AI Agents for Ecommerce: What They Are and Where to Start
· 7 min read · AskMerra
"AI agent" has become the label of the year in ecommerce. Helpdesk tools, chat widgets, ad platforms and ChatGPT's shopping features all use it, for very different things. If you run a small online store, that makes a simple question hard to answer: which of these do you actually need, and in what order?
This guide sorts the field into four kinds of agents, explains agentic commerce without the hype, and ends with what every agent needs from you and the questions to ask a vendor. Where it helps, we use AskMerra, the on-site shopping assistant we build, as the example, and we say plainly where it stops.
What makes software an "agent"
A chatbot writes replies. An agent also decides what to do next and acts through tools: it searches a catalog, opens a document, fills a cart or hands a case to a person. The AI model chooses the step, and the tools set the limits of what it can touch.
AskMerra's assistant is a small, concrete example. When a shopper asks something, the model can search the catalog (with price limits and an in-stock filter), fetch one product's full details, read one of your policies, show product cards and hand the conversation to your team. It cannot change prices, issue refunds or place orders. That boundary is a design choice, and it is the first thing to look at in any agent: what can it do on its own?
Four kinds of AI agents a store meets
1. The on-site shopping assistant
It works on your website, before the purchase. A shopper describes a need ("a gift for my dad, who cycles, under 60 euros") and gets products from your catalog with the current price and stock, plus answers about shipping, returns and payment taken from your own policies. It does the job of a good sales assistant at the counter. This is the category AskMerra belongs to, and our article on what an AI shopping assistant is explains how one works, step by step.
2. The customer service agent
It works after the purchase: where is my order, can I change the address, how do I send this back, when do I get my refund? To answer those, it needs access to your order system, your carrier's tracking and often your helpdesk. Some can also act, for example by opening a return.
Be clear about which kind you are buying. AskMerra answers policy questions (how returns work, how long delivery takes) but does not look up individual orders. When a shopper needs help with a specific order, it hands the conversation to your team. Our guide to reducing support tickets about shipping and returns shows how far good policies alone can go.
3. External shopping agents
These do not live on your site at all. A shopper asks an AI assistant such as ChatGPT to find a product, and the assistant searches, compares and, in some cases, checks out on the shopper's behalf. This is what people mean by agentic commerce, covered in the next section.
4. Back-office agents
They work for your team rather than with shoppers: enriching and tagging products, writing or translating descriptions, sorting support emails, reading conversations to spot patterns. They are less visible and often part of another tool. AskMerra's catalog enrichment is one example: when you import products, AI reads each one and extracts what shoppers ask about, such as use cases, skin or hair types, key ingredients, specs and safety notes. You can review and edit the result for every product.
Agentic commerce, without the hype
Agentic commerce means an AI agent shopping for a person, inside the agent's own app. OpenAI's documentation for its Agentic Commerce Protocol is a concrete reference for how this works today:
- Merchants share a product feed (CSV or JSON) with identifiers, descriptions, prices, inventory, media and fulfillment options. Integration starts with a sample feed for validation, followed by daily snapshots.
- Checkout runs on the merchant's systems. ChatGPT displays the checkout, but the merchant validates the order, calculates tax, charges the payment through its own payment provider and accepts or declines the order. OpenAI is not the merchant of record.
- Instant Checkout in ChatGPT is available to approved partners, who apply through OpenAI.
Source: OpenAI, Agentic Commerce key concepts, read on 9 October 2026. Programs like this change often, so check the current terms before you plan around them.
For a small store, the useful conclusion is not "apply tomorrow". It is that every external agent reads the same thing: structured, accurate product data. If your Google Shopping feed is complete and current, you have already done most of the preparation, whichever agent ends up sending you shoppers.
Which one to start with
There is no single answer, but there is a good way to decide: start where your customers already ask questions. Read a month of emails and chat messages and sort them.
| What your inbox and chats show | Start with |
|---|---|
| "Which one should I pick?", "Will it fit?", "Does it work with...?" before people buy | An on-site shopping assistant |
| "Where is my order?" and "How do I return this?" after people buy | Clearer order and tracking emails first, then a service agent connected to your order system |
| Shoppers who mention ChatGPT or other AI assistants | A clean, complete product feed |
| Hundreds of products with thin descriptions and no attributes | Catalog enrichment, which also improves every row above |
The first row deserves extra attention, because pre-sale questions are the ones that stand between a visitor and an order. An on-site assistant also has a practical advantage: it runs on data you already have, your catalog and your policies, and you can measure it. In AskMerra, for example, you can see which sales came after a chat and which suggested products sold.
Whatever you choose, start with one agent, measure it for a few weeks, and only then add the next.
What every agent needs from you
The agents differ. What they need from you is almost the same.
- A clean catalog. Stable product ids, specific titles, real attributes (size, material, compatibility, ingredients), accurate stock and current prices. In AskMerra, price and stock changes apply as soon as they sync, products missing from your latest feed stop being recommended, and out-of-stock products are never suggested.
- Written policies. Shipping costs and times, returns, payment methods and an FAQ, written once and kept current. An agent can only quote what you wrote down.
- A way to reach a person. Every agent meets questions it should not answer alone. Decide where they go: a contact form, email or live chat with your team.
- Limits. A spending cap, rate limits and protection against bots that run up your bill. Our article on protecting your AI chat budget explains the layers AskMerra uses.
- Measurement. A baseline from before you start, and a way to see the orders that followed.
- Honesty with shoppers. An agent that talks to people has to say it is an AI. Our guide to EU AI Act chatbot disclosure explains what that means for a store.
Questions to ask a vendor
Take these to any demo, ours included:
- Where do product answers come from? Can the agent ever show a product you do not sell, or a price you do not charge?
- How fresh is the data? Are prices and stock read live, and how often does the catalog sync?
- What can it do on its own? Add to cart, apply discounts, issue refunds, cancel orders? Which actions need a person's approval?
- What happens when it cannot help? How does the conversation reach your team, and with what context?
- How is it priced? Per seat, per conversation, per resolved ticket or by usage? Is there a hard monthly cap?
- How is abuse handled? What stops a script from sending thousands of messages at your expense?
- Where is the data hosted, and who processes it? Ask for the sub-processor list and the data processing agreement.
- How does it tell shoppers it is an AI? Ask to see it in each of your languages.
- Which languages are included in the plan you are considering? At AskMerra, for example, Start includes two, Basic three, Growth and Plus four, and Pro and above all six.
- How will you measure results? Which numbers will you see after the first month?
A sensible first step
If pre-sale questions are what you see most, an on-site shopping assistant is the natural first agent. Start your 14-day free trial to connect your catalog and policies and test the assistant in the playground before it goes live, or book a demo and bring the questions above. You add a card at checkout and pay nothing if you cancel before the trial ends. The documentation explains how the assistant answers, from routing to escalation.
Put AskMerra to work on your store
Connect your catalog, start your 14-day free trial and test the assistant in the playground before you go live.
Start free trialKeep reading
- What Is an AI Shopping Assistant? How It Works on Your StoreAn AI shopping assistant answers shoppers in their own words and recommends real products. Here is how one works on an online store, from question to sale.
- EU AI Act Chatbot Disclosure: A Guide for Online StoresSince 2 August 2026, the EU AI Act says people must be told when they chat with an AI. What that means for your store's chat, with examples and a checklist.
- Live Chat Meets AI: Let Your Team Take Over When It MattersAskMerra can now hand a chat to your team when a shopper asks for a person: a live queue, opening hours, sounds and a support role that sees only live chat.