What Is an In-Store AI Shopping Assistant?

A shopper stands in front of 600 wines looking for a red under $30 for spicy food. The shelf offers labels, prices, promotions, and fragments of tasting notes. A web search offers products the store may not carry. The employee who really knows wine is helping someone else.

The store knows what every bottle costs and what eventually sells. It still does not know what this shopper is trying to accomplish.

Retailers have spent decades digitizing the operation around the shopper: catalog, inventory, pricing, promotion, loyalty, checkout, labor, merchandising, and media. The purchase decision itself remains largely analog.

Direct answer: An in-store AI shopping assistant is software that turns a shopper's natural-language intent into a grounded, decision-ready recommendation using the store's actual products and context. It explains the choice, helps the shopper act, connects to people or commerce systems when needed, measures the outcome, and improves through evaluation and experimentation.

The governing shift is simple: retail digitized the store's operations. In-store AI digitizes the shopper's decision.

That is the shopper-facing edge of a larger change: AI-mediated commerce. Instead of forcing people to navigate catalogs, filters, signs, menus, and checkout systems, AI lets them state a goal and organizes the relevant products, trade-offs, and actions around it.

Over the next decade, carts, connected shelves, associate copilots, retail media, consumer agents, and robots will all feed or execute against that decision layer. The largest retail-tech shift is not one device. It is AI becoming the interface between shoppers and products, online and in physical stores.

Why does physical retail need a decision layer?

E-commerce represented 16.9% of U.S. retail sales in the first quarter of 2026.¹ Online retail has trained shoppers to expect search, comparison, recommendation, reviews, personalization, and measurable action. Inside a store, the same shopper often receives packaging, price, placement, promotion, and whatever human assistance happens to be available.

The problem becomes acute in categories where the choice is consequential but difficult to decode: wine for a meal, a supplement for a routine, skincare across ingredients and claims, hardware that must fit, electronics with competing specifications, or a duty-free purchase in an unfamiliar language.

The shopper does not need more content. The shopper needs the store's information organized around one immediate question: What should I choose here, for this purpose, right now?

Today, the shopper bears the integration cost. They reconcile packaging, generic search results, reviews, price, promotion, availability, and inconsistent human advice. When that work becomes too difficult, they buy the familiar option, choose on packaging or price, wait, guess, or abandon the category.

A useful assistant removes that work. It narrows the assortment to a small set the store actually carries, explains the meaningful trade-offs, answers refinements without restarting, shows the next action, and preserves the context when a person needs to step in.

Six jobs separate a real assistant from a novelty

The first two make the assistant useful. Grounding and action make it store-native. Measurement and learning turn it into a commerce platform.

Many products marketed as AI shopping assistants perform the first two jobs. The category changes when the system also owns store truth, action, outcome, and learning.

What does the shopper experience look like?

A good interaction is short and literal:

• A visible reason to engage. The shopper sees a specific benefit: find the right bottle for dinner, compare two products in plain language, locate an item, or get help without waiting.
• A natural request. The shopper speaks, types, or taps. Voice, text, and touch share one decision state.
• Immediate momentum. The assistant provides an initial direction or small candidate set before asking for more information than it needs.
• Refinement without restart. The shopper can say "less expensive," "not earthy," "compare these two," or "something more unusual" while the original mission persists.
• Grounded explanation. Product facts are translated into consequences for this shopper and this use.
• Action. The system shows a location, saves a shortlist, messages an associate, presents an offer, or initiates a commerce action.
• Outcome. The system records pickup, purchase, redemption, handoff, explicit feedback, or another credible signal.

Shopper loop: visible proposition → request → initial recommendation → refinement → explanation and comparison → action → outcome

The experience should feel like immediate access to an informed, patient, nonjudgmental specialist. It earns engagement through the shopper benefit, not through the novelty of AI. A QR code is an access mechanism, not a reason to scan.

The hard part begins when the conversation sounds easy

A fluent conversation can hide a weak commerce system. The two foundations that most often fail are product identity and store truth.

Product identity is not back-office cleanup

Suppose a shopper asks about a bottle of Dassai sake. The store catalog may identify it by retailer SKU, UPC, distributor description, Japanese name, translated name, pack size, image, or merchant-specific handle. Two entries may describe the same bottle. One may combine a bottle and gift box. Another may omit the grade. Availability may be stale.

When identity is wrong, every downstream system is contaminated: recommendation, price, availability, shelf location, explanation, brand eligibility, transaction attribution, and learning.

Failure chain: wrong identity → wrong recommendation → wrong price or location → wrong attribution → corrupted learning

A real product graph must reconcile retailer SKUs, variants, UPCs, distributor and brand records, pack sizes, vintages, images, merchant handles, and other identifiers while preserving source and confidence.

Store truth must be deterministic

Models can interpret intent, expand concepts, compare products, and explain trade-offs. They cannot be allowed to invent whether a product is stocked, what it costs, whether it is promoted, or where it sits. Those facts and actions require authoritative sources, precedence rules, validation, and explicit uncertainty.

The same discipline applies to category knowledge. A vintage-weather fact matters only when it changes likely style, quality, value, or drinking window. A hardware specification matters only when it changes compatibility or performance. The system's job is not to display everything it knows. It is to know what matters now.

How the system actually works

The interface is the visible edge of a seven-layer operating system. Each layer can be sourced from different partners, but the shopper should experience one continuous decision.

Figure 1. A useful assistant closes the loop from intent to outcome; it does not stop at an answer.

Governance cuts across every layer: permissions, product truth, commercial policy, privacy, brand controls, and failure handling. A foundation model can make the conversation fluent. It does not replace the commerce context, decision rules, action integrations, or measurement needed to make the answer reliable.

The surface is not the product

The same decision system can appear on a shopper's phone, shared screen, retailer app, smart cart, associate tool, wearable, or eventually a robot. The surface changes discovery, continuity, privacy, context, cost, and available actions. It does not change the core product.

What Broadway Spirits taught Onki

In February and March 2026, Onki tested AICap as a kiosk at Broadway Spirits in New York. The system ran for tens of hours. It could hold a voice conversation, recommend products from the store, explain choices, and participate in a real physical-store sale.

The live results were still poor. Only a small number of shoppers engaged, and the test produced very few sales. A capable assistant sitting in a store did not automatically become part of shopping behavior.

The kiosk asked a shopper to notice an unfamiliar machine, infer what it could do, decide that talking to it was socially comfortable, and begin a conversation before receiving value. The technology worked better than the invitation.

That result drove AICap toward mobile-first testing and a sharper operating model:

Engagement equation: surface + placement + visible promise + immediate shopper benefit = earned interaction

A phone lowers hardware cost, provides privacy, moves with the shopper, and preserves the conversation, shortlist, offer, or location. But a QR code is not a strategy either. The shopper scans because the store promises something useful now: a fast recommendation, a clear comparison, a product location, an additive benefit, or help without waiting.

Five systems are converging on the shopper's decision

No company began with the complete category. Consumer AI companies began with language, voice, vision, and personal context. Smart-cart companies began with item recognition and checkout. Connected-store platforms began with shelves and spatial infrastructure. Retailers began with catalog, loyalty, and transactions. Associate copilots began with employee productivity.

Instacart: commerce infrastructure moves into conversation

Instacart is the most direct strategic competitor. Its announced Cart Assistant extends across retailer websites, apps, and Caper Carts, while its broader platform already connects catalog, basket, location, advertising, and checkout. Its Catalog Engine has extracted more than 1.3 billion product-data points using vision-language models and human verification.²

By March 2025, Caper Carts were available in grocery stores across more than 60 U.S. cities, and Instacart said shoppers averaged more than 30 minutes with a cart. Its in-store ads can be inventory- and aisle-aware and can extend the same campaign from digital commerce into the aisle.³ Its disclosed computer vision is concentrated on products, shelves, inventory, interactions, heat maps, and cart location—not on a public claim of facial, demographic, gaze, appearance, or emotion understanding.⁴

Amazon: the transaction closes inside the cart

Amazon Dash Cart uses computer vision and sensor fusion to recognize items added or removed, maintain a running total, integrate a shopping list, surface nearby recommendations and promotions, and let the shopper leave through a dedicated lane.⁵ Its advantage is not the deepest consultative advice. It is turning in-store guidance into a completed physical transaction.

Vusion: the store becomes spatially legible

Vusion's EdgeSense AI combines 3D locationing, computer vision, connected shelf infrastructure, and natural-language models. Vusion and Qualcomm describe an AI-native store in which products, shelves, associates, shoppers, phones, wearables, and AI applications operate against a real-time spatial layer.^6,7 Vusion is not primarily building an expert seller. It is making the physical store machine-readable for whatever intelligence layer wins.

Lowe's: AI-assisted expertise reaches operating scale

Lowe's Mylow Companion starts with the associate. By May 2026, employees had asked it more than five million questions; Lowe's had added voice detection and Spanish support.⁸ Lowe's also reported that customer satisfaction increased about 200 basis points when associates used Mylow Companion and that online conversion more than doubled when customers engaged with Mylow.⁹ These are company-reported operating results rather than a disclosed randomized evaluation, but they show that AI assistance can earn repeated use and matter commercially.

Walmart and consumer agents: multimodality becomes ambient

Walmart's Sparky points toward agents that compare products, synthesize reviews, personalize decisions, and expand into voice, camera, reordering, and service booking.¹⁰ Google has made live camera-and-voice interaction widely available through Gemini Live.¹¹ Consumer agents may offer the lowest-friction interface because the shopper already carries them.

Their gap is authoritative store truth: exact assortment, current price, inventory, promotion, shelf location, retailer policy, staff routing, and closed-loop measurement. Consumer and retailer-controlled agents may compete, interoperate, or become different surfaces for the same decision.

The strategic control point is not necessarily the company that manufactures every cart, shelf sensor, screen, or robot. It is the layer that preserves the shopper's mission across product guidance, store action, transaction, and outcome.

Multimodal does not mean watching the shopper

Multimodal means maintaining one decision across voice, text, touch, product scans, images, location, and other useful context. The shopper should be able to point at a bottle, ask a question by voice, compare two products on screen, and continue by text without the system forgetting the mission.

The signals form a hierarchy. Direct input—what the shopper says, types, taps, scans, or rejects—is the clearest evidence of intent. Object and environment signals can identify the product, shelf, basket, or location. Interaction signals can show hesitation, repetition, abandonment, or acceptance. Inference about the person is the most speculative layer and should earn its place through measurable value.

A camera recognizing a bottle does not understand the shopper. A heat map does not know the mission. Facial, demographic, gaze, appearance, or emotion inference is neither required nor automatically useful. Add a signal only when it reduces shopper effort, improves recommendation quality, enables action, or strengthens measurement enough to justify the complexity.

Surface choice therefore determines the signal portfolio. A phone may have less ambient store context than a fixed camera but gains privacy, continuity, movement, and lower deployment cost. A smart cart gains basket and location context but may be weak at expert explanation. The system should combine signals around the decision—not collect signals simply because sensors exist.

Learning begins after the product is measurable

Adaptive selling begins only after the product is measurable. The practical progression is:

• Define stable shopper states, system actions, hard constraints, and outcomes.
• Instrument the complete session and connect it to downstream action.
• Evaluate recommendation, explanation, and factual quality offline.
• Run controlled experiments on propositions, questions, explanations, offers, and handoffs.
• Introduce bounded exploration or contextual bandits where traffic and attribution support it.
• Use broader reinforcement learning only when the reward, safeguards, and evidence are strong enough.

Contextual bandits are useful when the system must balance exploiting actions already likely to work with exploring alternatives that may improve future decisions. Retail research is actively testing these methods for personalized offer selection.¹² They are narrower and easier to bound than a general agent learning an entire selling policy.

The hardest problem is the reward. Clicks, time spent, recommendation acceptance, purchase, gross-margin dollars, repeat use, and trust can point in different directions. A system rewarded only for engagement may learn to prolong the conversation. A system rewarded only for immediate margin may damage fit and repeat use.

The rule is simple: learning chooses among eligible actions. It does not rewrite product truth, ignore hard shopper constraints, or make an unsuitable product suitable. Until reliable outcomes and sufficient traffic exist, structured rules, supervised ranking, offline evaluation, and controlled experiments are more credible than claiming the AI learns from every interaction.

The economic outcome is incremental gross-margin dollars

A shopper can spend five minutes with an assistant, compliment it, and buy exactly what they would have bought anyway. That is a pleasant interaction, not proven economic value.

An in-store assistant can create incremental GM$ by converting a shopper who otherwise would not buy, preventing category abandonment, completing a basket, improving product fit and mix, increasing confidence to try an unfamiliar product, making a promotion relevant to a real mission, enabling a productive associate intervention, selling from the backroom or an extended assortment, or improving repeat purchase through better first-time fit.

Controlling question: What happened because the assistant existed?

Onki uses a planning target of 50% more GM$ among eligible shopper opportunities AICap causes to engage. At a 10% engaged-opportunity share, the arithmetic implies a 5% category or storewide lift:

Planning model: 10% engaged-opportunity share × 50% incremental GM$ lift among those opportunities = 5% overall GM$ lift

These are AICap targets, not achieved results or generic industry benchmarks. The denominator must align: counting sessions against total door traffic while measuring margin against category buyers creates false precision.

Why engaged versus unengaged is not proof

Engagement is self-selected. People who begin a session may already be more curious, uncertain, or likely to purchase. A credible design uses randomized invitation or exposure, randomized time blocks or surfaces, a stepped rollout, or matched stores and periods. It links the interaction to a transaction or another credible outcome and separates conversion, basket, mix and margin rate, incentive cost, variable system cost, substitution, cannibalization, and repeat behavior.

The operating funnel should track eligible opportunity, exposure, session start, useful interaction, accepted recommendation or shortlist, action, purchase, GM$, and repeat use. Quality metrics should separately track product-identity errors, unsupported facts, unavailable-product recommendations, hard-constraint misses, misleading commercial influence, latency, and failed handoffs.

Retail media becomes decision media

Traditional in-store media begins with inventory: a screen, shelf strip, cart display, or app placement that needs a message. Conversational assistance begins with a shopper need: "I am choosing tequila for a party. I want something recognizable, but I do not want to overpay."

Decision-media unit: expressed need → eligible products → useful education or offer → consideration → action → measured outcome

A brand can fund product education, relevant participation in a consideration set, an additive shopper benefit, explanation or message tests, qualified engagement, incrementality measurement, and aggregated insight into missions and objections.

Retail-media standards are moving toward unified in-store definitions, closed-loop measurement, and incrementality rather than impression counts alone.¹³ The IAB's in-store playbook similarly emphasizes shopper missions and closed-loop measurement.¹⁴

The governing rule is that commercial influence must remain inside product truth, store availability, and hard shopper constraints. Payment may affect which eligible information or offer appears. It must not make an unsuitable product suitable or guarantee the top recommendation.

The strategic flywheel

• Useful guidance earns shopper participation.
• Participation creates retailer GM$.
• Expressed intent creates qualified brand opportunities.
• Brand funding supports free retailer distribution.
• Broader distribution expands product, intent, and outcome data.
• Better data improves guidance and measurement.
• Better evidence attracts more retailers, partners, and brand budgets.

This is why decision media is more than a monetization feature. It can become the economic engine that makes a shopper-first product free and broadly distributed.

AI does not eliminate the associate. It makes the associate arrive prepared.

An excellent specialist can outperform software in a difficult interaction. A person can judge nuance, retrieve an item, handle an exception, open a case, and create trust. The problem is that the right specialist is not always available at the right shelf in the right language.

The assistant is strongest at immediate discovery, repetitive questions, broad product knowledge, comparison, multilingual access, private interaction, shortlist creation, and context capture. The associate is strongest at trust-heavy decisions, exceptions, retrieval, sensitive situations, and nuanced closing.

The handoff should carry the shopper's need, constraints, shortlist, and location. The operating question is not "AI or associate?" It is "Who should do the next job?"

Why wine and East Asian alcohol are strong wedges

Wine makes the missing interface visible: large assortments, unfamiliar product names, specialized vocabulary, personal taste, occasion and food context, wide price and margin ranges, and uneven access to expertise. The shopper rarely wants wine information in the abstract. They want a bottle that fits a meal, gift, budget, remembered preference, or willingness to explore.

East Asian alcohol sharpens the problem. Identity and naming are fragmented across sake, shochu, soju, baijiu, Japanese whisky, pack sizes, translations, distributors, and merchant catalogs. The category rewards real normalization and explanation rather than generic model fluency.

The category is a wedge, not the product boundary. The same decision structure transfers to supplements, skincare, hardware, electronics, beauty, pet care, duty-free, and other high-consideration categories:

Reusable system: identity → intent → eligibility → comparison → explanation → action → outcome

What Onki has learned by building AICap

Onki is building AICap as a device-independent multimodal engagement and conversion platform for physical retail. AICap is not defined by a kiosk, QR code, camera, voice interface, wine category, foundation model, or future robot. Those are surfaces, inputs, wedges, infrastructure choices, or execution systems. The product is the decision intelligence behind them.

Status as of July 27, 2026

AICap has not yet proved repeatable shopper engagement, causal GM$ lift, scaled retailer distribution, or a production decision-media network. That is not a footnote. It defines the work.

The next milestone is not a more impressive demo. It is one fully instrumented loop:

Proof loop: eligible shopper opportunity → visible proposition → useful session → recommendation or action → purchase → incremental GM$

Once that loop works, distribution and learning can begin to compound.

How should a retailer test the category?

Do not begin with a broad "AI innovation" objective. Begin with one confusing category, one specific shopper job, one defined surface, and one falsifiable economic outcome.

• Can the assistant earn engagement? Define eligible opportunities, exposure, proposition, placement, surface, and incentive. If shoppers do not start, recommendation quality is irrelevant.
• Does it improve the decision? Measure usefulness, recommendation acceptance, confidence, time to decision, refinement burden, and observed action. Review failures, not only positive sessions.
• Is it grounded in store truth? Test exact product identity, assortment, price, inventory, promotion, location, hard constraints, and provenance.
• Can it cause economic value? Predefine the control, transaction linkage, margin calculation, incentive cost, and decision threshold before the pilot begins.
• Can the operating model scale? Record catalog work, data cleanup, integration work, staff burden, incidents, variable cost, failure modes, and time to deploy.

The pilot should end with a decision: expand, revise, or stop. "We learned a lot" is not a success metric.

The strategic conclusion

Stores have spent decades becoming better at knowing what products they have, where those products are, what they cost, and what sold.

The next step is understanding what the shopper is trying to accomplish before the transaction happens.

That requires more than a chatbot, another screen, or a robot placed in an aisle. It requires a decision layer that connects intent to exact product identity, store truth, explanation, action, and measurable economic value. The same layer can move across phones, retailer surfaces, smart carts, associate tools, wearables, and eventually embodied systems while preserving the shopper's mission and the retailer's economics.

That is the category Onki is building.

Frequently asked questions

Is an in-store AI shopping assistant just a chatbot?

No. A chatbot can answer questions. A complete in-store assistant also understands the shopper's mission, knows the store's actual products and context, recommends and explains a decision-ready set, drives a physical or commerce action, measures the outcome, and improves through evaluation and experimentation.

Does it require a kiosk?

No. A kiosk is one surface. Phones, retailer apps, smart carts, associate tools, partner screens, wearables, and future robots can expose the same decision system. The right surface depends on discovery, continuity, shopper friction, available context, cost, and the required action.

Does it need computer vision?

No. Vision can recognize products, shelves, pointing, location, and other interaction signals when those capabilities reduce effort or improve measurement. Product recognition is not the same as understanding the shopper, and demographic or emotion inference is not required for a useful assistant.

Does AICap use reinforcement learning today?

No. Reinforcement learning is a long-term AICap direction. The current priority is to define stable states, actions, constraints, outcomes, attribution, and evaluation. Structured rules, supervised ranking, experiments, and bounded bandit methods should precede broader reinforcement learning.

How should success be measured?

Measure incremental gross-margin dollars against a credible control. Diagnose the result through eligible opportunity, exposure, session start, usefulness, recommendation acceptance, action, purchase, product mix, incentive cost, variable system cost, and repeat behavior.

Can brands pay to participate?

Yes. Brands can fund relevant education, offers, qualified engagement, experimentation, and measurement. Commercial influence must remain inside product eligibility, shopper constraints, store truth, and a clear governance policy. Payment cannot make an unsuitable product suitable.

---

Sources

1. U.S. Census Bureau, Quarterly Retail E-Commerce Sales, Q1 2026, May 18, 2026. Source
2. Instacart, Instacart Announces New Enterprise AI Solutions to Democratize AI for Grocers of All Sizes, November 4, 2025. Source
3. Instacart, Instacart Expands In-Store Advertising to All Brands on Caper Carts, March 25, 2025. Source
4. Instacart, Introducing Store View and Second Store Check, March 27, 2025. Source
5. Amazon, Amazon Dash Cart: New Features and Whole Foods Market Expansion, July 11, 2022. Source
6. Vusion, VusionGroup Unveils EdgeSense AI, October 20, 2025. Source
7. Vusion and Qualcomm, The AI-Native Store Vision, February 25, 2026. Source
8. Lowe's, Q1 2026 Earnings Call Transcript, May 20, 2026. Source
9. Lowe's, Q3 2025 Earnings Call Transcript, November 19, 2025. Source
10. Walmart, The Future of Shopping Is Agentic. Meet Sparky, June 6, 2025. Source
11. Google, Gemini App Updates from Google I/O 2025, May 20, 2025. Source
12. N. Tankovic et al., Scalable and Interpretable Contextual Bandits: A Literature Review and an Experimental Framework for Personalized Retail Promotions, 2025. Source
13. Interactive Advertising Bureau, Standards and Guidelines: In-Store Retail Media and Advanced Measurement, accessed July 26, 2026. Source
14. Interactive Advertising Bureau, Digital Out of Home & In-Store Retail Media Playbook, May 21, 2024. Source