Ecommerce AI agent
An ecommerce AI agent that shops your catalog and fills the basket
It reads what a shopper wants, searches your live catalog, checks the results and builds the basket. The shopper confirms before anything is ordered.
In short
An ecommerce AI agent is software that shops a store's catalog for a customer: it searches, checks what it found, adds to the basket and prepares the checkout. Hephon Commerce is an ecommerce agent for supermarkets, grocery delivery, quick commerce and other retailers with a product catalog.
It works inside the retailer's own apps and website, on its WhatsApp Business number and on phone calls, in Arabic, Arabizi, English and French, and it never places an order without the shopper's confirmation.
Updated
What is an ecommerce AI agent?
An ecommerce agent is the difference between being told where the milk is and having it put in your basket. Three kinds of software get called "AI for ecommerce", and they do different jobs.
A chatbot answers questions. It is usually trained on help-centre pages and scripted flows: opening hours, delivery fees, return policy. It talks; it does not shop.
Site search ranks products for what a shopper types in the search bar. It is fast and essential, and it is built for short keywords. "Snacks for six people watching the game under 20 dollars" is not a keyword.
An ecommerce AI agent uses tools. It searches the live catalog with the shopper's branch prices and stock, checks that what it found matches the request, adds products to the basket, changes quantities and builds a checkout. This is what people mean by agentic shopping: the agent does the legwork, and the shopper approves the result.
| Chatbot | Site search | Ecommerce AI agent | |
|---|---|---|---|
| What it does | Answers questions | Ranks results for a typed query | Searches, checks, adds to the basket, builds the checkout |
| What it reads | Help-centre articles, scripted flows | The catalog index | The live catalog, the shopper's branch, the basket |
| Input it handles best | Common questions | Short keywords | Sentences: lists, budgets, recipes, follow-ups |
| Changes the basket | Rarely | No | Yes, and only orders after the shopper confirms |
What should an ecommerce agent automate first?
In grocery, start with the requests that are slow for shoppers and that a search bar handles badly.
- Search in shoppers' words
- Requests as people write them: Arabic, Arabizi, English, misspelled, mixed. "7aleeb" finds milk; "tomato" finds Tomatoes.
- Lists in one message
- "Milk, bread and eggs" returns a row per item, so the shopper picks each without asking three times.
- Recipes to baskets
- A dish and a number of people becomes ingredient lines, the right packs and one tap to add them all.
- Budgets and cheapest
- "Under 50 riyals" and "cheapest first" are understood and held for follow-ups.
- Orders on WhatsApp
- Product cards, a basket and a checkout the shopper confirms with a tap on your own number.
- Orders by phone
- A caller orders by voice; the basket is read back and the keypad confirms.
Grocery baskets are long, repeat often and depend on details a search bar ignores: pack sizes, the branch that will deliver, budgets, and what a family is cooking tonight. Those are the requests where an agent saves the most time, because each one would otherwise take several searches and taps.
Leave to people what needs judgement or authority: complaints, a late or wrong delivery, refunds, a large catering order, anything to do with an account or a payment dispute. The assistant can recognise these and hand the conversation to your team, who see it in the same inbox with everything said so far. A shopper can ask for a person at any time.
How does it act safely?
An agent that can change a basket needs harder limits than a chatbot that only talks. Hephon Commerce puts them in the server, not only in the instructions to the model, so no wording from a shopper can talk its way past them.
The shopper confirms every order
No tool the model can call places an order. The agent fills the basket, in sync with your app's cart, and builds a checkout. The order is placed only by the shopper's explicit confirmation: a tap on WhatsApp, a signed checkout page, or pressing 1 on the keypad after the basket is read back on a call. A typed or spoken "yes" alone never places an order. The order then goes to your platform through its API, or to your team as an order ticket.
Your rules are enforced on the server
A retailer can hide alcohol, pork and tobacco, or all three for a halal store. Those products are removed by the server before the agent sees the catalog, on every channel. Each product is judged by an AI model, not a word list, and a product that has not been checked yet stays hidden. On WhatsApp, alcohol and tobacco are always excluded. Out-of-stock products cannot be added, and prices and stock come from the shopper's own branch.
It is tested on your catalog first
Before launch, and before any change to the prompt, the model or search goes live, a test suite generated from your own catalog runs: every department covered, retrieval of real products, basket changes, the language of each reply, safety, and WhatsApp flows. Generic benchmarks say little about whether an agent finds your 1 L Almarai milk when a shopper writes "7aleeb".
Off-topic results are dropped
Every search's top results are checked against the shopper's actual message before they are shown. A request for snacks for a basketball night shows snacks, not basketballs. The agent also never assumes food: a store that sells phones gets phones.
Support agents, shopping agents and search engines compared
Most retailers need more than one. This is what each kind is built for, stated generally; check each vendor's current features.
| Support agents (e.g. Gorgias, Zendesk AI, Intercom Fin) | Shopping agents (e.g. Hephon Commerce) | Search engines (e.g. Constructor, Algolia) | |
|---|---|---|---|
| Main job | Resolve customer service questions and tickets | Turn a shopper's request into a basket | Rank products in the search bar and category pages |
| Typical questions | Where is my order? Can I return this? | Kabsa for 6. Cheapest milk. Snacks under 50 riyals. | A typed keyword: milk, rice, shampoo |
| Works from | Help centre, policies, order data, tickets | Live catalog, branch prices and stock, basket | Catalog index, behaviour data, merchandising rules |
| Acts on the basket | Usually not | Yes: adds, changes quantities, builds checkout | No, the shopper clicks |
| Where it lives | Help widget, email, social inboxes | Your app, website, WhatsApp, phone calls | Your site and app search |
| Where it is stronger | Ticketing, agent workflows, support analytics | Multi-item requests, recipes, budgets, dialects, voice | Speed at scale, merchandising control, search analytics |
| Use it alongside the others? | Yes, for after the order | Yes, for shopping in conversation | Yes, for the search bar |
Support agents such as Gorgias, Zendesk AI and Intercom Fin are strong at what happens after the order: tracking questions, returns, policies, routing tickets to the right person, and reporting on support quality. If most of your chat volume is "where is my order", start there.
Search engines such as Constructor and Algolia power the search bar and category pages at large scale, with merchandising controls and analytics a team can tune. They are the right tool for the millions of short queries a big store gets, and both offer conversational features of their own.
A shopping agent sits in the conversation: it takes a sentence, a voice note or a call and turns it into a basket. Hephon Commerce is built for that, for grocery in particular, and for shoppers who write and speak Gulf, Saudi or Lebanese Arabic, Arabizi and English.
Where does the agent live?
Shoppers should not have to go somewhere new. The same agent, with one basket, works in three places:
- Your app and website. An SDK for iOS, Android and the web, with React Native and Flutter, plus a hosted chat page per retailer. The basket stays in sync with your app's cart.
- Your WhatsApp Business number. Product cards, a basket, a checkout confirmed with a tap, voice notes and WhatsApp calls.
- Phone calls. Voice ordering in Gulf Arabic and more than 30 languages, with replies in under a second and a person on hand for any caller who wants one.
When Hephon Commerce is not the right fit
- You have no online catalog. The agent works from your live products, prices and stock. Without them it has nothing to shop.
- You want to replace site search. Hephon Commerce does not power your search bar or category pages.
- You want self-serve sign-up. Access is by invitation, after a demo on your own catalog.
Live in three steps.
- 1
Connect your catalog
Shopify, WooCommerce, a product feed or your API, with each branch's prices and stock.
- 2
Set the rules and test
Persona, languages, products you never sell, handover. Then a test suite built from your own catalog.
- 3
Go live channel by channel
In your app first, or WhatsApp first; add phone orders when you are ready.
Ecommerce AI agents, answered.
See it on your own catalog.
We set up a demo with your products, your branches and your shoppers' languages.
