文章Articles
传统行业 AI 落地的实战洞察与避坑经验。Field notes on making AI work in traditional industries.
让 AI 干活之前,先把流程写下来Before AI Takes Over, Write Down the Process
一家财务 AI 公司推出了 61 个「AI 员工」岗位,背后是 515 份人审过的 SOP,每次按同样的步骤执行,留痕可查。在不能出错的地方,AI 靠得住,靠的是写下来、审过、定了版本的流程。Kognitos has introduced 61 AI finance roles backed by 515 human-approved SOPs. The idea worth examining is how approved processes, controlled execution and audit trails can make AI more dependable.
一个活儿能不能交给 AI?先想清楚三个问题Can This Task Be Handed to AI? Three Questions to Think Through First
常有人问:哪些活儿可以交给 AI?我认为,判断一件事是否可以交给 AI 来做,要先想清楚三个问题:如果错了,能不能马上发现?错了的话,代价有多大?需不需要精确计算?问题想清楚了,通常会发现,一个活儿需要拆开来交:有的交给 AI,有的交给规则,有的留给人。People often ask: which tasks can be handed to AI? In my view, to judge whether something can be done by AI, you first need to think through three questions: if it goes wrong, can you spot it straight away? If it goes wrong, how costly is that? Does it need exact calculation? Once you have thought these through, you will usually find that a task needs to be split up: some parts go to AI, some to rules, and some stay with people.
供应商说「资料安全看你用哪家模型」,只说对了一部分A Vendor Says Data Security Depends on Which Model You Use. That Is Only Partly Right
对接 AI 产品时,资料安全不只取决于用哪家模型。资料从公司系统到模型,中间还要经过供应商自己的产品:它读哪些资料、存不存、能不能改,这些要供应商自己回答。自己买服务器跑模型,也解决不了这一段。When a company connects an AI product, data security does not depend only on which model it uses. On its way from the company's systems to the model, the data also passes through the vendor's own product: what it reads, whether it keeps a copy, whether it can make changes. Those questions are for the vendor to answer, and running a model on your own servers does not solve this part either.
AI落地缺的是什么?Anthropic投1亿美元培训FDE,只答了一半What Is AI Adoption Missing? Anthropic's $100 Million FDE Programme Answers Only Half
Anthropic 宣布投 1 亿美元培训 1 万名 FDE,理由是能让 AI 在真实业务里跑起来的人最难找。但FDE可能只是一半答案,另一半答案在企业自己这边:厂商的 FDE 管把产品用起来,企业得有自己的人,管该不该做、做到什么算好、厂商走了谁接着管。Anthropic has announced a $100 million programme to train 10,000 FDEs, saying the people who can make AI work in real business are the hardest to find. But FDEs may be only half the answer. The other half lies inside the company: a vendor's FDE gets its product used, but the company needs its own person to decide whether to do it, what counts as good enough, and who carries on after the vendor leaves.
从Muse到dots:AI有了自己的电脑,企业该管好什么?From Muse to dots: What Businesses Need to Manage When AI Has Its Own Computer
从云端小龙虾到Muse、dots,Agent的产品化正在把持续执行、安全隔离、账号授权和动作审批整合起来。企业有机会减少部署维护和重复操作,但也需要重新判断:哪些资料可以进入厂商环境,哪些动作可以交给Agent,关键决定由谁负责。From cloud-hosted OpenClaw to Muse and dots, agent products are bringing ongoing execution, isolation, account permissions and action approvals together. Businesses may spend less time on setup and repetitive work, but still need to decide which data can enter a provider's environment, which actions to delegate, and who owns key decisions.
公司的资料,到底能不能发给 AI?付费版也可能拿数据去训练Can company material go into AI? Even paid plans may train on your data
员工多半已经在用 AI 了,公司要做的是把这件事管起来:先分资料,定下哪些不能放进任何外部的 AI;再定工具,优先选条款写明不拿数据训练的;最后管账号,工作资料只放在公司的正式账号里。Your staff are probably using AI already, so the job is to govern how it's used: first sort the material, deciding what must never go into any external AI; then choose the tools, preferring ones whose terms say your data won't be used for training; then control the accounts, keeping work material in company accounts only.
DeepSeek 这次开源的 Harness 不是模型,是骨架What DeepSeek open-sourced this time isn't a model — it's a harness
DeepSeek 上个月开源了 DeepSeek Harness。对企业最有用的信息在名字里:开源的不是模型,是骨架——包在模型外面那层运行系统。它做对了三件事,也划了两条边界。Last month DeepSeek open-sourced DeepSeek Harness. The most useful thing about it for a company is already in the name: what shipped is not a model, it's the harness around one. Three of its design calls are worth borrowing — and so is the boundary it draws around itself.
传统企业缺的不是模型,是一套新的操作系统What traditional companies lack isn't a model — it's a new operating system
工具买了、会员开了、供应商也找了,可业务该怎么转还怎么转。我看过的传统企业,缺的从来不是模型——缺的是数据底座、组织架构、方法论这三样。The tools are bought, the subscriptions are live, the vendors have been met — and the business runs exactly as it did before. What these companies lack was never the model. It's three things: a data foundation, an org structure, and a method.
传统企业的 AI 项目,为什么大多数都白做了?Why most AI projects at traditional companies end up wasted
传统企业的 AI 项目失败,绝大多数不是败在技术,是败在"落地方法"——一个 8 年 CTO 总结的三个最常见的坑。Most AI projects at traditional companies don't fail on technology — they fail on execution. Three traps I keep seeing, from 8 years as a CTO.