Chatbots vs AI agents: how SMEs should choose support automation
The difference is where the reply stops. A chatbot only sends text. It is an answering machine plus search: the guest types a line, and the bot returns a fixed sentence or a passage from the files you gave it. It does not change an order. An agent can also act inside the scope you allow: look up one record, send one notice, or finish one fixed flow. Product names change. When a small firm chooses, look at whether your own process crosses systems. Do not buy because the deck says “agent.”
Why this keeps coming up
More support teams now run at least one system that handles conversation. Industry write-ups often say many teams already have some kind of auto-reply or agent. Those figures move with the sample. An owner should not set this quarter’s budget from one headline. The question for you is whether the counter still drowns in the same questions at peak hour. Product circles keep separating agents that take actions from plain chat. You can watch account and product updates on @OpenAI and @thsottiaux. What you ship still follows your process: which questions stay text-only, and which actions wait for a person to confirm.
Three levels — do not buy them as one SKU
1. Menus and fixed replies
The guest presses 1 for the address and 2 for a person. This layer is cheap and easier to control, because every sentence was written by you in advance. It barely understands free text. “Where is my order?” often misses the menu. It fits a unit whose facts rarely change: address, hours, how to get there. Once exceptions pile up, this layer is not enough.
2. Knowledge-base support
A knowledge-base bot answers only from FAQs, price lists, and hours you have approved. It does not fill gaps from the open web. It fits a firm whose questions repeat and whose answers already live in files. LINE Official Accounts often start here. The bottleneck is not how new the model is. It is whether the person you named edits the files every week. If nobody edits, the bot keeps quoting the old price. Questions the file does not cover go to a person. Do not let the bot add a sentence that only sounds true.
3. Agents that act
An agent looks up an order, changes a time slot, or opens a ticket only inside the permission you drew. It has to connect to a back office or another system. Permissions and failure paths are designed first. Do not drop in a plugin and call it done. Write the sentence it must use when a lookup misses, and whether it may guess a nearby order number. Actions the rule does not name stay off.
Write this on paper before you buy
The owner or the support lead writes this page. A vendor deck does not replace it. Decide which level to buy after the page exists. Channel, permission, and who switches the bot off are the lines decks usually skip, so write them here too.
- List the top 20 questions guests actually ask. Mark refunds and disputes as never auto-answered. Those go to a person.
- Knowledge comes only from approved FAQs, prices, and hours. Name one person to edit weekly, and write the edit date on the file.
- Write when a chat must hand off, where after-hours messages are stored, and who replies the next day.
- If you need order or stock lookup, name the system the data comes from, and whether the agent may only read or may also write. If write is not on the page, it may only read.
- After go-live, the same person reads two tables: the handoff rate, and the list of questions the bot could not answer. Next week those questions go into the file.
- Pilot one channel first, for example LINE only. Do not connect the site, the social page, and email on day one.
- Write the name of the person who may turn auto-replies off. Do not hunt for that person after a complaint arrives.
How this sits with the site and the back office
If the support bot, the website, orders, and membership each keep their own copy, the books will not match later. On this kind of job we draw where each field is read from, and which table the reply writes back to, before we talk about a model or a plugin. Ship the knowledge base, confirm the handoff actually runs, then add lookup. That is usually steadier than buying “full auto support” in one shot. See AI services and the LINE checklist.
FAQs
Do we need a full agent?
Not always. If the questions repeat and the answers already sit in approved files, a knowledge base plus a human handoff is enough. Add an agent when you must look up an order, change a booking, or write into another system, and only inside the scope you set.
Will it invent answers?
It can, if you leave the knowledge scope open. Limit it to approved files. If the file has no answer, hand the guest to a person. Each week the owner adds those misses to the file. Do not let it invent a sentence on the spot.
How long to see results?
Open one channel first, for example LINE. Run two to four weeks. The person in charge reads the handoff rate and the miss list, then decides whether to add order lookup. Do not add features in the first week.
Scope a support pilot AI agent overview
Product chatter: @OpenAI, @thsottiaux. Adoption percentages belong to each survey’s primary source. This post is a selection guide only.