Business
The problem: search boxes fail shoppers who don’t know what to type
People who land on an online store don’t always know the exact name of what they need. They type “a present for my dad, he loves cooking” and the search box returns nothing, or a list that doesn’t help. In a physical shop, a good sales assistant would ask two questions and show three products.
Questions about size, availability or whether a product fits go unanswered, and the shopper leaves. The owner never even finds out what they wanted. What that shopper needed was AI for shopping that asks what they need.
A generic conversational chatbot isn’t enough either. If it invents a price, promises a delivery date the store can’t meet or asks for payment details, the problem stops being commercial and becomes a legal one.
Gray Wool Oriental Rug
€183,28
Gray Wool Oriental Rug
AI Shopping Assistant · Powered by AllHub
Features
What it does: a salesperson with a method, not a chatbot
Every reply blends conversation with product cards shoppers can buy: an AI shopping agent that works as conversational AI for retail, not as a search box with a chat bolted on. These are the capabilities that matter most to a store, shown with AllHub’s original demos.
It understands the request and answers in the shopper’s language
It picks up intent even when shoppers don’t use the exact words, and it copes with typos. It is multilingual conversational AI: it replies in 12 languages, and a final check makes sure the answer comes out in the language of the conversation.
Cotton Duvet Cover Set
€69,95
AI Shopping Assistant · Powered by AllHub
It sells honestly
Every reply acknowledges what the shopper asked for, adds something new and asks a single question that moves the conversation on. If the store doesn’t stock what someone wants, it says so, suggests a close alternative and logs that demand for the owner.
It talks like your store, not like a robot
Each store gives it a name, a personality, a welcome message, its own logo and colours, and chooses what each product card shows.
Store logo
Primary color
AI Agent Name
Welcome Message
Battery Outdoor Table Lamp
€26,57
Battery Outdoor Table Lamp
AI Shopping Assistant · Powered by AllHub
One catalogue for every channel
A widget on the website and a store link for social media, bios or WhatsApp. It works as an AI shopping assistant for Shopify, WooCommerce and Wix, with catalogue, prices and policies synced automatically; Amazon is in a read-only pilot.
Battery Outdoor Table Lamp
€26,57
Battery Outdoor Table Lamp
AI Shopping Assistant · Powered by AllHub
Architecture
How it works inside: nine layers per message
Every shopper message goes through a layered pipeline, always in the same order. An automated test guards that order so nobody changes it by accident. Generative AI only steps in at one of the nine layers.
Shopper message
Data shieldInput
Blocks sensitive personal data before it reaches the AI.
Manipulation guardInput
Stops attempts to change the agent’s instructions, in several languages and in disguise.
Shopper stateContext
Works out where the shopper is in the purchase and prepares suitable shortcuts.
Deterministic repliesContext
When the question fits a known flow, such as a complaint, it answers without AI: faster and always consistent.
Store knowledgeContext
Retrieves information from that store only and strips hidden instructions from descriptions or reviews.
AI draftingAI
A model from a European provider, with a pinned version, drafts the reply. The system attaches products from the real catalogue.
Output shieldOutput
Checks that the reply leaks no internal instructions or personal data and contains no fraudulent links.
Purchase guardrailOutput
If a reply implied the purchase was done or asked for payment details, it fixes it and points to checkout.
Language checkOutput
Makes sure the reply comes out in the shopper’s language. If not, it regenerates it.
Reply with real catalogue cardsStore checkout
The underlying idea is defence in depth: nothing relies on a single barrier. Input is filtered, knowledge is sanitised and output is checked before the shopper sees it.
One more detail matters: the products in each reply are not written by the AI. The system attaches them from the real catalogue, so a wrong price or a product that doesn’t exist can’t slip into a card.
Security
Security: an agent with no dangerous hands
An agent that talks to strangers on the internet will face manipulation attempts. The design starts from that assumption.
Instructions that can’t be changed
It doesn’t reveal its configuration, doesn’t switch roles, doesn’t talk about other stores or the infrastructure, and treats any text trying to sneak in as content, never as a command.
Read-only access to the catalogue
It never writes to it, doesn’t browse the internet and never touches orders, refunds or inventory. Wishlist actions are always confirmed with the shopper before they run.
Explicit permissions, deny by default
Every AllHub agent has an identity and a closed list of what it may do; everything else is denied. An operator can quarantine one agent without stopping the rest, and sensitive actions can require a single-use human approval.
Payments out of reach
It never asks for or accepts card details. Payment happens in the store’s own checkout.
Every store isolated
Every query is filtered by store, and automated tests fail if a read ever loses that filter. Connector credentials are stored encrypted.
Tested against attacks
Automated test suites for prompt injection, data exfiltration, robustness, personal-data detection and output checks, plus an internal red-teaming round in July 2026. There are also usage limits per session and per store.
GDPR
Privacy and data (GDPR)
The rule is simple: the agent doesn’t need to know who the shopper is to help them choose.
Sensitive data never reaches the AI
The data shield blocks card numbers, IBANs, ID documents or phone numbers before the model, and the history the browser sends back is scanned again on every turn.
Pseudonymised conversations
The session identifier is stored as a hash, never in clear, and conversations are deleted automatically after about six months.
It doesn’t ask for what it doesn’t need
It never asks for an address, a card or billing details.
Analytics without raw personal data
The owner sees what is being asked and what the catalogue is missing, not who asked.
EU-first design
A language model from a European provider and infrastructure designed for the European Union.
Responsible AI
Transparency and responsible AI (EU AI Act)
Shoppers have the right to know who they are talking to, and owners to understand why the agent said what it said.
It says it is an AI
If a shopper asks whether they are talking to a person, it honestly says it is the store’s automated assistant, in line with Article 50 of the EU AI Act.
It doesn’t make things up
It only mentions products that exist, copies the exact price, respects real stock and, when it doesn’t know something, says so and offers to connect the shopper with support.
Every reply can be rebuilt
For each turn it records which cards were shown, in what order and why, with which model and which version of the instructions. A specific conversation can be reviewed step by step.
It monitors its own bias
It measures whether the agent treats a language, a type of shopper or a product source worse than others.
It protects minors
It won’t help a minor with alcohol, tobacco, gambling, weapons or adult products, and it never asks for a date of birth.
Documented
The technical documentation is organised following Annex IV of the EU AI Act.
People in charge
Human control: the agent knows when to step aside
A good salesperson knows when to call a colleague. So does the agent.
It escalates when it can’t solve something
If the shopper gets stuck or repeats the question, the agent hands over to a human contact or the store’s support channels.
Nothing runs without confirmation
Any action, such as adding or removing a product from the wishlist, is confirmed first.
Shoppers rate it
Every message can get a thumbs up or down.
The owner sees what happens
The dashboard shows what is being asked, what converts and which demand goes unmet.
Conversation Flow
Category → Intent → Outcome. Width = conversation volume.
Honesty
Where it stands, and its limits
What works today and what it doesn’t do, said plainly.
Works today
- Live in production at AllHub, connected to Shopify, WooCommerce and Wix stores.
- Amazon in a read-only pilot.
- Widget and shareable link, in 12 languages, with voice dictation.
- If a store runs out of quota, shoppers still get answers and relevant cards, and never see billing notices.
Limits
- It is pre-sales only: it doesn’t look up orders or handle after-sales. Faced with “where is my order?”, it points to support.
- It doesn’t process payments, by design.
- Continuous voice conversation is switched off for now.
- Impact metrics such as conversion or questions resolved have not been published yet. That’s why you won’t find figures on this page.
Being upfront about limits is part of the design: an agent that says what it won’t do earns more trust than one that promises everything.
For your company
What this case shows
Putting AI in online shopping in front of real customers isn’t solved with a good prompt, and neither are any other conversational agents. You have to decide what the agent can do, what it will never do, what gets logged and when it steps aside. That is the work I put into every custom AI agent, whether it faces customers or runs inside the company.
When the challenge is a whole business process, I start by mapping how it works today with the MAPA method, and only then do we decide which part deserves an agent.
What sits behind it
- Features designed to sell without pushing or making things up.
- Layered security, with least privilege and deny by default.
- Privacy by design: sensitive data never reaches the AI.
- Verifiable transparency: every reply can be rebuilt.
- People in charge whenever the agent shouldn’t decide.
FAQ
Frequently asked questions
What is an AI shopping assistant?
An AI shopping assistant is conversational AI built to sell: an AI-powered shopping assistant that helps people buy from an online store: it understands what customers ask for in their own words, shows them real catalogue products with price and stock, answers their questions and takes them to checkout. Unlike a search box, it asks, compares and explains why it recommends each product.
How is it different from conversational AI for customer service?
Conversational AI for customer service answers questions. This agent sells with a method: it detects where the shopper is in the purchase, shows product cards they can buy and logs the demand the store doesn’t cover. And it works within strict limits: it doesn’t invent prices, ask for payment details or touch orders.
Can the agent take payments or manage orders?
No, by design. It finds and recommends products, and the purchase always happens in the store’s checkout. It never asks for or accepts card details, and it doesn’t look up or change orders, refunds or inventory.
What happens to shoppers’ personal data?
Sensitive data such as card numbers, IBANs or ID documents is blocked before it reaches the AI. Conversations are stored pseudonymised and deleted automatically after about six months, and the owner gets analytics without raw personal data.
Does it comply with the EU AI Act?
It is designed around the Act’s transparency obligations: it identifies itself as an AI when asked, logs every decision so it can be rebuilt, and its technical documentation follows Annex IV. It isn’t a certification; it’s a design aligned with the regulation from the start.
Can you build an agent like this for my store or company?
Yes. I design custom AI agents for specific tasks, customer-facing or internal, with the same rules on security, privacy and human control. We start with a call to understand your case, and I’ll tell you honestly whether an agent makes sense.
Custom AI agents
Want an agent like this in your business?
Tell me which task you want to solve. On a call we’ll see whether a custom AI agent makes sense and how I would design it, with your sign-off at every stage.
Original source
The Conversational Storefront product page on allhub.io describes the agent as AllHub presents it. This page explains how it is designed inside.
allhub.io/en/ai-agents/conversational-storefrontThis article was created with the help of AI and reviewed by José Galán. I take great care over every post and every translation, but the odd mistake can still slip through. If you find one, write to me: you will be helping me improve.