Notes

From the counter to the hall to the court: which action should AI land on first?

2026-11-14 · Yunqing's OPC

When you land AI in a concrete place, the first choice is not the model. It is the action. The action must be small enough to repeat in a week, and large enough that finishing it is visible. A counter, a hall, and a court do not look alike. The method for choosing the action is the same. Choose badly, and you get a brilliant internal demo. Choose well, and four weeks later someone in the job does one less overtime task.

In a supermarket or bakery, do not start with “make the whole store intelligent.” A better first action is often the weekly poster and price tags: a short list of items to change, prices that must match the system, and a store manager who can review the work. Run it for four weeks and watch only two things: whether the draft is still slipping to Wednesday night, and whether a published price is wrong again. Draft answers to frequent questions can wait. When the official wording changes and no one reviews, the tool will get it wrong. Stock answers are the same: until permissions are clear, do not let a tool tell the public whether an item is still on the shelf.

In a park, a tourism site, or a civic exhibition, a better first action is often one stop of a tour script, not a digital person for the whole route. Write down the three objects this stop must mention, the history or figures that must not be improvised, and who may change the text. A weekly new-media calendar can run in parallel, but still with one channel and one rhythm. “Every platform must post daily” will break a pilot. Civic tone, especially, cannot be improvised on the spot. Get one stop right, then extend the array to the next stop. That survives inspection better than a full rollout.

At a tennis or pickleball venue, the first action can be as small as an event notice and a court reminder: when, which court, who it is for, what to bring, and where to look if the weather changes. Automatic tagging of training clips sounds attractive, but it usually needs steady filming and consent rules first. It is a poor first pilot. Let sign-ups and reminders go wrong less often, then talk about stroke analysis. Members will not renew because the model is new. They will get less angry when notices are accurate and changes are clear.

The three scenes share one checklist: is the action repeating, is the result observable, is the reviewer in the role, and can the materials keep up with change. If you choose well, after four weeks you will have an ordinary sentence, such as “posters no longer mean overtime to fix prices,” or “the tour no longer names a closed exhibit.” That sentence is closer to landing than a page of demo screenshots. What Solo Array is for is to help you use this checklist in a real place, and to stay with it for four weeks, instead of being pulled away by the next, prettier demo. Without those four weeks of evidence, it is too early to talk about the next model or the next platform. Finish this one action first.

Back to notes