# AI Is Not a New Tool. It Is a New Division of Labor. Canonical URL: https://theunclej.com/blog/human-in-the-loop-01-ai-is-new-division-of-labor Markdown URL: https://theunclej.com/blog/human-in-the-loop-01-ai-is-new-division-of-labor.md Description: Human in the Loop, Part 1. Once the agents are running, the real question is not which model you bought. It is who decides, who signs, and who is on the hook. Category: AI Organization Tags: AI, Organization Design, Human in the Loop, Enterprise AI Published: 2026-05-22 --- # AI Is Not a New Tool. It Is a New Division of Labor. ![Human in the Loop | AI Is Not a New Tool, It Is a New Division of Labor](/imgs/posts/human-in-the-loop-01-ai-is-new-division-of-labor-en/01-cover.webp) ![V01 Human in the Loop Is Not Human Reviewing AI](/imgs/posts/human-in-the-loop-01-ai-is-new-division-of-labor-en/02-figure.webp) > This is Part 1 of the _Human in the Loop_ series. ## That night at the terminal That night at the terminal. My back felt off. Not because the system was lagging. The opposite. The system was running too smoothly. I had broken the same org problem into pieces and pushed each piece to a different agent. One pulled sources. One built the structure. One checked facts. One polished the voice. One ran risk review. On the tool layer, it looked beautiful. Sources, in. Structure, in. Review report, in. Every loop could honestly say it had "made progress." Then I stared at the screen and gave a small, dry laugh. I came up through HR. I am not a CTO. What little code I can read, I often need AI to hold my hand through. But that background is exactly why I can see this clearly. Once AI smooths out the motions, what surfaces next is not a technical problem. It is an organizational problem. Who decides? Who decides which paragraph makes it into the final draft? Who decides whether a fact has actually been traced back to source? Who notices that "writing it safer" is really just "writing it more boring"? Who has the authority to look at a compliant-looking draft and call it NO-GO? The model can keep producing. The organization cannot keep pretending the signing seat does not exist. --- Look outside, and the signals get harder. OpenAI announced its Deployment Company in May 2026. The official post is blunt about it: more than $4B in initial investment, with Forward Deployed Engineers working alongside business leaders, operators, and frontline teams to identify where AI lands, and to rewrite organizational infrastructure and core workflows. Around the same window, Anthropic, Blackstone, Hellman & Friedman, and Goldman Sachs announced an AI-native enterprise services firm. Press reports put it at roughly $1.5B. The FIS announcement is more concrete: Anthropic's Applied AI team and FDEs embed inside FIS, co-design a Financial Crimes AI Agent, and transfer knowledge so FIS can keep building and scaling more agents on its own. Add NVIDIA's reported 30,000 engineers using AI coding tools to the same picture, and this stops looking like "another round of tool procurement." I pulled these public signals back into my own terminal and saw one thing. $4B. $1.5B. 30,000 engineers. Money, people, engineers, workflows, knowledge transfer — all crowding onto the same table. That table is not a tool table. It is a division-of-labor table. The boss thinks they bought a faster tool. What they actually bought is a colleague who will quietly take a slice of the judgment work. --- A lot of bosses still talk about AI through an old spreadsheet. Which model do we buy. Who gets seats. How many training sessions. How many headcount can we save. It sounds shrewd. It is thin. A person is not a stack of motions. A person catches judgment, exceptions, accountability, and post-mortems inside a workflow. Hand the motions to AI without rewriting those four things, and the organization does not level up. It just gains one more machine with no name on the signature line. McKinsey's 2024 enterprise AI survey reported that 65% of respondents say their organization regularly uses generative AI in at least one business function. Gartner, at the same year's IT Symposium/Xpo, reminded CIOs that AI value cannot be judged on point productivity alone — cost, risk, value, and business outcomes have to be priced in together. In the projects I have seen up close, the question is rarely "are you using AI." It is: you put AI in, and the old process is still there, the old incentives are still there, the old chain of responsibility is still there. In plain language. AI is no longer an "are we on it" question. A lot of companies are already on it. The real question is whether the organization can catch what comes back. A cash-rich company hits a decision vacuum first. Budget approved, tokens granted, GPUs bought. Now what — layoffs, redeployment, more output, or routing the freed capacity into new lines of business? A cash-poor company hits a different vacuum, a cognitive one. Employees learn AI for weeks, and the boss still cannot say in one sentence how it shows up on the P&L. So this series does not start from "how to use AI." It starts from a harder question. **Once AI is in, what does the organization do?** If you cannot answer that, AI is at best a more expensive plugin. If you can, it starts to become a new division of labor. ## AI strategy meetings keep dying on the first spreadsheet A lot of enterprise AI strategy meetings look right at first glance. The CEO is there. The CIO is there. Finance is there. HR is there. The business owners are there. Senior table. What ends up on paper, most of the time, is three things. Which model to buy. How many seats. How to schedule training. That is awkward. A room full of senior people, sitting in a strategy meeting, running a procurement meeting. It is not that they are not trying. The problem was filed in the wrong drawer from the start. Tool procurement has a mature playbook. Selection, pilot, rollout, usage metrics. A lot of software fits that mold. ERP, OA, CRM — most of them, at the bottom, are about making motions faster inside an existing organization, getting information to flow more smoothly, getting people to fill in fewer forms. That matters. But it never really touches one thing. Who decides. That is where AI gets messy. AI does not just help people move faster. It participates in judgment. A customer service agent judges intent. A recruiting agent judges fit. A compliance agent judges whether a document needs escalated review. The moment AI starts to participate in judgment, a procurement meeting is not enough room. A procurement meeting can pick a vendor. It cannot settle a chain of responsibility. A procurement meeting can negotiate price. It cannot negotiate, once a role gets split into pieces, which piece goes to a person, which piece goes to a machine, and which piece must stay with a human reviewer. Gartner, at IT Symposium/Xpo 2024, told CIOs that AI value cannot be priced on productivity uplift alone. Cost, risk, and business value have to ride in the same model. The version I see in projects is more blunt. An AI pilot cannot only prove the model runs. It has to prove the organization can catch it. Otherwise, once it goes live, the system spits the organizational problems straight back. The boss asks about ROI. Finance asks about cost. HR asks about roles. The business asks who is holding the bag. Each question is reasonable. None of them can actually be closed inside a procurement meeting. So the first mismatch sits right here. The company thinks it is running an AI strategy meeting. It is still filling out a procurement form from the old world. ## A role is not a person. It is a bundle of accountability. The difference between a tool and a division of labor is not in the spec sheet. It is in the responsibility relationship. A hammer is a tool. It lets a person hit harder, but it does not change who the carpenter is. ERP is a tool. It lets data flow more smoothly, but it does not automatically decide who is on the hook for inventory turnover. Most software systems raise the efficiency of motions inside an old organization. Roles still exist. Reporting lines still exist. Responsibility boundaries still exist. AI is different. The moment AI enters a workflow, it hits judgment. Judging whether a customer is high-risk. Judging whether a candidate is a fit. Judging whether a contract has anomalies. Judging whether a piece of content can be published. These look like execution. They carry adjudication inside them. The moment adjudication shows up, responsibility shows up with it. So AI is not a simple speedup of existing roles. It is splitting roles open. A customer service role used to look like one person: greet, identify, answer, escalate, hold the bag. Once AI enters, that one person gets split into pieces. The machine identifies intent. The machine drafts the reply. A human reviews the edge cases. A human handles exceptions. The organization carries the final accountability. The role does not necessarily disappear. But the division of labor inside the role has already changed. A lot of companies misread their first wave of AI exactly here. They manage AI the way they would manage "rolling out a tool." They measure it by usage. They explain it through "hours saved." Then they find out the hardest part is not getting people to open the tool. The hardest part is restating one sentence. From today on, who and who together actually do this thing? If that sentence is not clear, the organization will instinctively domesticate AI into an assistant plugin. Employees use it to draft, look things up, summarize. It looks lively. Workflows, accountability, and incentives stay where they were. There is some value in that. It will not change the organization's capability. To let AI change the organization, you have to admit it is not entering at the tool layer. It is entering at the labor layer. Move the labor layer, and five things follow. Roles need rewriting. Workflows need re-sequencing. Knowledge needs structure. Responsibility needs re-labeling. Governance needs to be auditable and reversible. This is not grand narrative. This is the messy ledger that a boss is going to meet, sooner or later, in an operating review after AI goes live. ## Org OS is not a concept. It is five places where things leak. If AI were just one more tool, a company would only need to hand out seats, run training, and track usage. The moment AI enters the judgment chain, the organization has to answer something more specific. Where does this piece of judgment live? Ask that question, and five leak points show up. First, roles. A customer service role used to look like a complete person. Greet the customer. Understand the problem. Pull the materials. Offer a solution. Decide whether to escalate. Stand behind the result. Once AI enters, that role gets opened up. Intent is identified by a machine. Materials come out of a knowledge base. Replies are drafted by a model. Exceptions go to a human reviewer. The commitment and the bag-holding still sit with some role inside the organization. The machine does the work. The role description does not move. Headcount is asked to be more productive. The responsibility boundary does not move. That is where it starts to leak. Second, workflows. A workflow used to be people walking it step by step. Now the workflow grows new nodes — machine generation, machine retrieval, machine judgment, human review, exception escalation, result acceptance. Leave the workflow diagram untouched, and AI ends up running in a grey zone. When something goes wrong, everyone discovers, only then, that the workflow never wrote down which step can pass automatically, and which step must stop for a human to look. Third, knowledge. A lot of companies say they have a knowledge base. What they actually have is a pile of documents. A document library answers "where is the file." Organizational memory answers "who, in what situation, can pull what knowledge, what judgment that produces, and how it gets updated when it turns out wrong." If knowledge cannot be retrieved, cited, updated, and traced back to accountability, it is not organizational memory for the AI era. It is just a bigger folder. Fourth, responsibility. The suggestion AI gave — who adopted it? Who reviewed it? Who let it pass to the next step? Who is accountable to the customer, the employee, the regulator, the P&L? Leave these unwritten, and the more successful the AI project, the more blurred the responsibility. Fifth, governance. Permissions, audit trails, logging, rollback — they look like back-office IT plumbing. They are actually the floor that decides whether an organization can keep using AI over the long run. Without governance, AI keeps leaning on individual conscience. Any system load-bearing on individual conscience turns into risk at scale. So Org OS is not a pretty concept. It has to catch at least five layers. Roles. Workflows. Knowledge. Responsibility. Governance. Leave any one of them empty for long, and AI will leak out through that layer. ## I am not over here inventing the concept on my own I cannot let this judgment rest on my voice alone. That turns into writing-room concept. I do not like that kind of thing. The external signals are already loud enough. Anthropic openly publishes research on how its own internal work is being changed by AI. OpenAI launched its Deployment Company and stated outright that FDEs work alongside business leaders, operations teams, and frontline teams to rewrite core workflows. FIS and Anthropic baked knowledge transfer into the agent deployment, so the customer can keep extending more agents independently afterward. Google Cloud is writing post-mortems on what changes between prototype and production — governance, evaluation, safety, model selection — once you cross into enterprise territory. YC's AI Native Company curriculum offers a different angle: a startup can design how the company runs around AI capability from day one. I quote them not to find a patron. I quote them because, inside my own systems and projects, I am seeing the same direction. What these materials share is not "everyone is buying tools." It is that the leading institutions have already pushed the question to a deeper place. How does a company get re-organized by AI. Two facts deserve to be kept apart. FIS is one line — knowledge transfer and agent deployment inside a regulated context. Anthropic with Blackstone, Hellman & Friedman, and Goldman Sachs is a separate enterprise-services firm line. The dollar figure on that one can only be cited at the press-report level, and cannot be written up as the FIS contract value. These signals all matter. They are still fragments. Anthropic leans research. OpenAI leans delivery. Google leans technology and cloud practice. YC leans startup. None of them automatically becomes an organizational blueprint that an ordinary Chinese enterprise can take back and copy. The second face is my own working ground. This batch of articles is not one person opening a doc and grinding it out. The shape is: break the proposition, gather sources, trace facts back to origin, draft in sections, then run review gates that block hallucination, RED-04 risks, redaction issues, and bad voice. That process is itself a small Org OS. It is not AI writing for me. It is me putting AI inside a system that has division of labor, accountability, review, and rollback. The third face is the cautionary side of a lot of enterprise AI projects. The model runs. The demo plays. Training is done. Roles did not change. Workflows did not change. Knowledge did not get structured. Responsibility was never written down. Governance never made it into daily operations. In the end, people discover: AI did not fail. The organization failed to give it a runnable seat. Put the three sides together, and the conclusion is plain. AI transformation is not an "is there a model" question. It is whether the organization can carry a new division of labor. ## Stop asking what else AI can do, for a minute So, stop asking what else AI can do, for a minute. The question matters. It also pulls people straight back into the tool view — model capability, plugin capability, automation capability, cost capability. Follow it long enough, and you are back to vendor shortlists, training calendars, and usage dashboards. The question worth pressing is the other one. Now that AI can do these things, how is the organization going to catch them? Are roles getting split. Are workflows getting redrawn. Is knowledge getting re-structured. Is responsibility getting rewritten. Is governance going to step out of back-office rules and into daily operations. Leave these unanswered long enough, and the stronger AI gets, the more grey zones grow inside the organization. What actually pins a lot of companies down is not the absence of a model, and not unwilling employees. It is that the old organizational operating system cannot run the new division of labor. The old system is still asking people to work to the old roles, old processes, old reporting lines, old responsibility lines. The new AI capability has already started slicing tasks, judgment, and knowledge apart on a different grid. So this piece does not close on a slogan. I leave one board-level judgment question. Which box does your company look like, right now? Box one. The OpenAI / Anthropic / NVIDIA path. Put AI into the division of labor, the workflows, and the governance on purpose. Accept that it is not a tool. It is a new way of organizing work. Box two. The Klarna path. Believe replacement first. Discover that quality, accountability, experience, and the cost of walking it back all come home together. Re-hire people. Re-add process. Box three. The most common one. The AI strategy meeting becomes an IT procurement meeting. Seats bought. Training scheduled. Vendors picked. The board never re-signed the agreement on judgment, responsibility, and how the upside gets shared. I am not picking your answer. That is what the next piece is about. Not why AI is not strong enough. Why the Org OS cannot run. --- ## Read on - Series hub: [Human in the Loop](/blog/human-in-the-loop) - Next: [Why Org OS Cannot Run AI](/blog/human-in-the-loop-02-why-org-os-cannot-run-ai)