The Company Buys the Work, Not the Person公司买的是工作,不是人公司買的是工作,不是人 · 09
Can the Signer Actually Stop It?
签字的人,真的能叫停吗
簽字的人,真的能叫停嗎
A person in the workflow does not prove that judgment remains human. Information, time, authority, and protection for exercising judgment must become real.
流程里有一个人,不等于判断已经在人手里。信息、时间、权限与不因改判吃亏的条件,要一起到位。
流程裡有一個人,不等於判斷已經在人手裡。資訊、時間、權限與不因改判吃虧的條件,要一起到位。
Imagine an analyst checking automated rating drafts and signing them. The company says saved time will support thought. His caseload doubles, however, and every change counts as disagreement in his assessment. The signature remains while the space for judgment may be shrinking.
The question is what the reviewer can actually do. Can he obtain and understand evidence? Can he change the outcome? Will a justified revision damage his performance rating? Without those conditions, a step named review can become little more than pressing approve.
The source addresses the deploying company first: if a role carries judgment, make it real. It does not command every worker to become a decision maker or call someone backward for preferring skilled manual work. Work arrangements and a person's desired direction are separate matters.
A fatal event makes the distinction weightier. In the 2018 Tempe automated-test-vehicle crash, the NTSB identified operator distraction as probable cause and company deficiencies in risk assessment, supervision, and automation-complacency safeguards as contributing factors. Its investigation concerned safety, not determination of every legal liability.
Company deficiencies do not erase an operator's actions; an occupied seat does not prove adequate company conditions. Both can be true on separate grounds. A fatal crash and an office metric differ enormously in consequences. The comparison is confined to formal powers and practical conditions.
First, a person must be able to judge. Required information must be available and intelligible, with training, support, and time. Supervising a system that is usually right creates difficulties of sustained attention. An instruction to concentrate cannot be the whole arrangement.
Second, judgment must change something. Within the role, a person can revise, pause, or initiate effective escalation. Not every position needs power to shut an entire system down. But the case at hand cannot be compelled to continue automatically despite a recognized problem.
Third, exercising that authority must not itself bring disadvantage. A justified inquiry should not lose points simply for taking time or differing from the model. This concerns metrics as well as personal retaliation. An arrangement can punish care without anybody intending to punish it.
Justification depends on information, authority, and conditions available at the time, not only hindsight. Protection does not exempt deception or abuse of power. Those require independent grounds. A safeguard for judgment is not a promise that operators can never be at fault.
Adequate conditions still do not certify substantive review in every instance. Agreement can follow examination; disagreement can be mechanical. Evidence considered, people heard, and reasons recorded matter. Acceptance and rejection rates cannot become tests of an employee's inward life.
Imagine a kiln workshop allowing workers to alter or halt a firing curve without automatically charging spoiled materials against wages. Yet the displayed records describe the previous clay batch. Authority exists, but information about the new material is missing. Permission to intervene does not complete the conditions.
That example and the rating example expose different gaps: missing information and potentially punitive assessment. Missing conditions cannot collapse into a general charge of employee failure. If judgment is assigned, necessary time and training are requirements of the role, not optional benefits offered as kindness.
In Moffatt v. Air Canada in 2024, a website chatbot misdescribed bereavement-fare applications. British Columbia's Civil Resolution Tribunal found the airline had not taken reasonable care over accuracy and rejected treating the chatbot as separately responsible. This was a particular decision, not a universal allocation of liability for all systems.
It illustrates the source's demand that a company not remove its deployed information channel from the account by attributing words to AI. Configuration, deployment, use, and supervision remain traceable choices. An algorithm is not a replacement actor arriving to carry every consequence.
Nor does corporate responsibility automatically transfer to every individual. A final signer lacking control or practical review conditions cannot carry the whole decision merely because of a name. Separate personal responsibility requires independent grounds, while the corporate name cannot conceal independently attributable personal actions.
An inquiry should therefore continue past the person closest to the outcome. Who fixed caseloads, designed agreement metrics, approved release, and knew information was missing? Each position accounts for its actual contribution. Calling something a process failure must not mean nobody is responsible.
Volume, speed, and agreement metrics are not offenses by name. They can supply useful signals. Their incorporation matters: do they remove review time, penalize justified deviation, or make checking and suspension practically impossible? These are questions about operation rather than terminology.
High agreement might indicate a good model or nobody able to revise it. Low agreement might indicate active review or poor output. Figures need supporting facts. Certifying human judgment from one percentage merely delegates judgment to another automatic signal.
Automation can genuinely change roles. Someone who watched a kiln overnight may learn a new task or not want that direction. A company can alter work within grounded, challengeable authority without making a person's development path an additional command.
Wanting to retain a skill belongs to the person; retaining every particular position is a shared work arrangement. Neither answers the other. This essay guarantees no permanent execution jobs and gives technical progress no automatic exemption from explaining costs, reasons, and departure.
If automation already saves the company money, transformation cannot automatically become a crisis requiring sacrifice from everyone. A real shortage needs evidence. Change must still preserve basic conditions and choice; savings for the company cannot by themselves justify transferring every burden to its workers.
Training can provide conditions; indispensable role support cannot be downgraded to a favor. Additional help also cannot require adopting the firm's designated future as a personal wish. Acceptance, refusal, and departure leave concrete shared matters to handle. A growth slogan cannot answer for the person.
Monitoring follows work needs rather than technical capacity. Imagine a camera intended to inspect kiln color also recording a rest corner. Changing its angle is a small correction with a significant distinction: reasons to observe the fire are not reasons to observe every minute of workers' lives.
Delegation to a product has a similar boundary. Choosing less effort and allowing a tool to select can itself be the user's judgment. Every detail need not be personally performed. The scope should be visible, and unassigned matters should not be decided incidentally.
The company should explain how authorization is interpreted and executed and answer for information it provides. Stopping future delegation needs a usable route. A simple authorization should not lead to an obscure maze when the person wants to withdraw it.
Completed and ongoing actions require separate handling. Revoking authorization does not erase third parties' rights or automatically undo every transaction. Stopping and correction depend on scope and arrangements, while a route into review remains available. Withdrawal is not a magical deletion of history.
Is this an imaginary perfect company? The source accepts that flawless practice everywhere is a direction, not a promised destination. But access to evidence, effective suspension, and penalties for revision can be checked now. The ideal need not arrive before a particular condition becomes answerable.
A name may still need to go on the form. The important question is whether its owner could see, pause, and change something before signing. A company assigning responsibility for judgment must provide more than a job that leaves a human name beneath a system's decision.
设想一个分析师每天打开系统给出的评级草稿,检查后签字。公司说,自动化节省的时间可以用来思考;同时,他负责的数量翻倍,考核又把每次改动计作不一致。签名仍在,判断的空间却可能越来越小。
问题不是人工介入有几个字,而是这个人实际能做什么。有没有取得并理解资料的时间,能不能改变结果,认真改判会不会损害自己的考核?没有这些条件,把最后一步叫复核,也可能只剩按下通过。
原作把要求先说给部署系统的公司:交给人判断,就要把判断的位置做实。它不要求每个人都转型成判断者,也不把愿意继续做熟练手工的人看成落后。工作的安排与人自己想往哪里去,是两件事。
一件真实事故让区别更沉重。2018年,Uber自动驾驶测试车在坦佩撞死一位行人。NTSB的调查把安全员分心列为可能原因,同时列出公司风险评估、监督及应对自动化松懈机制不足等促成因素。报告是安全调查,不是对所有法律责任作裁决。
这里既不能用公司流程有问题抹掉操作者的动作,也不能只看安全员坐在那里,就说公司已经提供全部判断条件。两面可以同时成立,需要各自根据。严重事故与普通办公指标后果不同,比较只限于角色权限与实际条件的区别。
做实首先是能判断。该知道的资料取得到,内容理解得了,有训练、支援与足够时间。监督一个大多时候正确的系统,也需要应对长期注意的困难,不能把“应该认真”当成全部设计。
其次是能改变结果。在职责范围内可以改、可以暂停,或启动真正有权处理的升级通道。不必每个岗位都能关掉整家公司;但手里这一件,不能明知有问题也只能自动传下去。
第三是行使权力不因此吃亏。一次有根据的核查,不应只因多花时间、与系统不同,就被扣分。这既防个人报复,也看指标结构;没有人存心惩罚,考核仍可能稳定惩罚认真判断的人。
是否有根据,要按当时能取得的信息、权限与实际条件看,不用最后结果倒推。保护判断也不替欺瞒或滥用权限免责;那些需要独立根据,不能把一个保障写成操作者从此永远不会有错。
条件齐备也不能认证每次真的判断。有人同意系统,是检查后同意;有人否决系统,只是机械反对。要看过什么资料、听了谁、留下怎样理由,而不是把接受或反对的比例当成内心与能力测验。
设想一家窑坊允许当班者修改或叫停烧成曲线,也不从工资里直接扣坏料损失。安排看似周全,可系统旁列的仍是上一批原料记录。新料不同,曲线出错,权限在手也不代表判断所需信息已经给全。
这与开头评级的设例相照:一个缺信息,一个可能被考核压住。不能把缺少条件与没有使用条件混成一句“员工失职”。公司若要求岗位承担判断,相应训练与时间是履职所必需,不是可给可不给的额外福利。
再看归责。2024年的 Moffatt诉加拿大航空裁决中,官网聊天机器人给出错误的丧亲票申请信息。卑诗省民事解决仲裁庭认定,加航未尽合理注意确保回答准确,不能把聊天机器人当成独立承担行为的实体。这是该案的具体裁决,不是一条替所有系统完成归责的通用判例。
它照亮原作的一条要求:公司不能只说“那是AI讲的”,就把自己部署的信息通道从账上拿掉。系统输出仍须追到谁设定、谁部署、谁采用、谁监督,不能把算法当成一个会来承担全部后果的替身。
但这也不等于所有个人责任都直接从公司转过来。最后签字的人若没有控制权、没有实际判断条件,不能只凭签名让他承受全部决定。个人另负什么,须有独立根据;公司名义也不能遮蔽实际操作者的独立行为。
因此调查不能停在离结果最近的那个人。谁定了复核数量,谁设计一致率指标,谁批准发布,谁知道资料缺口,都在决定链上。各位置就自己实际承担的部分说明理由、纠正错误,不能把“流程问题”说成无人负责。
考核中的处理量、速度和一致率本身不是罪名。它们可以是信号,也可能帮助发现问题。需要看怎样进入安排:有没有挤掉复核时间,是否让有理由的偏离吃亏,暂停查证是否仍实际可行。
一致率高,也许模型很好,也许没人能改;一致率低,也许判断活跃,也许系统很差。数字需要自己的事实支撑。用某个比例直接验收“人仍在判断”,只是又把判断交给另一个自动信号。
执行交给系统后,岗位确实可以变化。原来守夜的人可能学习新任务,也可能不想走那条路。公司能在有根据的权限与可问安排里调整工作,不能把个人成长方向也一并写成命令。
想继续原来的技能,是人的选择;公司是否保留每一个原岗位,是共同工作安排。两件不能互相代答。本文不许诺一切执行岗位永久存在,也不让“技术进步”自动免掉对调整理由、代价与退出的说明。
若自动化已经给公司节省成本,更不能只借“转型”把变化说成所有人应共同分担的难关。有没有真实困难,需要资料。调整仍要守住基本生存、可离开与可另选方向;公司得到的节省,不能独自成为把所有代价推给员工的理由。
培训可以提供条件,所需的基础履职支持不能降格为恩惠。除此之外的帮助,也不能附带要求员工把公司指定未来当作自己的愿望。接受新工作、拒绝或离开,都有具体共同事项需要交代,不是由成长口号替他回答。
监控也应按工作需要,而非技术能力划范围。设想窑房摄像头原为观察火色,却照到休息角落。调整角度可以很小,却承认一个重要区别:公司想知道火怎样烧,不等于有理由知道员工每一分钟怎样生活。
用户把事情委托给产品,也有相同边界。少操心、让工具替自己选,可以就是用户的判断,不需要强迫每个细节都亲自完成。关键是授权范围看得清,没交出去的事不被顺手代定。
公司应说明系统怎样理解与执行授权,并承担它向用户提供信息的相应责任。停止今后的代办,要有实际可发动的路;不能让一键授权后,撤回变成找不到入口、说不清对象的一串障碍。
已经办成或正在办理的事则要另看。取消授权不使第三人的权利消失,不等于所有已完成交易都自动撤销。如何停、如何纠正,仍照原授权与相关安排处理,同时保留转入审查的通道。
这是不是想象中才有的公司?原作承认,方方面面永远做得完美是一种方向,不许诺现实必然实现。但某一岗位有没有资料、能不能暂停、改判是否被扣分,今天就能核,不必等理想公司先出现。
最后那张表仍需要有人签。值得问的不是纸上有没有名字,而是签名前那个人是否真有机会看见、停顿与改变。若公司要他为判断负责,就也要让他拥有不只是替系统留下一个人名的工作。
設想一個分析師每天打開系統給出的評級草稿,檢查後簽字。公司說,自動化節省的時間可以用來思考;同時,他負責的數量翻倍,考核又把每次改動計作不一致。簽名仍在,判斷的空間卻可能越來越小。
問題不是人工介入有幾個字,而是這個人實際能做什麼。有沒有取得並理解資料的時間,能不能改變結果,認真改判會不會損害自己的考核?沒有這些條件,把最後一步叫複核,也可能只剩按下通過。
原作把要求先說給部署系統的公司:交給人判斷,就要把判斷的位置做實。它不要求每個人都轉型成判斷者,也不把願意繼續做熟練手工的人看成落後。工作的安排與人自己想往哪裡去,是兩件事。
一件真實事故讓區別更沉重。2018年,Uber自動駕駛測試車在坦佩撞死一位行人。NTSB的調查把安全員分心列為可能原因,同時列出公司風險評估、監督及應對自動化鬆懈機制不足等促成因素。報告是安全調查,不是對所有法律責任作裁決。
這裡既不能用公司流程有問題抹掉操作者的動作,也不能只看安全員坐在那裡,就說公司已經提供全部判斷條件。兩面可以同時成立,需要各自根據。嚴重事故與普通辦公指標後果不同,比較只限於角色權限與實際條件的區別。
做實首先是能判斷。該知道的資料取得到,內容理解得了,有訓練、支援與足夠時間。監督一個大多時候正確的系統,也需要應對長期注意的困難,不能把「應該認真」當成全部設計。
其次是能改變結果。在職責範圍內可以改、可以暫停,或啟動真正有權處理的升級通道。不必每個崗位都能關掉整家公司;但手裡這一件,不能明知有問題也只能自動傳下去。
第三是行使權力不因此吃虧。一次有根據的核查,不應只因多花時間、與系統不同,就被扣分。這既防個人報復,也看指標結構;沒有人存心懲罰,考核仍可能穩定懲罰認真判斷的人。
是否有根據,要按當時能取得的資訊、權限與實際條件看,不用最後結果倒推。保護判斷也不替欺瞞或濫用權限免責;那些需要獨立根據,不能把一個保障寫成操作者從此永遠不會有錯。
條件齊備也不能認證每次真的判斷。有人同意系統,是檢查後同意;有人否決系統,只是機械反對。要看過什麼資料、聽了誰、留下怎樣理由,而不是把接受或反對的比例當成內心與能力測驗。
設想一家窯坊允許當班者修改或叫停燒成曲線,也不從工資裡直接扣壞料損失。安排看似周全,可系統旁列的仍是上一批原料記錄。新料不同,曲線出錯,權限在手也不代表判斷所需資訊已經給全。
這與開頭評級的設例相照:一個缺資訊,一個可能被考核壓住。不能把缺少條件與沒有使用條件混成一句「員工失職」。公司若要求崗位承擔判斷,相應訓練與時間是履職所必需,不是可給可不給的額外福利。
再看歸責。2024年的 Moffatt訴加拿大航空裁決中,官網聊天機器人給出錯誤的喪親票申請資訊。卑詩省民事解決仲裁庭認定,加航未盡合理注意確保回答準確,不能把聊天機器人當成獨立承擔行為的實體。這是該案的具體裁決,不是一條替所有系統完成歸責的通用判例。
它照亮原作的一條要求:公司不能只說「那是AI講的」,就把自己部署的資訊通道從帳上拿掉。系統輸出仍須追到誰設定、誰部署、誰採用、誰監督,不能把算法當成一個會來承擔全部後果的替身。
但這也不等於所有個人責任都直接從公司轉過來。最後簽字的人若沒有控制權、沒有實際判斷條件,不能只憑簽名讓他承受全部決定。個人另負什麼,須有獨立根據;公司名義也不能遮蔽實際操作者的獨立行為。
因此調查不能停在離結果最近的那個人。誰定了複核數量,誰設計一致率指標,誰批准發佈,誰知道資料缺口,都在決定鏈上。各位置就自己實際承擔的部分說明理由、糾正錯誤,不能把「流程問題」說成無人負責。
考核中的處理量、速度和一致率本身不是罪名。它們可以是信號,也可能幫助發現問題。需要看怎樣進入安排:有沒有擠掉複核時間,是否讓有理由的偏離吃虧,暫停查證是否仍實際可行。
一致率高,也許模型很好,也許沒人能改;一致率低,也許判斷活躍,也許系統很差。數字需要自己的事實支撐。用某個比例直接驗收「人仍在判斷」,只是又把判斷交給另一個自動信號。
執行交給系統後,崗位確實可以變化。原來守夜的人可能學習新任務,也可能不想走那條路。公司能在有根據的權限與可問安排裡調整工作,不能把個人成長方向也一並寫成命令。
想繼續原來的技能,是人的選擇;公司是否保留每一個原崗位,是共同工作安排。兩件不能互相代答。本文不許諾一切執行崗位永久存在,也不讓「技術進步」自動免掉對調整理由、代價與退出的說明。
若自動化已經給公司節省成本,更不能只借「轉型」把變化說成所有人應共同分擔的難關。有沒有真實困難,需要資料。調整仍要守住基本生存、可離開與可另選方向;公司得到的節省,不能獨自成為把所有代價推給員工的理由。
培訓可以提供條件,所需的基礎履職支持不能降格為恩惠。除此之外的幫助,也不能附帶要求員工把公司指定未來當作自己的願望。接受新工作、拒絕或離開,都有具體共同事項需要交代,不是由成長口號替他回答。
監控也應按工作需要,而非技術能力劃範圍。設想窯房攝像頭原為觀察火色,卻照到休息角落。調整角度可以很小,卻承認一個重要區別:公司想知道火怎樣燒,不等於有理由知道員工每一分鐘怎樣生活。
使用者把事情委託給產品,也有相同邊界。少操心、讓工具替自己選,可以就是使用者的判斷,不需要強迫每個細節都親自完成。關鍵是授權範圍看得清,沒交出去的事不被順手代定。
公司應說明系統怎樣理解與執行授權,並承擔它向使用者提供資訊的相應責任。停止今後的代辦,要有實際可發動的路;不能讓一鍵授權後,撤回變成找不到入口、說不清對象的一串障礙。
已經辦成或正在辦理的事則要另看。取消授權不使第三人的權利消失,不等於所有已完成交易都自動撤銷。如何停、如何糾正,仍照原授權與相關安排處理,同時保留轉入審查的通道。
這是不是想象中才有的公司?原作承認,方方面面永遠做得完美是一種方向,不許諾現實必然實現。但某一崗位有沒有資料、能不能暫停、改判是否被扣分,今天就能核,不必等理想公司先出現。
最後那張表仍需要有人簽。值得問的不是紙上有沒有名字,而是簽名前那個人是否真有機會看見、停頓與改變。若公司要他為判斷負責,就也要讓他擁有不只是替系統留下一個人名的工作。
Source & reading note原论文与阅读说明原論文與閱讀說明
Ten essays draw on the Company prequel, six main papers, the platform-work and AI essays, and the dialogue with corporate theory. They examine a company direction publicly treating people as ends, not a classification of entire real businesses. Basic protections do not disappear when a company changes direction. Unattributed factories, platforms, workshops and contracts are imagined. Linked examples distinguish company testimony, regulatory findings, judgments and settlements within their stated dates and scope, not current legal or investment advice. Organizational accounts, role duties and independently grounded personal responsibility remain separate. Detailed law, the full historical dialogue and unresolved thresholds remain in the source papers. English is independently written.
十篇据《公司论》前置篇、六篇正文、平台与零工及AI专论、传统对话改写。主要讨论公开把人作为目的的公司方向,不给现实企业贴整体标签;对人的基本保护不因公司换方向而消失。未注明史料来源的工厂、平台、窑坊和合同均为设想。真实案例另列出处,区分公司自述、监管认定、判决与和解,按材料所记日期和范围说明,不作现行法律或投资建议。公司账、岗位职责与有独立根据的个人责任分开。详细法条、全部思想史对话和尚无统一答案的阈值留在原论文。简繁中文同文,英文独立重写。
十篇據《公司論》前置篇、六篇正文、平台與零工及AI專論、傳統對話改寫。主要討論公開把人作為目的的公司方向,不給現實企業貼整體標籤;對人的基本保護不因公司換方向而消失。未註明史料來源的工廠、平台、窯坊和合同均為設想。真實案例另列出處,區分公司自述、監管認定、判決與和解,按材料所記日期和範圍說明,不作現行法律或投資建議。公司帳、崗位職責與有獨立根據的個人責任分開。詳細法條、全部思想史對話和尚無統一答案的閾值留在原論文。簡繁中文同文,英文獨立重寫。
Read the source paper ↗阅读原论文 ↗閱讀原論文 ↗
NTSB report on the Tempe automated-test-vehicle crashNTSB坦佩自动驾驶测试车事故调查报告NTSB坦佩自動駕駛測試車事故調查報告 ↗