Does Your Mini Program Need an AI Assistant?

Somebody suggests putting an AI assistant in the Mini Program, and the conversation immediately becomes a conversation about models. Which one, how good is the Chinese, can it handle follow-up questions.

Wrong conversation. The useful one starts from a question with a checkable answer: what happens to a Chinese enquiry that arrives at three in the morning your time, today, without an assistant? Whatever that is, the assistant has to beat it.

The short version: an assistant is a shift, not a feature, and it earns its place only where three things hold — you have a real coverage gap, most of your questions have stable factual answers, and somebody owns keeping those answers current. Judge it on what it does when it does not know something, because most of what it will face lands there. And treat the log of what people asked it as the actual product; it tells you what your pages failed to answer — worth more than the replies.

A shift, not a feature

An assistant does not sell anything. What it does is occupy a seat during hours when the seat is empty, and answer the questions a person would have answered.

That reframing kills most of the bad versions of this project immediately. If your team already replies inside an hour during Chinese browsing hours, an assistant is competing with your own staff and will lose. If the first human reply currently lands fourteen hours later, it is competing with silence, and almost anything beats silence.

Which arithmetic applies to you is not a matter of opinion. China runs on UTC+8, the overlap with European and American working hours is narrow or absent, and we worked the numbers through in staffing WeChat enquiries across time zones. Do that first. The staffing answer determines whether there is a job here at all.

The three conditions

A gap the assistant actually covers. Not "we could be faster" — a stretch of hours where nobody replies. Count them honestly across a week, including weekends, where the real hole usually sits.

Questions with stable answers. Open your last fifty Chinese enquiries and sort them. One pile will be things like: does it run in January, how far is the walk, do you collect from hotels in this district, is there a Mandarin-speaking guide, what happens if it rains. Those have answers that stay true for a season. The other pile reads: we are five people, two of them in their seventies, arriving on the 14th, can you make something work. Nothing in a knowledge base answers that, and an assistant that tries produces something confident and useless.

The ratio between those two piles is your answer. Most tourism operators find the first pile is larger than they expected, and that pile is the argument for doing this. Some find the opposite. A real result, and it should stop the project.

An owner for the source of truth. The assistant knows what you gave it. When the winter timetable changes and nobody updates the material behind it, the assistant keeps answering from the old one, in fluent Chinese, with complete confidence. Name the person who owns that before you build anything.

Judge it on what it does when it does not know

Most of what an assistant meets is outside its knowledge, so the handling of that case is the product.

What good looks like is unglamorous: it says it cannot answer that one, it takes the enquiry, and it states when a person will reply. Three sentences. A traveler who gets that at 3am has been served — they know the message landed, they know when to expect an answer, and they have not been left guessing whether anybody is there.

What bad looks like is an assistant that improvises, or one that loops, or one with no route to a human at all. A dead end at 3am is worse than no assistant, because the traveler now believes they have tried.

Two design rules follow. The assistant is never the only path — the enquiry form stays exactly where it was. And it should be scoped narrowly: the twenty or thirty questions you genuinely receive, answered from your own material. Breadth is what makes these things unreliable, and nobody is asking your tour company about the weather in Lisbon.

The log is the real product

Here is the part most operators do not anticipate. The transcript of what people asked is more valuable than the answers.

Every question that arrives is a question your pages failed to answer. A week of logs tells you exactly which facts are missing, in the words your customers actually use, which is information you cannot get any other way — people do not write in to tell you your page is unclear, they just leave.

We run an assistant on this site, so a concrete example: reading our own logs showed that when it routed people to the wrong article, the model was rarely the problem. The index behind it described each post in our own formal phrasing instead of the words a client would type. Rewriting those descriptions fixed the routing. The log surfaced that; nothing else would have.

Read the logs weekly for the first month. Most of what you find belongs on a page, not in the assistant.

When it does not earn its place

• During hours you are already covering. It will be worse than your team and it will intercept people who would have reached them.
• When most answers are "it depends". Bespoke, consultative products do not decompose into a knowledge base, and pretending otherwise produces answers that sound settled and are not.
• When nobody will maintain it. An unmaintained assistant degrades into a liability on a schedule set by how often your details change.
• As a signal that you are modern. Nobody has ever booked a tour because the operator had a chatbot.

The failure worth its own treatment

An assistant can state something false in perfect Chinese — a price, an inclusion, a date — to a customer with no reason to doubt it. That is a different subject with its own mechanics and its own remedies, and it deserves more room than a paragraph here. We will take it on its own.

For now the relevant part is narrow: decide in advance what the assistant is allowed to state as fact, and keep prices and availability out of that set unless they come from a live source.

The honest edge of this

How well any of this works depends on your material, your question mix and your maintenance discipline, none of which we can see from here. We have no benchmark to offer you and would not trust one that was offered to us — the useful measurement is your own log after a month.

What does hold: an assistant cannot know anything you did not give it, and it cannot be held responsible for what it says. Both of those are yours.

Where CN1X fits

We build the Mini Program and we build this when it belongs in it, which starts with reading your last fifty enquiries and telling you whether the split supports the idea. Sometimes it does not, and that is a shorter conversation.

What we will not do is put a number on how many enquiries it deflects. Attribution inside the Chinese ecosystem is genuinely broken, and any figure we quoted would be part measured and part invented — we would rather tell you which. If you want the enquiry split looked at before anybody builds anything, send us a season's worth.