Magento 2 AI Chatbot & Live Chat – eChat

Mostly, and you don't have to choose. Across 3,766 production conversations handled with no human agent behind them, only 0.66% of substantive conversations asked for a live person. When most requests come down to finding information and handing it to the customer, the AI covers the channel.

For the rest, live chat is built in: your team takes over a conversation in the same widget, and the AI drafts replies for them. We see the AI as an information helper, not a sales consultant. If your chats are complex and unique, or your staff actively sell through chat, today's models don't reach the level of human communication, so let your team take over those conversations.

Yes. Turn on live chat in the eChat backend, and your team sees who is on the store right now, takes over any conversation, and replies in the same widget, with files and product cards. The AI stays silent while a person answers and resumes when the conversation is handed back.

You choose whether the AI answers first and your team joins when a shopper asks for a person, or your team answers while someone is available. When nobody is online, shoppers can leave a message with their email. There are no per-agent fees; the manual covers the setup.

Product questions dominate, which surprises merchants who expect the chat to offload support. In our production data, first messages broke down like this:

• 38% product selection, specs, and compatibility
• 25% greetings and small talk
• 29% other topics such as policies, returns, and contacts
• 8% order status and delivery details

Order status stays low because a logged-in Magento shopper is two or three clicks away from it. The chat works as a product-consultation channel, not only as support.

It fully resolved about 50% of conversations and fully or partially resolved about 85%. We deliberately keep those as two numbers instead of collapsing them into one smooth resolution rate, because full resolution and partial help are different things.

One caveat on the method: the conversations were scored automatically by a separate, more capable model, not read by hand. Only about 5% of shoppers explicitly wrote that the bot had helped. The full methodology is published here.

Cents. In our production data the median conversation, usually three to five question-and-answer pairs, came to $0.028 of model cost. Your AI provider bills that on your own API key; eChat adds no per-conversation fee of its own. The driver is not message length but how much information the bot has to read before it can answer.

One question is rarely one model call, so a hard product question costs more than a policy question. The heaviest conversations in our data were product consultations, the ones you least want cut short, and they still stayed within single dollars, which is why we don't recommend capping a conversation by cost. Abuse is a separate problem with a separate answer: dedicated layers cut off text dumps and attempts to pull the chat away from your store's topic.

Hallucinations are mostly a reaction to a gap in your content: unable to find the information, a model starts filling it in. eChat doesn't rely on instructions alone to prevent that. An answer passes through a guard layer that checks its facts against your knowledge base before the shopper sees it.

Even that layer isn't perfect. In production it lets a false answer through roughly once every thousand conversations, and we would rather name that number than pretend it's zero. Every conversation is in your history, and the Correct action on a wrong answer turns the right one into an FAQ entry or a Guidance rule, so you fix the source.

The steps are short: install the extension with Composer, connect it to eChat from the Magento admin, add an API key from your AI provider (OpenAI, Anthropic, Google, or xAI), and run the first sync or wait for the nightly one. Nothing needs training: the assistant answers from your synced content as soon as the sync finishes.

A large catalog takes longer to sync, and the content work described below matters more than the setup. The quick start walks through each step.

Mostly content. Over 90% of what the bot knows is your catalog itself, and the remaining 10% comes from CMS pages and content you add on purpose. Two questions tell you how much work is left: what do customers ask about most, and is the answer already in your admin panel?

If your product descriptions are complete and detailed, you may need no preparation at all. If support answers customers from a source the site has never seen, that source has to reach the bot's knowledge before launch, not after, or the bot will expose the gap in its first week. The information most often missing is unglamorous: live stock, prices and promos, store addresses, shipping terms, and self-service instructions.

The language is detected from the shopper's first word, and the whole conversation runs in it with no configuration on your side. In production, shoppers chatted in 30+ languages while the catalogs and knowledge bases behind them stayed in English. When the language can't be detected, the assistant falls back to a preferred language you pick from a list of 65.

Stores with several store views connect each one as its own workspace, with its own synced content and texts, as the multi-store guide explains.

Add a second provider as a fallback, and eChat switches to it when the first key runs out of quota, hits a rate limit, or is revoked. If every provider fails, the assistant switches itself off instead of showing shoppers errors, and you can have eChat email you when that happens.

The eChat service runs in Canada, which the EU recognizes as giving personal data adequate protection, and replies come from the AI provider you connect with your own API key. The nightly sync sends your catalog, not your customers or orders. Order details are looked up live, only when a shopper asks, and guests confirm their email first.

You decide how long conversations are kept, can delete any conversation or erase a shopper by email for an erasure request, and can hold the widget back until cookie consent (Cookiebot, Magento's cookie notice, or any other consent tool). A Data Processing Agreement is published at echat.mirasvit.com/dpa, and the manual lists exactly what is stored and where.