AI Can Write Software. In a Store, You're Still on Your Own.

Walk into a wine store and ask for something simple:

"I need a red under $30 for spicy food. Interesting, but not weird."

Six hundred bottles stare back.

A web search recommends wines the store does not carry. Shelf tags offer scores, sale prices, and fragments of tasting notes. The employee who really knows wine is helping someone else. You pick up one bottle, put it down, search your phone, lose the thread, and finally buy the label you recognize.

Or you leave with nothing.

This is absurd.

AI can write code, interpret images, summarize a thousand pages, and hold a natural conversation. Retailers can track products from warehouse to checkout. Brands can target an advertisement to a household across the country.

But when a shopper is standing three feet from the shelf — at the exact moment a purchase decision is being made — the store suddenly becomes dumb.

Physical retail has computers everywhere except where the shopper needs one.

That missing interface is what the in-store AI shopping assistant can become. Not another digital sign. Not a chatbot trapped inside a kiosk. Not a robot wandering an aisle because robots look futuristic.

A real assistant understands what the shopper is trying to accomplish, knows what this store actually sells, narrows hundreds of products to a decision-ready few, explains the trade-offs, helps the shopper act, and tells the retailer whether the interaction created value.

It may appear on a phone, screen, smart cart, retailer app, wearable, or eventually a robot. The device is not the product.

The product is a better decision.

Retail digitized everything around the decision

Physical stores are not technologically primitive. Retailers have spent decades digitizing catalog, inventory, pricing, promotion, loyalty, checkout, labor, fulfillment, loss prevention, merchandising, and advertising. E-commerce represented 16.9% of U.S. retail sales in the first quarter of 2026.¹ Online shopping did not replace stores. It changed what shoppers expect from a buying interface.

Online, a shopper can describe a need, filter an assortment, read reviews, compare products, save candidates, receive recommendations, and buy without finding an employee. In a store, the same shopper often gets packaging, price, placement, promotion, and whatever human help happens to be available.

The gap becomes painful whenever the category is confusing: wine for a dinner or gift, a supplement for a routine, skincare across ingredients and claims, a hardware part that must fit, 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, prices, promotions, availability, and human advice. When that work becomes too difficult, they buy the familiar option, choose on price or packaging, wait, guess, or abandon the category.

An in-store AI assistant removes that work.

A real assistant does more than converse

A product can produce eloquent answers and still be useless in a store. The complete system has six jobs:

The first two jobs make the assistant helpful. Grounding and action make it commerce software. Measurement and learning make it a platform.

Imagine the wine shopper sees one clear promise: Find the right bottle for dinner in two minutes. No app required. The shopper scans, speaks, types, or taps: "Red. Under $30. Spicy food. Nothing too heavy."

The assistant responds with three bottles the store actually carries. It explains which is brightest, which is the safest crowd-pleaser, and which is more adventurous. When the shopper says, "Not earthy — compare the first two," the system keeps the original mission, removes the rejected direction, explains the remaining trade-off, shows the shelf location, and preserves the shortlist for an associate.

That is not a chatbot answering a question. It is a decision system moving through a physical purchase.

We built the assistant. Shoppers still had to choose to use it.

We learned this the expensive way — not in money, but in time and assumptions.

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, and explain the choices. In April, AICap participated in what we describe internally as the first voice-AI-assisted sale in a physical store.

The live results were still not good. Only a small number of shoppers engaged, and the test produced very few sales.

We were right about the shopper problem and wrong about the first engagement surface.

We had built a capable assistant and assumed that putting it in a store would create usage. It did not. The kiosk asked shoppers to notice an unfamiliar machine, infer what it could do, decide that talking to it in public was socially comfortable, and begin a conversation — all before receiving value.

The technology worked better than the invitation.

That failure changed our direction. A shopper phone lowers hardware cost, provides privacy, moves with the shopper, and preserves the conversation, shortlist, offer, or product location. No app or account is required. A visible prompt can open the experience through a QR code or link.

But a QR code is not a strategy either. Nobody scans because scanning is exciting. A shopper scans because the store promises something useful now: a fast recommendation, a clear comparison, a product location, help without waiting, or an additive benefit such as a gift service or accessory.

What earns engagement: visible shopper problem + credible promise + low-friction access + immediate value

The lesson is larger than kiosk versus mobile. A surface cannot compensate for a weak proposition. And a strong assistant creates zero economic value when shoppers do not begin.

Our next test at Broadway is therefore not simply a different screen. It is a test of the complete invitation: placement, promise, access, incentive, useful interaction, action, and measurable outcome.

The hard part begins when the conversation sounds easy

Natural conversation creates the illusion that the rest of the system is simple. It is not.

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

One identity error spreads everywhere: recommendation → price → availability → shelf location → explanation → brand eligibility → transaction attribution → learning

Product identity is not back-office cleanup. It is the foundation of trust.

The same discipline applies to store truth. Models can interpret intent 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 need authoritative sources and deterministic controls.

Category knowledge also has to become useful rather than encyclopedic. 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.

Multimodal does not mean watching the shopper. It means preserving one decision across voice, text, touch, product scans, images, location, and other useful context. A signal belongs in the product only when it reduces effort, improves the decision, or strengthens measurement.

Reinforcement learning is an eventual capability, not a starting point. Before AICap can optimize a selling policy, it needs stable actions, reliable outcomes, causal measurement, and enough traffic. Until then, evaluation and controlled experiments are more credible than claiming the system "learns from every interaction."

Every major retail system is moving toward the same moment

No company began with the complete category. Smart-cart companies began with item recognition and checkout. Retail-data companies began with catalogs and inventory. Connected-store companies began with shelves and location. Retailers began with loyalty and transactions. Consumer AI companies began with language, voice, and vision.

Now those systems are converging on the shopper's decision.

Instacart is the strongest direct strategic competitor because it combines retailer relationships, grocery transactions, advertising, checkout, a massive catalog engine, Caper Carts, and an announced conversational assistant. Its Catalog Engine uses more than 1.3 billion product data points; by March 2025, Caper Carts had reached more than 60 U.S. cities and average sessions exceeded 30 minutes.^2,3 Amazon leads physical transaction execution through Dash Cart.⁴ Vusion is making products, shelves, and store space machine-readable.^5,6 Lowe's reports more than five million associate questions and a 200-basis-point customer-satisfaction increase when Mylow Companion was used.⁷ Walmart and consumer agents are making multimodal shopping increasingly ambient.^8,9

Each player approaches the decision from a different asset. The strategic control point is not necessarily the cart, shelf, model, or robot. It is the layer that preserves the shopper's mission across product guidance, store action, transaction, and outcome.

The economic prize is not engagement

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.

The controlling business outcome should be incremental gross-margin dollars caused by the assistant. AICap can create them by converting a shopper who otherwise would not buy, preventing category abandonment, completing a basket, increasing confidence in an unfamiliar product, improving product mix, making a promotion relevant to a real mission, or bringing in an associate at the right moment.

Onki's planning target is ambitious: increase GM$ by 50% among eligible shopper opportunities AICap causes to engage. At a 10% engaged-opportunity share, the arithmetic produces a 5% category or storewide lift:

Planning target — not an achieved result: 10% engaged-opportunity share × 50% incremental GM$ lift = 5% overall GM$ lift

The distinction between target and result matters. Engaged shoppers are self-selected. They may already be more curious, more uncertain, or more likely to buy. Comparing them with everyone else can manufacture a lift.

The question is not, "How much did engaged shoppers buy?" It is, "What happened because the assistant existed?"

A credible test therefore randomizes invitation or exposure, alternates time blocks or surfaces, stages rollout across comparable stores, or uses matched stores and periods. It links the interaction to a transaction or another credible outcome and separates conversion, basket, product mix, margin rate, incentive cost, variable system cost, and cannibalization.

That proof standard is harder than reporting sessions. It is also what makes the product economically consequential.

Retail media becomes decision media

Most in-store media begins with inventory: We have a screen, shelf strip, cart display, or app placement. What advertisement should fill it?

Conversational assistance begins with a shopper need: I am choosing tequila for a party. I want something people will recognize, but I do not want to overpay.

That difference creates a new media unit:

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

A brand can become useful inside an actual decision. It can fund product education, an additive shopper benefit, explanation or message testing, qualified engagement, incrementality measurement, and aggregated insight into shopper missions and objections.

The governing rule is simple: payment can influence which eligible information or offer appears. It cannot make an unsuitable product suitable, override a hard shopper constraint, invent store truth, or guarantee the top recommendation.

Retail-media standards are already moving toward common in-store definitions, closed-loop measurement, shopper missions, and incrementality rather than raw impression counts.^10,11

Retail media sells exposure near a decision. Decision media participates usefully in the decision and measures what changed.

This creates the strategic flywheel behind AICap:

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

The loop works only while the assistant remains useful. If monetization makes the recommendation less trustworthy, the network destroys the shopper behavior on which its economics depend.

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

An excellent specialist can outperform software in a difficult conversation. The problem is that the right specialist is not always available at the right shelf in the right language.

AI is strongest at immediate availability, broad product knowledge, repetitive questions, comparison, multilingual access, and context capture. People are strongest at trust-heavy decisions, exceptions, physical service, retrieval, and nuanced closing.

The system should decide who performs the next job — and carry the shopper's mission, constraints, shortlist, and location across the handoff. A shopper who already spent three minutes explaining the problem should not have to begin again when a person arrives.

Why wine is the right wedge

Wine makes the missing interface impossible to ignore. The assortment can be huge. Product names are unfamiliar. Vocabulary is specialized. Taste is personal. Food and occasion matter. Price and margin vary widely. Expertise is unevenly available. The package cannot explain everything.

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

That is why AICap begins with East Asian alcohol and broadens through wine and beverage alcohol. Wine exposes the problem clearly enough to build the system. It does not define the platform's boundary.

What Onki is actually building

Onki is building AICap as a device-independent multimodal engagement and conversion platform for physical retail.

Today, AICap has working conversational voice, text/touch prototypes, store-specific alcohol recommendations, comparison and shortlist capabilities, plus large-screen and mobile UX work. Onki has also built Shopify catalog extraction and is developing the product and data foundation required for broader deployment.

Broadway disproved the assumption that technical capability automatically earns repeated shopper engagement. AICap has not yet proved causal GM$ lift, scaled retailer distribution, or a production decision-media network.

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

The proof loop: eligible shopper opportunity → visible promise → useful session → recommendation or action → purchase → incremental GM$

Once that loop works, the same application and intelligence layer can move across a shopper's phone, retailer screen, partner surface, smart cart, associate tool, wearable, and eventually a robot or humanoid. The interface changes. The shopper's mission and the decision system persist.

The next interface is the decision

That is why the opportunity is larger than a wine assistant, a kiosk company, or a robot. The long-term control point is the application layer that understands the shopper, knows the products, selects the next useful action, and learns what actually created value.

Retailers already know what products entered the store, where those products sit, what they cost, and what sold. They still know remarkably little about the decision that happened — or failed to happen — in front of the shelf.

AI can make that decision interactive. It can understand why the shopper came, organize the store's actual products around the mission, explain the trade-offs, bring in a person when needed, connect the interaction to a transaction, and learn whether the help mattered.

We have already learned that technical capability is not enough. The system must earn the shopper's participation before it can improve the decision. That is the work now: one useful, measurable loop from intent to action to incremental gross-margin dollars.

Once that loop works, the phone, screen, cart, associate tool, or robot becomes a surface. The decision layer becomes the platform.

Three questions that matter

Is this just a chatbot?

No. A chatbot can answer questions. A complete in-store assistant also knows the store's actual products and context, narrows the assortment to 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, context, cost, and the action required.

Does it replace store associates?

No. The strongest model combines AI availability and knowledge with human trust and physical service. AICap should handle discovery, comparison, multilingual access, and context capture, then bring in an associate when judgment or execution adds value.

---

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. Amazon, Amazon Dash Cart: New Features and Whole Foods Market Expansion, July 11, 2022. Source
5. Vusion, VusionGroup Unveils EdgeSense AI, October 20, 2025. Source
6. Vusion and Qualcomm, The AI-Native Store Vision, February 25, 2026. Source
7. Lowe's, Q1 2026 Earnings Call Transcript, May 20, 2026. Source
8. Walmart, The Future of Shopping Is Agentic. Meet Sparky, June 6, 2025. Source
9. Google, Gemini App Updates from Google I/O 2025, May 20, 2025. Source
10. Interactive Advertising Bureau, Standards and Guidelines: In-Store Retail Media and Advanced Measurement, accessed July 26, 2026. Source
11. Interactive Advertising Bureau, Digital Out of Home & In-Store Retail Media Playbook, May 21, 2024. Source