Online store support drowns in the same three messages: "where is my order", "can you deliver tomorrow evening instead", "the size is wrong, how do I return it". The owner decides to "add a bot" and quickly discovers that the word covers two very different products. One is a button-based chatbot that walks the customer through a pre-built menu. The other is an AI agent that reads free text, looks up the order itself and takes an action. This guide explains how they differ in practice, which customer requests each one can handle, and how to choose without paying for capability you do not need — or saving money in the one place where you should not.
Two different tools behind one word
A scripted chatbot is a decision tree. The customer taps "Delivery" → "Timing" → "Dubai" and gets a prepared answer. If someone types "can you get it to Marina by tonight?", the bot only copes if the phrase happens to match its keywords. The upside is predictability: it says exactly what you wrote and nothing else.
An AI agent is a language model connected to your data and tools: the order database, delivery and returns policies, the product catalogue, the CRM. It reads a message written however the customer likes, calls the right tool ("find order by phone number"), gets the system's answer and replies in plain language. It can act as well as inform: move a delivery slot, open a return request, or hand the conversation to a person with a ready summary.
The real difference is not intelligence but access. A bot with no access to orders can only repeat your policy. So can an AI agent with no access. The first question is therefore not "which model?" but "which systems are we willing to connect?"
What customers ask, and who can handle it
Export last month's support conversations and sort them by type. The picture usually looks something like this:
| Request | Button bot | AI agent | Stays with a person |
|---|---|---|---|
| Delivery times and fees by area | Fine if pricing is simple | Answers a free-form question with an address | Keeping rates up to date |
| "Where is my order?" | Tracking link only | Finds the order by phone, gives status and courier window | Orders stuck with the carrier |
| Reschedule or change address | Creates a ticket for an operator | Changes the slot if the system allows it and confirms | Orders already out for delivery |
| Returns and exchanges | Shows the policy | Checks eligibility, collects photos and reason, opens a request | Disputed cases and refunds |
| Product choice, sizing, compatibility | Poor: too many branches | Answers from product pages and size charts | Complex or high-value orders |
| Complaints, damage in transit | Not suitable | Records details and photos, escalates immediately | Investigation, compensation, apology |
If most of your volume sits in the first row, buttons will do. If it sits in rows two to four, buttons will mostly frustrate people: they click through the menu and then type "let me talk to a human" anyway.
Returns: where AI helps and where only a person decides
Returns are the sensitive part. A mistake costs money and trust, so the agent's authority should be written down explicitly.
Safe to give the AI agent:
- checking whether the purchase is inside your return window;
- asking for the reason and for photos of the item and packaging;
- explaining how to send it back: courier pickup, drop-off point, in store;
- opening a request in the CRM with order number, reason and attachments;
- reporting the status of an existing request.
Keep with a person:
- approving refunds and any compensation;
- exceptions ("two days late, but a loyal customer");
- damage and defects where someone has to decide who is responsible — the store, the supplier or the carrier;
- conversations with a customer who is already upset.
Return terms come from consumer law in your market and from your own store policy. The agent should read them from your documents rather than "know" them, so it never promises something your policy does not offer. We cover answering strictly from documents in our piece on RAG systems.
What that conversation looks like
Customer: Hi, the trainers run small, I'd like a size up.
AI agent: Hi! Could you share the phone number the order was placed with? I'll look it up.
Customer: +971 50 xxx xxxx
AI agent: Found it: order from 14 October, trainers, size 42. An exchange is possible — you are still within the return window, and size 43 is in stock. Please send a photo of the pair and the box so we can confirm they haven't been worn. A courier can collect them and bring the new pair at the same time; the next slot is tomorrow, 2–6 pm. Does that work?
Customer: Tomorrow is fine. Photos sent.
AI agent: Thank you. I've created the exchange request and passed it to a manager for approval. I'll message you with the request number as soon as it's confirmed.
Notice the agent never says "approved". The final yes belongs to the manager, but it arrives as a complete request, and approving it takes a minute instead of fifteen messages back and forth.
When a simple bot is enough
- A small catalogue and repetitive questions: delivery, payment, store address.
- Orders tracked in a spreadsheet, so there is nothing to connect.
- Low volume, and a manager can reply within the hour.
- The main goal is lead capture — name, phone, product — followed by a call.
A good button bot is a sensible first step. It will also show you which questions fall outside the menu, and that list becomes the brief for an AI agent later.
When an AI agent pays its way
- Customers write long messages or send voice notes, often in more than one language.
- A large share of requests are about a specific order.
- Orders arrive through several channels — website, WhatsApp, Instagram, marketplaces — and answers must be consistent.
- Peaks after promotions and around holidays overwhelm the team.
- Your order system has an API or at least an export, so the agent has somewhere to look.
For channel-specific details, see AI assistants for Instagram DMs and how a chatbot takes over first-line support.
Honest limitations
- It makes mistakes. It can misread a question or answer confidently without data. The safeguards: answer only from the knowledge base, say "let me check with a colleague" instead of guessing, and spot-check conversations.
- It is only as good as your data. If order statuses update once a day, the agent will report yesterday's status.
- Actions need limits. Editing, cancelling and refunding should each be defined: does the agent do it, or only prepare it for a person?
- Someone has to own it. A team member updates policies, reviews escalations and adds missing answers. Without that, quality slips.
- Personal data. Phone numbers, addresses and order history are personal data; decide where they are stored and who can see them before launch.
Pre-launch checklist
- Count what share of last month's requests concerned a specific order.
- Find where order status can be read: store API, CRM, carrier, spreadsheet.
- Put delivery, payment, return and exchange rules into one document.
- Decide which actions the agent performs and which it only prepares.
- Name who receives escalations, and during which hours.
- Agree how you will judge results: response time, share resolved without a person, complaints.
If the checklist points to an AI agent, see how we build AI agents around business processes, browse more industry guides, or get in touch — the cost depends on scope and data, and the first conversation is where we work it out.
FAQ
Can we start with a bot and move to an AI agent later?
Yes, and many stores do. The bot gathers question statistics, and the policy documents you write for it become the agent's knowledge base. Just avoid building a huge tree with hundreds of branches that you will have to replace.
Does an AI agent handle Arabic, Russian or mixed-language messages?
Current models handle most major languages and code-switching well, but test on your own conversations first: take 50–100 typical messages and review the answers before going live.
What if the agent gives a wrong answer about a return?
Three things reduce the risk: answers only from your policies, no final decisions involving money, and a handoff to a person in any dispute. Review conversations closely for the first few weeks.
Does our store need to be on a specific platform?
No. What matters is that order data can be read — through an API, a CRM integration or a regular export. If there is no data at all, start by tidying up order tracking.