Live chat, AI chatbot or AI agent: what does your online store need?

Choose live chat when a person needs to make a judgement. Use an AI chatbot for recurring questions that your published information can answer. Add actions when the conversation needs a current record or a change in another system. Most small stores need a combination, with a clear path from automation to a person.

The labels are less useful than the actual work. A product called an “AI agent” might only answer questions. Another product sold as a “chatbot” might also check orders and propose bookings. When comparing tools, ask what each can read, what it can change and who remains responsible when something goes wrong.

Three ways to handle the same conversation

For this guide, live chat means a conversation answered by a person. An AI chatbot answers using the content you provide. An AI agent with actions can also use configured tools to retrieve information or propose a change. These are working definitions, not three mutually exclusive product categories.

Start by listing the last twenty questions your team answered. Mark which needed a public fact, a private record or a decision. This small exercise is more informative than selecting a product by the number of features on its homepage.

Situation 1: “Do you deliver to Ireland?”

This is a good knowledge-base question. The answer should come from the current shipping policy, including destinations, methods, prices and exceptions. A person can answer it, but does not need to make a new decision each time.

The important dependency is content quality. If a PDF says delivery is free above €40 and the website says €50, adding a more capable agent does not resolve the contradiction. Someone must establish the correct rule and update the sources.

Test more than the obvious wording. Ask about a discounted basket, an oversized item and a different destination. A useful answer applies the conditions or asks for a missing detail. See the ecommerce FAQ template for a structure you can adapt.

Situation 2: “Where is order 1042?”

An explanation of your usual dispatch time cannot establish whether this particular order has shipped. The conversation needs a current record from the order system, or someone who can inspect it.

A read action can retrieve that record when you have configured a suitable connection. Private records also need a real access rule. Asking the visitor to enter an email is different from verifying that they control it, and verifying an address still does not establish that an arbitrary order belongs to them. The lookup must be restricted to the right customer.

In Breezaro, Custom Actions connect configured HTTPS endpoints to the assistant. The setup includes response-field selection and options for visitor verification and ownership checks. Those controls need to match the endpoint you connect. The WooCommerce order-status guide walks through a concrete example.

An empty result should lead to a useful next step: check the order number, explain the limitation or involve a person. It should not become an invented shipping update.

Situation 3: “Cancel it and refund me”

This contains both an action request and a business decision. Before automating it, establish which cancellations are allowed, who approves exceptions and what the connected system can actually do. A friendly message saying “Done” is not evidence that a refund exists.

For any supported change, the visitor should see the intended operation and its details before committing. The system then needs to return a result that distinguishes success from failure. If the result is uncertain, the conversation must say so and offer help.

Breezaro's data-changing Custom Actions require a confirmation button in the web widget. They are not offered on WhatsApp, Messenger or Instagram, where that confirmation card is unavailable. This mechanism does not create a ready-made refund integration: the action still needs a suitable endpoint and your business rules. For complex complaints, human takeover may be the better first choice.

What a small team should evaluate

Availability. Live chat needs someone to answer it. Automated replies can cover questions outside staffed hours, but an escalation does not make an offline colleague available. Set expectations in the message shown to visitors.

Maintenance. Count the time spent updating sources, checking conversations and repairing failed connections. A bot with poor material creates more support work, regardless of the subscription price.

Control. Check whether the operator can see the preceding conversation, take over and stop automated replies. In Breezaro, the Take Control documentation explains the human side of that workflow. Confirm feature availability for your plan in the current pricing.

Evidence. Test the tool with your own questions and agreed expected answers. Include a missing answer, a customer requesting a person and a failed lookup. An attractive demonstration is useful; these cases show what your team will need to manage.

Start with one complete support journey

A sensible first release is a bot that answers delivery and returns questions from clean sources, searches a connected product catalog and can direct unresolved cases to your team. Add one live lookup when that workflow is working. Add changes to external systems only when you can verify the full path from request to recorded result.

You can explore that combination in Breezaro, using your own content and a short test set. The chatbot testing checklist provides cases for the next step. Choose the level of automation that completes the customer's task and that your team can confidently maintain.