# The New Labor Contract Canonical URL: https://theunclej.com/blog/human-in-the-loop-04-new-labor-contract Markdown URL: https://theunclej.com/blog/human-in-the-loop-04-new-labor-contract.md Description: Human in the Loop, Part 4. Once Skills, Agents, and anti-distill tools show up, the contract between company and employee stops being about hours. It gets repriced around injection, judgment residue, and accountability. Category: AI Organization Tags: AI, Organization Design, Human in the Loop, Enterprise AI Published: 2026-05-22 --- # The New Labor Contract ![Human in the Loop | The New Labor Contract](/imgs/posts/human-in-the-loop-04-new-labor-contract-en/01-cover.webp) ![V04 Layering: Rules, Models, Humans, and Accountability](/imgs/posts/human-in-the-loop-04-new-labor-contract-en/02-figure.webp) > This is Part 4 of the _Human in the Loop_ series. ## The colleague.skill cold open The first time I saw `anti-distill`, my first reaction was not "what a clever tool." It was: the hand attached to the old contract has finally reached too far. What the person in the corner office needs to look at this year is not whether employees are using AI. It is that someone on GitHub is already writing code to stop the company from distilling their judgment away. Let me start with a project whose name sounds light and whose stakes are heavy. It first showed up as `colleague-skill`, then got upgraded to `dot-skill`. The public page is direct about what it does: your colleague left, your mentor graduated, your teammate transferred — they took their working methods, their context, their judgment frames with them. The project turns the material they left behind into a Skill the AI can call. Two numbers sit on that page. The project crossed ten thousand stars. The README marks 15k stars on April 19, 2026. By the time I pulled it down, the public page showed 18.3k. This is not a tool news story. If a company says "we want to distill the methods of our best people," it sounds normal. The old names for it were SOP, knowledge base, best practice. The moment it becomes a Skill, the nature of the thing changes. SOPs are for humans to read. Skills are for agents to run. Knowledge bases are for new hires to study. Skills are for the system to reuse. What companies used to sediment was experience. What they want to sediment now is judgment. Of course employees can feel that. So another project showed up fast. It is called `anti-distill`. The title is even more direct: an anti-distillation Skill. The public page shows 2.2k stars. The logic in the README is clean: if the company asks you to turn your work into a Skill, you hand the file to this tool, and it outputs a clean, structured, professionally-worded version that looks like it does the job — while the real value, the scar tissue, the gut calls, the human knowledge, stay in your private backup. Do not read this as a joke. The joke version is "employees are scared of being replaced by AI." Code is not a joke. The code is saying something else: employees are scared of their judgment being pulled out and turned into an organizational asset that no longer belongs to them. The sense for a client's pace, the rhythm of pricing, the gut feel for a system, the touch in collaboration — things a person accumulated over years and kept inside their head — the employer used to buy their working time. The moment those things get written into a Skill, the employer is no longer buying time. They are buying a replicable judgment template. And the contract is not enough to hold that. The employee is not being precious. The company is not being evil. What is being traded changed. The contract is still in the old era. The conflict here is not in the AI. Employees are not anti-AI. The person who wrote `anti-distill` is using the Skill system themselves. They are not against the model, the automation, the productivity tool. They are against one specific thing: a hand with no contract boundary, reaching into the working process, pulling out judgment, turning it into a reusable asset, and still paying by the hour. That is the first crack in the new labor contract. The old contract was written around role, salary, hours. What is actually getting traded now is a different bundle: how much know-how the employee injects into the AI, and how much judgment the human still has to carry after the AI runs. The old contract has not caught up. Employees are already writing code to protect themselves. The anti-distillation tool is not technology news. It is labor-relations news. What it is really telling the person in the corner office is this: if you treat AI as a productivity tool, what you see is "employees won't share their experience." If you treat AI as a new division of labor, what you see is "the company is buying new assets with an old contract." Between those two readings sits a labor contract no one has written yet. ## The old human-time-wage schema is breaking Since industrial capitalism, the employment relationship has run on a quiet schema. Person. Time. Wage. A company hires a person. The contract names a role. The system tracks attendance. Finance issues pay. It looks like you are buying the person. The settlement actually lands on "how much of this person's time did we occupy." Monthly salary, hourly rate, overtime, bonus — different shapes, same assumption underneath: the time a person is present is a decent proxy for their attention, experience, and judgment. That assumption used to be enough. Because judgment was hard to pry out of time. How does a salesperson know whether to give ground on the third round of price pressure. How does an HRBP know that "the team morale is off" actually means the goals were broken down wrong. How does an engineer know that one architectural call will turn into three months of tech debt. None of that is easy to meter separately. The company had no choice but to bundle it inside role and hours. So the old contract was really buying a mixed bag. Attendance. Attention. Experience. Judgment. Accountability. AI walked in, and that bag got opened up for the first time. Part of the experience can be written as a prompt. Part of the workflow can be wrapped as an agent. Part of the judgment can be pre-run by a model. Material that used to take a senior half a day to assemble is sitting in a first draft inside minutes. Tradecraft that used to take repeated mentorship can be written into a Skill. Retros that used to depend on personal memory can now be sedimented into something queryable, reusable, iterable. This is not a simple efficiency lift. YC-adjacent discussion has pushed AI past "efficiency tool" toward "new capability, new company shape." Translated into contract language, the shift is this: buying capability by the hour is not stable anymore. Capability no longer sits only in the person. It gets continuously injected into the system. Judgment is no longer entirely produced inside a human head. AI runs the first half of it. The harder half gets pushed back to a human. So the "time" cell starts to leak. A lot of bosses think they are buying eight hours of an employee's day. Inside those eight hours, there are at least three buckets. One is attention. One is the act of injecting know-how into AI. One is the judgment residue the human still has to hold after the AI is done. The old contract bundles all three into one line item. The new division of labor pulls them apart at settlement time. The problem is, a lot of companies are still reading new work through the old ledger. An employee writes thirty high-quality rules that quietly let every agent downstream make one fewer mistake per day. In the old ledger that is "produced a document today." Another employee runs the tool in circles, generates a pile of lively-looking usage logs. In the old ledger that becomes "actively engaged with AI." The employee did not get worse. The pricing lens got distorted. The thing the person in the corner office should actually be looking at is not "how many people did AI replace." The harder question is: of your labor cost, how much is still buying attention, how much is buying reusable know-how, how much is buying the judgment backstop after the AI runs. Pile those three together and the labor cost line gets bloated. You will keep paying old prices for some roles. You will badly under-price what other people are actually creating. The problem with the old labor contract is not that it instantly stops working. The problem is that it stops being able to name what is being traded. What the employee delivers is not just time. What the employer receives is not just current output. There is a new layer in the middle — AI — and the old schema is still paying based on "how long did the person sit in that seat." That is why a new contract has to show up. ## Employee panic: from rumor to code On the employee side, the reaction shows up in two shapes. One shape is rumor. I have seen a story circulate. Employees, when their personal experience was being distilled inside the company, quietly embedded a Navier-Stokes trigger — anything that tried to extract their core know-how would get redirected into trying to solve a massive computational problem, burning the compute. This is an unverified anecdote. Nothing more. I do not have a reliable original post. I cannot prove it actually happened. I am not going to attach it to any specific company. Whether it is technically feasible is not the point of this piece. The value is not in factual load. The value is in emotional load. Why does this story feel true to people? Because it names, precisely, the first reaction an employee has when facing distillation. That reaction is not "I hate AI." It is closer to: "You cannot, at the same time, tell me this is part of my job, and pull my judgment out of me to build a system that won't need me later." The other shape is code. `anti-distill` is the harder expression. It does not run on rhetorical flourish. It turns the defensive move into a tool. Read the Skill file. Identify which parts carry replacement value. Replace the core knowledge with statements that read as correct but carry almost no leverage. Keep the actual judgment for yourself. Not a joke. Tradesmen used to protect their craft through apprenticeship, guilds, and not writing the critical step into the public workflow. Knowledge workers protected their judgment by "not spelling it out," "only going this deep with you," "covering the seams myself." Companies always disliked the black box. But the old contract carried a default: the employer is buying labor time, not buying the lifetime contents of your head. AI Skills change that default. When the company starts asking employees to turn their working methods into a callable Skill, the natural question is: the part I just wrote in — is that my reusable capital? When this Skill runs in my place a hundred times, are you still paying me at the old wage? If the system learns my judgment, who is on the hook when it makes a mistake? If it does not make a mistake, what is my value worth? The old contract cannot answer any of these. So employees answer in two ways. Rumor burns compute. Code seeds noise. One is exaggeration. One is real. One cannot be treated as a factual case. One already exists publicly on GitHub. If the person in the corner office only sees "employees won't share," they are reading shallow. What employees are actually resisting is unbounded distillation. The company asks you to turn over your experience, without rewriting asset ownership, revenue split, accountability, or career protection. You can call that unprofessional. You can call that uncooperative. The reaction itself is a market signal. The signal is simple. The old labor contract cannot protect a new kind of labor asset. Smart employees feel it first. They are not waiting for the law to update. They are not waiting for HR policy to catch up. They are building defenses inside their own docs, their own tools, their own deliverables. That defensive line will not always be elegant. But it is a warning. If the new contract does not get written by the company, employees will write a defensive draft on their own. ## Employer disorder: AI usage is not capability The employee side is defending. The employer side is moving too. The problem is not that bosses don't realize AI will rewrite work. The real problem is that a lot of companies, in the first move, treat "how much AI you used" as "how much capability you have." Once that gets written into the comp plan, things start going sideways. This is the easiest mistake for a boss to make. He thinks he is driving AI adoption. He has just replaced the old punch card with a new AI usage leaderboard. Meta's _Claudeonomics_ is a useful warning sample. Pragmatic Engineer, citing The Information's report, described how Meta built an internal token leaderboard covering more than 85,000 employees, listing the top 250 power users. Meta employees burned 60.2 trillion AI tokens in 30 days. Pragmatic Engineer later wrote that Meta pulled the leaderboard down after social-media blowback. The number is not the most interesting part of that story. The interesting part is that the metric is pointed the wrong way. Token usage is input, not output. Fuel, not mileage. Motion trace, not judgment quality. Let an agent run idle loops, and the leaderboard looks beautiful. Use one well-placed judgment to unjam a workflow, and the token record probably looks unremarkable. If a company treats AI usage as AI capability, it will reward the wrong moves. Employees are not stupid. Reward tokens, they optimize tokens. Reward call counts, they optimize call counts. Reward "looking AI-native," they will make the surface of the work look AI-native. The dashboard ends up full of pretty numbers. The actual judgment quality inside the organization does not move. This feels a lot like the ghost of old KPIs walking back in. The metric used to be online hours. People stayed online. The metric used to be visit count. People padded visits. Now the metric is tokens, call counts, AI usage rate. People will optimize those numbers. The moment a metric is wrong, the smart employee will work it faster than the average employee. This is not "AI wasted compute." That framing is too cheap. This is a contract clause that got written wrong. The old contract bought hours. The new slogan buys usage. Both bypass the actual question: what know-how did the employee actually inject into the system? After the AI runs, which judgments still belong to the human? Look only at usage volume, ignore injection quality and judgment residue, and what the company has bought is noise. I had wanted to put in some rumored cases of companies baking AI usage directly into performance evals. I do not have a reliable public source on a specific company yet, so I am not naming names. That boundary matters. We can discuss the management tendency: when a company turns AI usage into a comp signal, it is unilaterally adding new clauses to the labor contract. But without a verifiable source, I am not pinning that on a named company. The abstract is already enough. The thing the person in the corner office should be asking is not "are my employees using AI." That question is too shallow. Of course you want them using AI, the way you want them using Excel, Slack, the CRM. Tool usage was never the capability itself. Real capability lives in two places. One, whether the employee can inject structured experience into the system. Two, after the system runs, whether the employee can carry fewer but more expensive judgments. If neither of those is in the eval, and only usage is, you are inviting noise. The employee-side anti-distillation is a defense against unbounded extraction. The employer-side tokenmaxxing is a compliance reaction to a broken metric. They look far apart. They are two faces of the same problem. The new labor contract has not been written. The old metrics started running first. So the person in the corner office cannot hand "AI organization" over to a slogan. "AI for everyone" is not a contract. "What every role injects into the system, who approves it, who is on the hook when it fails, how the upside is priced" is a contract. The first one builds leaderboards. The second one builds organizational capability. ## The new contract formula A new labor contract has to count two things. First, how much know-how the employee injected into the system. Second, how much judgment a human still has to carry after the AI runs. I will write it as a rough formula. > New value = know-how injection × judgment residue This is not an academic formula. It is not a financial model. Treat it as a lens the person in the corner office uses to read the organization. You do not need to compute either side to two decimals. You do need to know what you are actually buying. The multiplication is on purpose. Either side near zero collapses the value. A person who sediments experience well but does not carry the key judgment after AI runs is closer to a knowledge-base librarian. A person whose judgment is sharp but never gets injected into the system is still a great solo operator from the old era — the organization cannot reuse it. What the new contract buys is both sides standing up at once. Know-how injection is the degree to which an employee turns their experience into a system-reusable capability. It is not "how many times you used AI." Using AI does not equal injecting know-how. A person who has the tool write their weekly update ten times a day has very low injection. Another person who organizes customer segmentation, pricing rhythm, exception handling, and risk judgment into reusable rules — so that every downstream agent makes one fewer kind of mistake — has high injection. This part used to have no price tag in the contract. The company would say "this is your work sediment." The employee would say "this is my professional capital." Neither side is wrong. The trouble is that AI is, for the first time, making this thing reusable at scale. You used to write a doc. Whether a new hire read it was a coin flip. Now you write a Skill, a rule, a prompt, a checklist, and the system runs it every day. So the high-injection employee is not selling time. They are selling reusable capital. The second piece is judgment residue. Judgment residue is not "things AI can't do yet." That definition is too static. AI moves fast. What it cannot do today, it may do tomorrow. The more accurate definition: after one pass of AI, the share of judgment a human still has to take over, correct, weigh, sign, and own the consequences for. This part also cannot be counted in old hours. A person might spend ten minutes overruling an AI proposal — and those ten minutes decide whether the project goes off course. Another person might spend three hours co-piloting an agent in circles, producing a long output, carrying no consequential judgment. The old contract logs both as "work time." The new contract has to pull them apart. Look at injection alone, you slide toward tokenmaxxing. The company says "use more AI, produce more prompts, sediment more templates," and the employee fills the motion to the brim. More templates, more calls, prettier records. Judgment quality is not necessarily up. The Meta token leaderboard is the magnified version of that risk: the metric rewards input, the employee optimizes input, the ledger looks busy. Look at hours alone, you under-price judgment residue. A lot of bosses are still weighing things with the old scale. How long did you sit. How many meetings did you attend. How many decks did you turn in. After AI, the most valuable part of the work might happen in the shortest time. Seeing where the model is wrong. Knowing when not to automate. Daring to sign for a contrarian call. That is why the new contract has to do double-entry accounting. The know-how injected into the system has to be visible. The judgment residue carried after the AI runs has to be visible. The first one decides whether organizational capability can be reused. The second one decides whether organizational accountability has someone holding it. Skip either side, the contract gets written crooked. For the person in the corner office, this is not a moral statement. What you want is organizational capability, not employee attitude. What you want is judgment quality, not enthusiasm for tools. What you want is reusable capital and accountable decisions, not a pile of "I used AI today" motion traces. The first table of the new labor contract starts here. ## Resetting the accountability layer The new labor contract cannot live only in policy. Policy will say: employees shall use AI reasonably, managers shall carry review responsibility, the company shall protect knowledge assets. Sounds right. Hits the ground as three empty sentences. A contract with teeth has to be written into the system. Peter Drucker, in _Managing Oneself_, talked about the knowledge economy — knowing your strengths, your values, your way of working. Put that in today's context and the next question is already there. When knowledge can be copied and called by AI, a person's value cannot stop at "what I know." It has to land on "which decision I am accountable for." The accountability layer has to be reset. The person in the corner office will eventually ask one plain question. Who pays when it goes wrong. That sentence is more useful than "AI ethics." Colder too. Because as long as accountability is not written cleanly, a lot of AI rollouts will, at the moment of an incident, fall back into the old logic. The employee will say the model suggested it. The manager will say they only reviewed the direction. The system log will say execution succeeded. The customer will only see the company made a mistake. You will discover, at that moment, the company has not become smarter. The company has only fragmented its accountability. The new contract has to unbundle that chain first. One. Every key write action must carry an actorId. AI can generate. Agents can execute. Workflows can run on their own. The moment an action touches customers, money, contracts, accounts, data, or external commitments, "system executed successfully" is not enough. The record has to carry three layers. What the AI did. What the human reviewed. Whose name the system finally signed under. This is not compliance neat-freakery. This is how you book judgment residue. The old contract chased accountability through job descriptions. The new contract chases it through event records. You cannot, on one hand, let AI carry judgment for employees and, on the other, swing the old role-based stick when something breaks. Who gave the instruction, who confirmed the boundary, who approved the write, who took over the exception — all of it has to be queryable. Two. AI failure has to become a first-class workflow node. A lot of companies treat human handover as the fallback for when the system can't cope. That framing is wrong. AI failure is not workflow trim. It is the most important accountability site inside the new contract. Because judgment residue tends to live exactly here. Model drift. Insufficient context. Value tradeoff conflict. Customer risk rising. Policy boundary ambiguous. These nodes do not belong inside chat logs. They belong inside SLAs. Inside tickets. Inside retros. Inside how a role is priced. An employee who consistently catches complex judgment at AI failure points cannot be valued on "tickets closed." A role whose failure points keep dropping and whose handovers keep standardizing has to be repriced too. Three. Injection has to be tracked as an asset. The rules, prompts, cases, checklists, and failure retros an employee writes should not just sit scattered across documents. That is still an old knowledge base. The new contract treats them as reusable assets. Who contributed. Which workflow it serves. How many times it was called. Which errors it corrected. What quality change it produced. Inside my own engineering setup, I have run similar moves. Write-action gates. Human takeover. Structured sediment at the close. In public writing this has to be sanitized — no customers, no repo names, no sensitive paths — but the grammar is sayable. The grammar is this. Judgment has to be tied to a person. Action has to leave a trace. Takeover has to be booked. The person in the corner office does not need an "AI usage policy." Policy manages attitude. The system manages accountability. If the new labor contract does not enter permissions, approval, logs, SLA, retros, and asset ledgers, it is a slide deck. I am going to say that line once more, harder. "Reasonable use of AI" written into policy will not save you on the ground. actorId, approval chain, rollback conditions written into the system will save the company at the moment things break. That is also why HR cannot write this contract alone. It needs the person in the corner office to define the value lens. It needs the business to define the judgment sites. It needs IT and engineering to write the accountability chain into the system. It needs HR to translate it into roles, comp, performance, and career paths. Skip any one of them, and you fall back to the old script — encourage usage, strengthen training, watch for risk. That is not a new labor contract. The core question of the new contract is not "how should employees use AI." It is: after AI participates, which slice of know-how becomes an organizational asset, which slice of judgment stays with the human, who is accountable when something fails, and which kind of accountability needs to be repriced. ## A Luddism comparison It is easy, at this point, to grab a lazy label. Anti-distillation employees — is that a new shape of anti-technology sentiment? The label does not stick. E. P. Thompson's _The Making of the English Working Class_ was published in 1963. The Google Books description sets it inside the formation period of the industrial revolution and how workers and craftsmen built a shared identity. Looking at Luddism today, the heavy weight is not "they smashed machines." It is why they saw the machines as a threat. A lot of modern readings make the point clearly. Luddites did not simply hate machines. What they opposed was something else. Capitalists were using machines to suppress wages, to dismantle craft standards, to bypass long-standing trade conventions, to suddenly destroy the skill and livelihood security workers had built over years. The machine was the visible object. The real conflict was in the invisible contract. That maps cleanly onto today's anti-distillation tools. Employees are not afraid of AI on sight. Many employees know AI better than their bosses, and have already wired it into their own workflows. What actually makes people tense is when "please become more efficient" quietly turns into "please hand over your judgment," and the handed-over judgment gets written into the system as a capability that can later route around the person. If you call that anti-technology, you have read the problem backward. Anti-distillation is not refusal of technology. It is self-help defense when a new contract is missing. The employee knows they cannot block AI. The employee knows the company will sediment knowledge. What they are trying to defend against is unbounded extraction — no revenue split, no accountability reset, no career protection, just "this is an organizational asset." The weavers in 1811 were not smashing abstract industrial progress. They were resisting an arrangement that turned a new tool into a multiplier on old power. Employees in 2026, seeding noise into Skill documents, are not resisting AI itself. They are resisting forced injection with no contractual constraint. The two events sit far apart. The form is very different. The skeleton of the problem is the same. Technology rewrites production. The old contract does not get rewritten in time. Labor protects itself first, in its own way. The person in the corner office should not read this section as "employees naturally rebel." The reading is different. When the company does not price the new capability, does not assign judgment residue, does not write an AI failure accountability chain, employees will build the boundary themselves. That boundary can be crude. It can hurt organizational efficiency. It can make truly valuable experience harder to sediment. So the answer is not to crack down on employees. Crackdown only makes anti-distillation more invisible. The actual answer is to write the contract cleanly. What knowledge must be sedimented. What judgment belongs to the employee as professional capital. What sedimentation earns a return. Which AI outputs must be human-signed. Which accountability cannot be swallowed by the system. The lesson from Luddism is not "technology triggers resistance." The more accurate lesson is this. Technology does not bring a new order on its own. The new order has to be written into contracts, boundaries, and accountability chains. Write it too slowly, and labor writes the defensive script first. ## Hooking to the Wulf matrix By here, this piece has only solved half the problem. We know what the new labor contract has to count. Know-how injection. Judgment residue. Accountability chain. We know why the old contract is not enough. It proxies judgment through time. It wraps accountability inside the role. It bundles attention, experience, and bag-holding into one wage line item. The person in the corner office cannot walk into an org meeting holding a formula. You cannot say the same sentence to sales, engineering, HR, finance, customer service, and ops. The know-how each role injects into AI is different. The judgment residue left after AI runs is different. The sales judgment may live in customer pacing. The engineering judgment in architectural boundaries. The HR judgment in organizational diagnosis. The customer service judgment in risk escalation. If a boss walks into the room with "know-how injection × judgment residue," the room will probably nod. Then go back and keep doing what they were doing. Because a formula can only state the problem. It cannot rewire each role. So the new contract has to keep being translated downward. It has to become a job description. A workflow node. A permission design. An approval table. A retro loop. In other words, it has to move from "labor contract semantics" into "organizational design matrix." That is the work of the next piece. In July 2025, the Wulf team described a spectrum of human–AI collaboration. Part 5 will open that up properly. This piece is not stealing the terminology, and it is not laying out the detail early. The thing left here is one question. If the new contract has to land on every role, what the person in the corner office needs is not the sentence "human in the loop." It is a table. The table has to answer three things. One. Where is the know-how injection point for this role. Experience rules. Customer corpus. Judgment cases. Exception handling. Workflow design. Two. Where is the judgment residue for this role. Approval after AI. Correction. Value tradeoff. Signature. Crisis takeover. Three. What kind of human–AI relationship does this role sit in. AI assists human. AI runs first, human reviews. Human directs multiple agents. Human only steps in at key nodes. Without that table, the new labor contract cannot enter the organization. You stay at slogans. AI for everyone. Roles redefined. Capability upgraded. None of it wrong. After the meeting, the employee still does not know which experience to hand over. The manager still does not know which judgment they have to carry. HR still does not know how to change the comp. IT still does not know how to wire the permissions. Part 4 says: the old contract has cracked. Part 5 will say: once it cracks, how each role gets rewired. This is not HR's table. This is not IT's table. This is the person in the corner office's organizational design table. Because only the corner office can decide which judgments must stay with humans, which know-how must become organizational assets, which AI failures must be written into the accountability chain. If that table does not get built, the company will keep papering over the real problem with three nice sentences. AI for everyone. Roles redefined. Organizational capability upgraded. The sentences are fine. Without the wiring table, they are fog. Next piece, I roll the table out. --- ## Read on - Previous: [Why the HR Three-Pillar Model Breaks](/blog/human-in-the-loop-03-why-hr-three-pillar-breaks) - Series hub: [Human in the Loop](/blog/human-in-the-loop) - Next: [AI Replaces Actions, Not Organization](/blog/human-in-the-loop-05-ai-replaces-actions-not-organization)