For years a mini-program was a small app: a few screens, a form, a list. You tapped, it responded, end of story. Then language models got small enough and cheap enough to put behind a tap, and the format quietly changed. A mini-program can now hold a conversation — answer a question, take an order in plain language, walk a user through a return at 11 pm. The screen stopped being the limit.

The teams experimenting here aren't building chatbots for the sake of it. They're moving tasks that used to need a menu and three taps into a single sentence. That's a real shift, and it's why "AI mini-program" stopped being a conference buzzword and started being a build decision.

What actually works today

Most of the impressive demos don't survive contact with real users. The patterns that do are narrower and more useful.

Conversational commerce. "Find me a red dress under forty dollars that ships tomorrow" is a query a search box can't take but a model can. The mini-program turns it into a filtered grid. The model does the parsing; the store does the selling. This is where AI earns its keep instead of just impressing.

On-demand help. A support mini-program that answers "where's my order" from order data, in the user's language, at any hour, removes a whole support tier. It's not glamorous, but it's the kind of thing that keeps a user.

Guided tasks. Returns, onboarding, picking a plan — anything with branching logic that normally means a wizard of eight screens — collapse into a back-and-forth. Users tolerate the conversation because it's shorter than the form.

Use caseWhat the model doesWhy it fits a mini-program
Conversational commerceParses a natural-language query into a filtered resultThe host already has the catalog and payment
On-demand helpAnswers from live order and account dataAlways open, in the user's language
Guided tasksReplaces an 8-screen wizard with a chatShorter than the form users dread

The two things that bite

AI in a mini-program lives or dies on two numbers. Latency — the user is on a phone, mid-task, and a three-second "thinking" pause reads as broken. And token cost — every conversation is spend, and a chatty model with no guardrails burns money on pleasantries. A shared runtime helps with the first: a pre-warmed engine and fast first screen mean the model's response is the only wait, not the app's startup on top of it. The second is on you — cap turns, cache, and keep the model doing work, not small talk.

The tradeoff

You're adding a dependency that can say something wrong, and you're paying per interaction. The teams that ship treat the model as a junior clerk with a manager: it handles the obvious cases, and the messy ones get a human or a clear exit. Done that way, an AI mini-program isn't a gimmick. It's the same small app you already shipped, except now it can talk.

Sources