Turn labor into compute: what we mean, and what we don't把人力转化为算力：我们指的是什么，不是什么
In most companies, capability lives in people and walks out with them. Here is the alternative, and why it is not about replacing anyone.在大多数公司里，能力存在人身上，人走了能力也跟着走。这篇讲另一种做法，以及它为什么不是要取代谁。
October 7, 2026 · 3 min read2026 年 10 月 7 日 · 阅读约 2 分钟
In most companies, capability lives in people. The buyer who knows which supplier slips every December. The accountant who knows why one customer's invoices always need a second look. The planner who knows which machine to avoid on a Friday night shift.
That knowledge is real, and it is fragile. It takes months to hire someone and longer to train them. When they leave, their experience leaves with them. And the people who know the most spend their days answering the same questions for everyone else.
"Turn labor into compute" is how we describe the alternative. It is easy to misread, so here is what we mean by it, and what we don't.
Labor scales by the person. Compute scales with the work.
A person's day has a fixed number of hours. If twenty colleagues need the same lead-time check at nine in the morning, the buyer who knows how to do it answers twenty times, or nineteen people wait.
A digital employee that has learned how that check is done answers all twenty at once, cites the purchase orders it read, and is ready again at 9:05. Its capacity grows with the work instead of with headcount. That is the economic shift: the cost of a routine answer stops being someone's afternoon.
What we mean
We mean that the way work is done in a company should become something the company keeps.
In practice, every department gets one digital employee and every digital employee has a named owner. The owner trains it the way you would train a new colleague: with the policies, the examples, the exceptions, and with corrections. Each correction becomes part of how it works the next day, for everyone in the department.
Over time, the digital employee takes on the lookups, the first drafts, the follow-ups and the weekly summaries. The person keeps the judgement: approving the quote, deciding on the supplier, signing the month-end report.
What we don't mean
We don't mean replacing people. The digital employee works in the team's chat, next to the team, and it is only as good as the people who train it.
We don't mean "fully automatic". Reading and drafting run freely, but anything that changes a rule, a data model or a record in a business system waits for a person to approve it, and every step is logged.
And we don't mean "set and forget". A digital employee that nobody corrects stops getting better. Ownership is part of the design, not a formality.
Where the technology actually is
A useful way to think about AI capability is the five levels OpenAI has described: chatbots that answer questions, reasoners that analyse problems, agents that get work done, innovators that discover new knowledge, and organisations of AI that run operations together.
Enterprise deployment today is at level three. Agents that take a task in a group chat, read the right systems, draft the result and route it for approval are running every day. Level four is starting to appear at the edges: at customers, digital employees that read the same operating data every day have begun to surface patterns nobody had asked them to look for. Level five is where this is heading, not where it is.
The question for a leadership team is not whether AI will replace jobs. It is where the company's know-how lives, and whether it compounds.
How it compounds
Know-how in people grows slowly and leaves suddenly. Know-how in a system grows with every correction and stays when people move on.
That is why a deployment grows in stages. In the first three months, every department has a digital employee and everyone uses it, even if only to ask questions. By month six, the work is modelled role by role and use cases are connected along the value chain; for a growth-stage group, the target is a network of more than a hundred agents. By month twelve, the company works from one set of numbers and makes decisions in the system, and the target is more than a thousand agents across the value chain. The actual scope is set in the deployment assessment.
None of this requires the company to change its systems. The ERP stays. The chat stays. What changes is where the experience lives.
The practical test
If one of your best people left on Friday, how much of what they knew would still be working on Monday? Today, for most companies, the honest answer is very little.
Turning labor into compute is the work of changing that answer.
在大多数公司里，能力存在人身上。那位知道哪家供应商每年十二月都会延期的采购；那位知道某个客户的发票为什么总要多核一遍的会计；那位知道周五夜班哪台机器最好别排的计划员。
这些经验是真的，也很脆弱。招一个人要几个月，带出来要更久。人一走，经验跟着走。而最懂行的人，每天都在替别人回答同样的问题。
「把人力转化为算力」是我们对另一种做法的说法。这句话容易被读偏，所以这里说清楚：我们指的是什么，不是什么。
人力按人头扩，算力随工作量扩
一个人一天的小时数是固定的。早上九点有二十位同事要核同一批交期，会核的那位采购就得答二十遍，否则十九个人在等。
学会了这件事怎么做的数字员工，可以同时答这二十个人，附上它读过的采购订单，九点零五分又能接下一件。它的容量跟着工作量长，而不是跟着人头长。这就是经济上的变化：一个常规问题的成本，不再是某个人的一个下午。
我们指的是什么
我们指的是：公司里「事情怎么做」，应该变成公司自己留得住的东西。
具体做法是，每个部门一名数字员工，每名数字员工有一位具名的负责人。负责人像带新同事一样带它：给制度、给例子、讲例外，并且纠正它。每一次纠正，第二天就成了它干活方式的一部分，整个部门都受益。
慢慢地，查数、初稿、跟进、周报交给数字员工；判断留给人：批报价、定供应商、签月报。
我们不是指什么
不是取代人。数字员工在团队的群里干活，和团队在一起；带它的人有多好，它就有多好。
不是「全自动」。查询和起草不受限，但凡是要改规则、改数据模型、改业务系统里的记录，都要等人批准，每一步都留痕。
也不是「装上就不管」。没人纠正的数字员工不会变好。负责人是设计的一部分，不是走过场。
技术真正走到了哪一步
看 AI 的能力，有一个好用的框架：OpenAI 提出的五个等级——回答问题的聊天机器人、分析问题的推理者、把事做完的智能体、发现新知识的创新者，以及一起运营的 AI 组织。
企业部署今天处在第三级。在群里接活、读对的系统、起草结果、提交审批的智能体，每天都在跑。第四级开始在边缘出现：在客户那里，每天读同一份经营数据的数字员工，开始找出没人让它去找的规律。第五级是方向，不是现状。
对管理层来说，问题不是 AI 会不会取代岗位，而是公司的经验存在哪里，会不会越积越多。
怎样越积越多
存在人身上的经验，长得慢，走得快。存在系统里的经验，每纠正一次长一点，人走了也还在。
所以部署按阶段长。前 3 个月，每个部门都有数字员工，全员在用——哪怕只是问答。到第 6 个月，按岗位把活建模，场景沿价值链连起来；对一家成长期的集团，目标是一张 100 个以上智能体的网络。到第 12 个月，全公司一本账、在系统里做决定，目标是覆盖价值链的 1000 个以上智能体。具体范围在部署评估里定。
这些都不需要公司换系统。ERP 还在，群聊还在。变的是经验存放的地方。
一个实际的检验
如果你最好的员工周五离职，这个人知道的东西，到周一还有多少在继续起作用？对今天的大多数公司，诚实的答案是：很少。
把人力转化为算力，就是去改变这个答案。
AIDC · AI Deployment CompanyAIDC · AI Deployment Company
