What problem does it solve?
Search boxes and filters work when the shopper knows the product name. They fail when the shopper knows the problem: "what do I need to fix a leaky faucet", "what should I wear to a wedding in Barcelona in November", "will these bindings fit these boots". In a store an experienced associate answers those questions and sells the right basket; online the shopper reads reviews in ten tabs, guesses, or leaves.
Retailers with large or technical assortments (home improvement, sporting goods, fashion, electronics) feel this most, and their best experts are scarce and seasonal. Earlier product recommendation engines ranked items from behaviour but could not hold a conversation or explain a trade off. Generative models grounded in the catalog, reviews and how to content can, and the larger retailers have now put them in front of hundreds of millions of shoppers.
How does it work?
- Understand the need. The assistant takes a free text or spoken question about a product, a project or an occasion, and asks clarifying questions (skill level, budget, size, location) the way a good associate would.
- Retrieve from the retailer's own data. It searches the product catalog, specifications, customer reviews, questions and answers and the retailer's how to content, and checks price and local stock.
- Recommend and explain. It proposes a short list or a complete basket, compares options and says why each item fits, including compatibility between items.
- Personalise with consent. For logged in customers it can use purchase history and preferences, and it adapts to the page the shopper is on.
- Hand over to the purchase. It adds items to the basket, reserves in store or books a service, and routes complex or high value questions to a human expert.
- Stay inside the rules. Prices, promotions and availability come from live systems, never from the model, and sponsored products are labelled.
- Audience
- Customer facing
- Autonomy
- Autonomous
- Adoption
- Early adopters
- Channels
- Mobile app, Web chat, WhatsApp, Phone and voice
What is it worth?
Benchmarks are computed from the public deployments below: one data point per organization per KPI, with who made each claim.
| KPI | Median | Reported range | Data points | Claimed by |
|---|---|---|---|---|
| Users served | Not pooled | 2 million to 350 million | 2 | 2 organization |
| Conversion uplift | Too few to pool | 3x | 1 | 1 vendor |
| Customer satisfaction | Too few to pool | 90% | 1 | 1 vendor |
| Satisfaction uplift | Too few to pool | 50% | 1 | 1 vendor |
Value drivers: Revenue growth, Customer experience, Lower cost to serve, Inclusion and access.
Indicative value
An online retailer with EUR 200 million in annual online sales
EUR 60,000 to EUR 720,000
Incremental gross margin from assisted sessions per year
How this is calculated
Formula: onlineSales * engagedShare * incrementalShare * grossMargin. The low scenario uses every low input, the high scenario every high input.
| Input | Low | High | Basis |
|---|---|---|---|
| Annual online sales onlineSales, EUR per year | 200,000,000 | 200,000,000 | The reference retailer. |
| Share of online sales from sessions where the shopper uses the assistant engagedShare, fraction of online sales | 0.02 | 0.06 | Editorial assumption for the first years; usage grows slowly because most shoppers still search and browse. Replace with your own adoption data. |
| Share of those sales that is truly incremental incrementalShare, fraction of engaged sales | 0.05 | 0.15 | Editorial assumption, deliberately far below the conversion multiples on this page (Sierra, the vendor, reports triple the conversion for Sun & Ski Sports shoppers who engage), because engaged shoppers are self selected. Measure with a holdout group. |
| Gross margin on incremental sales grossMargin, fraction of sales | 0.3 | 0.4 | Editorial assumption for a general merchandise retailer. Replace with your own margin. |
What it leaves out: A rough estimate of incremental margin only. It leaves out the cost of the assistant and its model usage, service contacts avoided, the effect on returns (better advice can lower them), and it assumes a holdout test confirms the uplift.
Who already uses it?
5 public deployments, strongest evidence first. Grades: A regulator or audit, B the organization itself, C vendor case study, D anonymous or estimate.
Lowe's
United States · Retail and ecommerce · 2025
Lowe's launched Mylow in March 2025, a customer facing virtual advisor built with OpenAI that answers home improvement questions, gives project steps and links the project to product discovery, with recommendations that can be refined by budget and zip code. In May 2025 it rolled out Mylow Companion, built on the same foundation, to associates in more than 1,700 stores, so staff on the floor get the same product, project and inventory answers. No outcome figures were published.
No outcome disclosed.
Walmart
United States · Retail and ecommerce · 2025
Walmart launched Sparky in June 2025 as an "Ask Sparky" button in its app across all categories. Sparky answers product questions, compares options, synthesizes reviews and recommends products for an occasion, and Walmart describes a roadmap toward reordering, service booking and multimodal input. It joins Walmart's earlier generative AI features for search, review summaries, product descriptions and comparisons. No outcome figures were published on the launch page.
No outcome disclosed.
Amazon
United States · Retail and ecommerce · 2024
Amazon launched Rufus in 2024 as a generative AI shopping assistant in its app, trained on the product catalog, customer reviews, community questions and answers and information from the web, to answer product questions, compare items and recommend products for an occasion or need. In 2026 it combined Rufus and Alexa+ into Alexa for Shopping, an agentic assistant that adds price history, price alerts and automated buying. Amazon reports over 350 million customers in twelve months, and says US customers who use Alexa for Shopping spend on average over 40% more per order than those who do not, a comparison between self selected groups rather than a controlled test.
- Users served: at least 350 million, the 12 months to Q2 2026
"Over 350 million customers have used it in the last 12 months, and engagement accelerated in Q2, with active users nearly doubling, and interactions up over 5x year-over-year."
Claimed by: organization
Zalando
Germany · Retail and ecommerce · 2023
Zalando opened a beta of its assistant to logged in customers in Germany, Austria, the United Kingdom and Ireland by November 2023, after testing it internally and with selected customers. Since October 2024 it gives logged in customers fashion advice in their local language across all 25 Zalando markets, using Zalando's own models and OpenAI's large language models. It interprets context such as occasion, location and weather ("what should I wear to my dad's 60th birthday in November in Barcelona?"), and an update in March 2025 connected it to customer accounts and shopping history and made it aware of the page the customer is browsing. In a pilot of the personalised version Zalando observed 40% more high value interactions, such as likes and add to cart actions.
- Users served: at least 2 million, cumulative since launch, as reported in March 2025
"So far, over 2 million customers have used it to get inspired and find items they love."
Claimed by: organization
Sun & Ski Sports
United States · Retail and ecommerce · 2025
Sun & Ski Sports, a Texas based outdoor retailer with a strongly seasonal business, started its AI agent Sunny on basic returns and order status questions and then extended it to expert product advice on skis, boards, boots and bindings on its product pages. Sierra, the vendor, reports higher satisfaction on conversations the agent handles than on those transferred to humans, higher conversion for shoppers who engage with it, and a winter season without hiring temporary service staff.
- Customer satisfaction: 90%, conversations handled by the agent, as reported in October 2025
"Sunny achieves 90% customer satisfaction compared to 68% for conversations transferred to human agents."
Claimed by: vendor - Satisfaction uplift: 50%, as reported in October 2025, three years after the CMO joined in 2022
"Three years later, Sunny, their AI agent, has improved CSAT by 50% and tripled product page conversion rates"
Claimed by: vendor - Conversion uplift: 3x
"Customers who engage with Sunny convert at triple the rate of those who don't."
Claimed by: vendor
How do you implement it?
A model agnostic playbook: what to prepare, the order to build in, and what goes wrong.
Data you need
- A product catalog with complete attributes and specifications, not only marketing copy
- Customer reviews and questions and answers, where the retailer owns and may use them
- Live price, promotion and stock by channel and store
- How to and project content, with an owner and review date
- Consent records for using purchase history in recommendations
Systems to integrate
- Product information management and search or recommendation engine
- Pricing, promotions and inventory systems
- Basket, checkout and store reservation APIs
- Customer profile and consent management
- Human expert chat or video channel for handover
Complexity: Medium
The conversation is the easy part. The work is in clean, rich product data, live price and stock, compatibility rules, and evaluation of recommendation quality at the scale of a full catalog.
- 1
Pick categories where advice matters
Start where shoppers ask compatibility or project questions and conversion is low (tools, sporting goods, fashion occasions), not in commodity categories where search already works.
- 2
Fix the product data first
Missing attributes produce confident but wrong recommendations. Measure attribute completeness per category and fill the gaps before launch.
- 3
Keep price, stock and promotions live
Call the pricing and inventory systems at answer time. Never let the model state a price or a discount from its own memory or from stale retrieved text.
- 4
Build an evaluation set per category
Write real shopper questions with the right answers, including compatibility traps and questions the assistant should refuse (medical, safety critical), and run them on every change.
- 5
Launch with a holdout group
Measure conversion, basket size and returns against shoppers who do not get the assistant, so the business case rests on incremental effect rather than on self selected users.
- 6
Connect the assistant to the humans
Offer a human expert for high value or complex projects, and give store associates the same assistant so online and in store advice agree.
Guardrails
- Prices, promotions and availability only from live systems, never generated
- Recommendations only from the retailer's current catalog, with sponsored items labelled
- Refusal and a safe pointer for safety critical, medical or legal questions
- Personal data used for personalisation only with consent, and never inferred sensitive traits
- Protection against prompt injection that tries to obtain unauthorised discounts or commitments
KPIs to instrument
- Conversion and average order value against a holdout group
- Share of recommended items in stock and correctly priced at answer time
- Return rate of products bought after an assisted session
- Satisfaction and thumbs down rate per category
- Share of sessions handed to a human expert and why
Human in the loop
Merchandising and category experts own the content and review a sample of conversations per category every week. Human experts take over complex or high value projects on request. Any new capability that commits the retailer (adding to a basket, reserving stock, booking a service) is signed off before launch.
Common failure modes
- Confident recommendations from thin data
- When attributes are missing the model fills the gap with plausible text. Measure data completeness and make the assistant say when it does not know.
- Commitments the retailer did not make
- Shoppers try to talk the assistant into prices, discounts or promises. Keep all commercial terms in systems of record and test for manipulation.
- Conversion claims built on self selection
- Engaged shoppers were already more likely to buy. Without a holdout the business case is overstated.
- Advice beyond the catalog's safety limits
- Recipes, chemicals, electrical work and health products need refusals and safety content, not creative answers.
What are the risks and rules?
EU AI Act
Limited risk (transparency)
A shopping assistant interacts directly with people, so under Article 50(1) shoppers must be informed that they are dealing with an AI system unless that is obvious. It is not listed in Annex III, so it is not high risk. Manipulative or deceptive techniques that materially distort a shopper's behaviour and cause significant harm are prohibited under Article 5(1)(a), which matters for how persuasion and urgency are designed.
Rules that apply
Guidance
- Article 50, transparency obligations for providers and deployers of certain AI systems (European Union, Europe). People must be informed that they are interacting with an AI system unless this is obvious from the context.
- Article 5, prohibited AI practices (European Union, Europe). Prohibits AI that uses manipulative or deceptive techniques to materially distort behaviour in a way that causes significant harm; relevant to persuasive recommendation design.
- Unfair commercial practices directive (European Commission, Europe). Misleading claims, hidden advertising and aggressive practices are unfair whether a person or an assistant makes them, so sponsored items and claims need the same care.
- Digital Services Act, Regulation (EU) 2022/2065, Article 27 on recommender system transparency (European Union, Europe). Online platforms that use recommender systems must set out the main parameters in their terms and conditions (Article 27), which applies to marketplaces that add a conversational recommender.
Controls to put in place
- AI disclosure and labelling of sponsored recommendations
- Price and stock answers traceable to the live system call
- Evaluation set per category run on every model or content change
- Consent check before any use of purchase history
- Monitoring of conversations for manipulation attempts and unsafe advice
When it went wrong elsewhere
- Incident 622: Chevrolet dealer chatbot agrees to sell Tahoe for $1. A dealer's sales chatbot was talked into "agreeing" to an absurd price and recommending a competitor's car, showing why commercial terms must stay outside the model.
Frequently asked questions
- Do AI shopping assistants increase sales?
- The published figures show strong associations. Sierra, the vendor, reports that Sun & Ski Sports shoppers who engage with its agent convert at triple the rate of those who do not, and Amazon says US customers who use Alexa for Shopping spend over 40% more per order. Those shoppers are self selected, so measure the effect with a holdout group before building a business case on it.
- How many shoppers actually use them?
- At the largest retailers, many. Amazon reports that over 350 million customers used its AI shopping assistant in the twelve months to its second quarter 2026 results. Zalando reported in March 2025 that over 2 million customers had used its fashion assistant, which has been live in all 25 of its markets since October 2024.
- Where should the assistant get prices and stock?
- Only from live systems at the moment of the answer. Prices, promotions and availability that come from the model or from stale text lead to wrong promises, and a dealer chatbot that "agreed" to sell a car for one dollar shows how easily a model can be talked into commitments.
How to cite this page
Blits.ai AI Use Case Library, "AI shopping assistant for product discovery and recommendations", last verified 27 September 2026, https://www.blits.ai/ai-use-cases/conversational-shopping-assistant. Licensed under CC BY 4.0. Method: how we verify use cases.
Changelog
- 27 September 2026: First published