How SITION runs on AI
SITION leaves most of the gathering, writing, publishing and checking for its three media to AI, and people hold only the direction and the final decisions. In this feature series, “AI, Starting from the Work”, we lay out, as procedures, how to build AI into each kind of work a business does. Before that, we show how we run ourselves, with numbers we actually counted.
■ What we run
SITION runs three media, each with a different character.
| Media | What it covers | Articles published in the last 30 days |
|---|---|---|
| LiveMakers | Financial intelligence that reads the next move in finance from leading indicators | 592 |
| SIPO.TOKYO | Publishing as a Cardano stake pool and DRep | 250 |
| LifeMakers | Lifestyle media on living and technology | The latest 50 articles span about 5 days (around 10 a day) |
Of LiveMakers' 592 articles, 160 are breaking news, 332 are news analysis, 30 are morning round-ups, 26 are long-form analysis, and so on. Per day, that comes to around 40 articles across the three media.
What supports this is not an editorial room with many reporters. It is AI, the rules that AI follows, and a person who makes the final decisions.
■ The workflow, and who does what
Broken down, the making of one article follows this flow.
| Stage | What happens | Who |
|---|---|---|
| Gather | Sweep official announcements, public institutions, news reports, X and more, five times a day | AI |
| Select | Rank by importance and freshness for readers, and remove duplicates of the same topic | AI |
| Verify | Check figures, dates and who said what against primary sources | AI (following the rules) |
| Write | Write to each media's style rules | AI |
| Images | Create a new thumbnail for each article | AI |
| Publish | Publish under conditions set for each type of article | AI and people (explained below) |
| Check | Count what was published and where things stopped, and raise alerts on anything abnormal | AI. A person reads the alerts |
What people hold: what each media should write about (direction), whether to start a new series, the articles that need a final decision on whether they can be published, and money and dealings with outside parties.
■ Seen in three levels
This series sorts ways of building in AI into three levels. ① AI drafts, and a person reviews everything. ② Hand it over with checks (a person looks at exceptions and spot checks). ③ Hand it over (a person watches the numbers). Applied to our own operations, it looks like this.
- ③ Hand it over: gathering and ranking information, creating an image for each article, counting what was published. Work where a mistake does not reach readers directly, or can be checked by machine
- ② Hand it over with checks: daily breaking news and news analysis. Only what passes automated checks that follow the rules goes out; anything that fails is stopped and a person is notified
- ① Drafts: long-form analysis, series, and articles that touch on direction. AI writes the draft, and a person reads all of it before it is published
Even within “writing articles”, we separate the levels by how much an article can affect readers. How much we hand over is decided not by the type of work, but by what happens when it goes wrong.
■ What we prepared — data, rules and people
Before handing work to AI, we prepared three things.
- Data: a list of where to take information from (official announcements, public institutions, primary sources), and a record of what we have written before. Without the record, the same story gets written again and again
- Rules: every writing rule is written down as a document. “Check figures against primary sources”, “Do not use figures from news reports as they are”, “Titles are written in headline style”, “Which articles may be published”, and so on. AI may not remember the next day what it was told in a conversation, so the documents are the source of truth for the rules
- People: one person is set as the final decision-maker. Points where the AI is unsure reach that person as short questions
■ What did not go well
The hard part of running on AI turned out to be not building things, but noticing when they stop.
- Stopping silently: when part of the system stops, there are ways of stopping that leave no error record. If you look at the records and conclude “there are no failures”, you notice late. Now we count every day whether the articles that should have come out did, and look first at the days that fall short
- Drifting style: where the rules were not written down, the writing wavered from day to day. We add to the documents as we notice
- Reported figures drifting from the source: figures in news reports sometimes differ from the original documents. That is why “Verify” is a stage of its own
■ How we measure the effect
What we count every day now is the number of articles published and the number stopped. On the other hand, we cannot show in numbers how many hours of human work AI has saved, because we did not record the working time before we started.
That lesson is why this series puts “before you start, measure how the work is done today” as the first step.
■ What this series will cover
From the next episode, each episode takes up one kind of work. We start with work that stays inside the organization, then move on to work that touches people outside, and then to work where money and accuracy matter most.
- Meeting minutes and tracking action items
- Internal help desk
- Sales materials and proposals
- Handling inquiries
- Ordering and order intake
- Monthly accounting
Every episode lays out, as a procedure, how it is done today, how to build AI in, what to prepare, what to watch out for, how to measure the effect, and what to do in the first two weeks.
■ Terms
- AI agent: a system that can be handed several tasks in a row, following a set procedure. Unlike using AI to produce a piece of text once, it carries on through “gather, select, write and notify”
- Primary source: the document information originally comes from, such as official announcements, statistics from public institutions, or laws and regulations. Not news reports or summaries
Where the numbers come from
- Article counts were taken on October 15, 2026 from each site's public information, for the 30 days from September 15 to October 14, 2026
- LiveMakers: articles listed in the sitemap whose URL date falls in that period
- SIPO.TOKYO: articles whose publication date recorded in the site's article list falls in that period
- LifeMakers: because 30 days of publication dates could not be counted from public information, we show the latest 50 articles in the RSS feed (October 10 to 15, 2026)