Universal Commerce Protocol, with Google Pay and customer approval for each transaction.
That is a useful counterexample to the claim that AI only advises and never completes a sale. It also shows why transaction scope matters. One approved purchase is not unlimited permission to shop autonomously.
Meta and Sierra’s Personal Agent Protocol, announced October 6, 2026, addresses how personal agents interact with businesses and obtain appropriate access. The initial specification was scheduled for later in October.
It belongs on an infrastructure watchlist. A protocol proposal, a reference implementation and a working integration at a shopping centre are three separate milestones. None should be substituted for the others in a vendor comparison.
The cases above report different outcomes. Westfield’s published figure concerns routes. Pandora’s concerns service cases. Lowe’s documents deployment to associates. These are not interchangeable measures of sales impact.
A useful pilot should record the shopper’s request, the information retrieved, the action attempted and whether that action completed. An answer such as “the shop is open” should be traceable to a current record. A failed lookup should stay a failed lookup rather than becoming a confident guess.
For navigation, measure the correctness of the destination and route before claiming visits. For shopping, distinguish a cart created from a paid order. For support, distinguish a conversation ended from an issue resolved. For staff tools, measure whether the proposed answer or plan was accepted and usable.
Physical analytics can help with a different part of the picture. Ouster and Digital Mortar’s Alltown Fresh case uses lidar and spatial analytics to study movement and inform store layouts. Ouster supplies sensing and perception technology; Digital Mortar supplies analytics. Ouster’s Gemini software is unrelated to Google’s Gemini assistant.
This is measurement infrastructure, not an AI shopping agent. The case does not establish attribution from a chatbot recommendation to a store visit. Linking those events would require an additional, explicitly designed measurement method with appropriate consent and controls.
Start with the promise the service will make. “Find a children’s shoe shop” requires different information from “reserve this shoe in size 28 and take me to the pickup counter.”
The first needs an accurate tenant directory, categories, trading status, hours and location. The second also needs branch-level stock, reservation authority and an order system. A mall should not imply that it has the second capability because it has built the first.
A proposal should identify who updates those records, how quickly changes reach the assistant and what happens when sources disagree. Accessibility information and temporary route closures need the same attention as store names. A scheduled opening must not become a recommendation to visit a store that is still fitting out.
Ownership matters too. The operator should know who can access conversation records, which requests may trigger actions, where a human takes over and whether sponsored recommendations are distinguished from organic ones. Product pages and third-party content should never be allowed to grant the agent new permissions.
Cost comparisons should include setup, data cleanup, map maintenance, integrations, usage and continuing support. Compare the cost of a completed task, not just the monthly price of a chat window. The cases reviewed here do not supply a standardized, comparable price-and-performance benchmark.
The most interesting leasing insight may eventually be what the assistant cannot satisfy. Repeated requests for a missing service could identify a gap in the tenant mix. But assistant users are a self-selected group. Treat those requests as a hypothesis to test against catchment research, tenant performance and other visitor feedback, not as a ready-made leasing recommendation.
A sensible first deployment makes a narrow promise and keeps it. Expansion follows when the data and the measured results justify it.
The commercial value is in the task completed, not the conversation started.
Miss A, Rowan and HEYTEA are on US mall opening lists in fall 2026. Their public records show what each…
Gross margins and selected profit metrics for Coach, Pop Mart, lululemon, Miniso, Five Below, TJX and Costco, plus where each…
NikeSKIMS, Birkenstock and Starbucks Reserve chose HKRI Taikoo Hui for major China debuts. Why the Shanghai complex keeps winning new…
Bluewater opens its first new anchor since 1999. Next takes the former House of Fraser space with about 132,000 square…
Opening in IMAX 70mm on July 17, 2026: Christopher Nolan's The Odyssey. Select shows went on sale one year in…
European retail is scaling AI, agentic commerce and retail media, but consumer trust is becoming the constraint. Four structural shifts…
