How AIDC runs on ADISAIDC 怎样用 ADIS 运营自己
We build a system meant to run companies, so we run our own company on it: code, pricing, research, operations and this website.我们做的是一套用来运营公司的系统，所以先用它运营我们自己：代码、定价、研究、运营，还有这个网站。
October 7, 2026 · 3 min read2026 年 10 月 7 日 · 阅读约 2 分钟
We build a system that is meant to run companies. The most honest test of that claim is to run our own company on it.
So we do. ADIS, the Autonomous Decision Intelligence System behind our products, drafts our code, our pricing, our research and the words on this website, and it reads our own operating data every day. People sign off on all of it. This post describes what that looks like in practice, and what it has taught us about deploying AI anywhere else.
Products: ADIS drafts, people merge
Most of the code in our platform is first written by ADIS. Nothing ships on that basis alone. Every change goes through review and tests before it is merged, and every release is recorded in the public changelog.
The lesson came quickly: once drafting is cheap, review becomes the bottleneck. So we invested where the bottleneck is. Every change runs the full test suite, every fix to an API comes with a regression test and an entry in a fix log, and independent reviews look for problems before a person does. The goal is not to write more code. It is to make every merge one a person can approve with confidence.
Pricing: drafted by ADIS, signed by the owner
Our pricing standard, three systems with a build fee and a per-person subscription, was drafted by ADIS and signed off by the owner of the company. It is public.
When you ask for an estimate on our website, ADIS reads your company's profile and the price comes from the published pricing model, not from a negotiation. The same numbers appear in the estimate, on the pricing page and in the answers Adis gives in chat, because they come from one source.
Research: experiments by ADIS, review by researchers
AIDC Research publishes work on agent systems, organisations of agents and the economics of AI. ADIS runs the experiments and writes the first drafts. Researchers review the method, the numbers and the claims. Every paper ships with its data and the code behind every number, so anyone can check our work, including us, later.
Operations: one dataset for the whole company
Our cloud bills, our cloud resources and our development work all live in Semantic, the same data layer we deploy at customers. Each requirement is an object, linked to the pull requests that implement it and to the person who accepted it.
That single dataset is what lets ADIS be useful in operations. It starts every review from the same numbers we do: which service costs more than last month, which resource nobody has touched in weeks, which piece of work has been waiting for acceptance. It proposes; someone decides whether to switch something off.
This website
ADIS organised this website and wrote its copy, in four languages, and it went live after sign-off. Ask Adis, the assistant in the corner of every page, answers questions about AIDC in seconds and sends the material that fits the question. When a visitor wants to talk to a person, it hands over to us.
The best argument we can make to a customer is not a demo. It is that we would not run our own company any other way.
What running on ADIS has taught us
Ownership has to be explicit. Every area above has a named person who signs off. When ownership was vague, quality drifted; when it was named, it held.
One set of numbers changes the conversation. Cost discussions got shorter the day our bills and resources lived in the same place as our work.
Corrections compound. Each time someone corrects a draft, the correction becomes part of how ADIS works next time. After a few months, the corrections we make are about judgement, not about format.
Review is the job. When drafting is nearly free, the scarce skill is knowing what good looks like and saying so clearly. That is the skill we hire for, and it is the skill we help customers build in their own teams.
Not everything we sell has a counterpart inside a company of our size. But the core of what we deploy, one dataset, digital employees with owners, actions with approvals and sign-offs on a rhythm, runs here first, every day.
我们做的是一套用来运营公司的系统。检验这句话最老实的办法，是用它来运营我们自己的公司。
所以我们就这么做了。ADIS——我们产品背后的自主决策智能系统（Autonomous Decision Intelligence System）——起草我们的代码、定价、研究和这个网站上的文字，每天读我们自己的经营数据。所有这些，都由人签收。这篇讲它在实际中是什么样子，以及它教会了我们哪些部署 AI 的道理。
产品：ADIS 起草，人来合并
平台里的大部分代码，初稿由 ADIS 写。光凭这一点什么都发布不了：每一处改动都要经过评审和测试才能合并，每一次发布都记进公开的更新日志。
这里的教训来得很快：起草变便宜之后，评审就成了瓶颈。所以我们把力气花在瓶颈上：每一处改动都跑全量测试；每一个接口修复都带一条回归测试和一条修复记录；人看之前，先有独立的复核找问题。目标不是写更多代码，而是让每一次合并，人都能放心地批。
定价：ADIS 起草，负责人签字
我们的定价标准——三套系统，建设费加每人每月订阅——由 ADIS 起草，由公司负责人签字，并且公开。
你在网站上要估价时，ADIS 读你们公司的资料，价格来自公开的定价模型，而不是谈出来的。估价、定价页、Adis 在对话里给的数，都是同一套数，因为它们来自同一个源头。
研究：ADIS 做实验，研究员审
AIDC Research 发表关于智能体系统、智能体组织和 AI 经济学的研究。ADIS 跑实验、写初稿；研究员审方法、审数字、审结论。每篇论文都附上数据，和算出每一个数字的代码，谁都可以检查，包括以后的我们自己。
运营：全公司一份数据
我们的云账单、云资源和开发工作，都放在 Semantic 里，也就是我们给客户部署的那一层数据。每一个需求都是一个对象，链着实现它的代码合并记录，和验收它的人。
正是这一份数据，让 ADIS 在运营上派得上用场。它每次复盘，和我们从同一组数字出发：哪个服务比上个月贵了，哪个资源几周没人碰了，哪项工作还在等验收。它提建议，关不关由人决定。
这个网站
这个网站由 ADIS 组织结构、撰写文字，四种语言，签收后上线。每一页右下角的问 Adis，几秒内回答关于 AIDC 的问题，按问题发合适的材料；访客想找人聊，它就把对话交给我们。
我们能给客户的最好论据，不是一场演示，而是：我们自己的公司，不会用别的方式来运营。
用 ADIS 运营自己，教会了我们什么
负责人必须写明。 上面每一块都有一位具名的人签收。负责人模糊的时候，质量会漂；写明了，就稳住了。
一本账会改变讨论。 账单、资源和工作放进同一个地方的那天起，成本讨论就变短了。
纠正会越积越多。 每一次有人纠正一份初稿，这次纠正就成了 ADIS 下一次干活方式的一部分。几个月后，我们做的纠正，已经从格式问题变成了判断问题。
评审就是工作本身。 起草几乎不花钱的时候，稀缺的能力是知道「好」是什么样，并且把它说清楚。这是我们招人看重的能力，也是我们帮客户在自己团队里建立的能力。
不是我们卖的每一样东西，在一家我们这样规模的公司里都有对应。但我们部署的核心——一份数据、有负责人的数字员工、带审批的动作、按节奏签收——每天都先在这里运行。
AIDC · AI Deployment CompanyAIDC · AI Deployment Company
