Chapter 10The Collapse of the Execution Layer: Evolving the Talent Structure

1. Four thousand people are not four thousand jobs taken by AI #

On February 26, 2026, Block announced that it would shrink the organization from more than ten thousand people to fewer than six thousand. More than four thousand left or entered a consultation process. The company had 10,205 full-time employees at the end of 2025, then reported to the U.S. Securities and Exchange Commission a reduction of more than 40%. By the second quarter of 2026, that restructuring plan had already ended (Block, 2026).

The scale of that change is real. In a letter to staff, Dorsey attributed the decision to a set of changes already accelerating: intelligent tools and smaller, flatter teams made it possible for the company to work in a different way. In a later letter to shareholders he went further: smaller teams, with those tools, could do more, and do it better.

Those facts establish only two things. Block did complete a large layoff, and the company did explain the layoff as a result of AI and organizational redesign. The public filings do not disclose which tasks the people who left had been doing, or at which layer they sat, and they do not prove that more than four thousand jobs had already been taken over by AI. Writing "four thousand people left" as "four thousand jobs replaced by AI" would take a company claim and miswrite it as a verified mechanism.

The more reliable unit of analysis is not an occupation, and not a job title. It is the tasks inside a job. In 2025 the International Labour Organization and NASK used nearly thirty thousand occupational tasks to estimate potential exposure to generative AI: about a quarter of global employment sits in some range of exposure, and the highest-exposure category is only 3.3%. The report distinguishes technical potential from unemployment that has already occurred. Because most occupations still contain tasks that need a person, transformation of a job is more likely than wholesale replacement of the job (ILO & NASK, 2025).

A support person may both look up an order and calm an angry customer. An engineer may both change a field from a checklist and decide a system boundary. A manager may both roll up progress and coach people, allocate resources, handle conflict, and carry accountability. AI can take some of those tasks. It does not, by itself, erase the whole job. Only when enough of the core tasks have already moved, and the remaining tasks can be folded into other jobs, can a standalone role disappear.

When standards and the context that belongs to a decision can be shared, information does not all have to be retold up and down a hierarchy. When the cost of execution keeps falling, much repeated processing also does not have to stay bound to a given job. The first things to loosen are the tasks that exist on a clear instruction, a mature standard, and the hauling of information. Coaching, accountability, conflict, allocation of resources, trust on the ground, and high-risk decisions do not disappear with them.

"Collapse of the execution layer" does not declare that a group of people has lost its value. It is a way of taking a job apart again: which pure-execution tasks are getting cheaper, which judgments and responsibilities still have to be carried by a person, and how the organization should recombine those tasks. The old gene being replaced is treating obedience, busyness, and volume of execution as the value of talent itself.

2. Three kinds of task, not three ranks of person #

S1, S2, and S3 mark how much judgment a task needs. They mark only the concrete task. They do not rank an occupation, a job, or a person.

• S1, execute on a clear instruction. The aim, the steps, the inputs, the timing, and the test of acceptance are already given. The person does the work and reports deviations. Entering a batch of fixed fields into a system, or generating a fixed report from a clear list, sits here.
• S2, execute proficiently to a standard. The aim and the standard are already given. The person can make local choices among mature methods, and does not have to handle a material exception, reopen a ranking of value, or revise the standard. Completing a routine page to a design spec, or handling an ordinary application by an audit standard, sits here.
• S3 and above, handle exceptions, make a trade-off, and take part in revising the standard. When the standard does not cover the case, when standards conflict, or when the conditions have already changed, the person has to state the trade-off, carry the result, and write the exception back into the next version of the standard.

The three kinds of task often sit in one job at the same time. A recruiter entering a candidate's details may be S1. Screening résumés to a standing standard may be S2. Finding that the original standard keeps missing a class of talent, and rewriting it, is S3. The same recruiter can finish all three in one day. Completing data entry in the morning does not make that person "an S1 person." Taking part in a revision in the afternoon does not pin a permanent "S3 rank" on them.

The classification also moves as the standard changes. The first time a class of customer complaint appears, someone has to investigate the facts, weigh the loss, and decide how to handle it. That is S3. After similar events accumulate, the organization writes the conditions of use, the range of compensation, the stop line, and the escalation. The next handling may drop to S2. A task moving from S3 to S2 is not a person falling back. It is the organization settling one judgment as a reusable standard.

The reverse also holds. After the market, the regulation, or the risk changes, a task that could be finished to a standard yesterday may show a conflict of standards today, and have to return to S3. An S level describes how much judgment this task needs under the present standard, the present context, and the present conditions of risk. It does not say which class the task belongs to forever.

That also limits what "collapse of a job" means. That a job contains some S1 and S2 tasks is not enough to judge that it will disappear. The organization first has to take the tasks apart, watch which can migrate, which responsibilities still sit there, and whether the remaining tasks are enough to make a standalone piece of work. Pure-execution tasks can collapse. Jobs will recombine. People should not be pressed into a table of high and low ranks.

Tool training is still useful here. It cannot replace a rebuild of the tasks. People need to learn to use AI, and they also need to learn to check AI, protect data, and handle exceptions. If the organization has never split the tasks, has no standard of acceptance, and has not reallocated responsibility, adding a prompt course to the old job will not make the old structure evolve on its own.

3. What moves first to AI or to outsourcing #

That a task can be written as a standard only shows that the organization has already put some of the judgment up front. The task sits closer to AI or to outsourcing. It has not, by itself, finished the rest of the road. Whether the task can actually migrate has to be run through at least six questions.

1. Can the work travel. Can the input, the process, or the output be digitized, or can an outside supplier finish it elsewhere and deliver it? Physical actions that have to be done on site, and work that depends on sensing in the room, usually move more slowly.
2. Is the standard usable. Have the action, the conditions of use, exception and escalation, the stop line, and the evidence of acceptance already been written clearly? An SOP of an ideal state, with no handling of anomalies, is still not enough to support a stable migration.
3. Can the context be supplied lawfully and in time. Can AI or an outside team obtain the same version of the facts the task requires? Privacy, security, trade secrets, data quality, and limits of permission can all turn "possible in theory" into "cannot be handed over in practice."
4. Can the result be accepted. Can the organization observe quality, sample the process, trace the source, and separate the result of the task from luck? If an error shows itself only months later, so-called automation may only have hidden the cost of checking until later.
5. Is the error controllable. Can a mistake be found, stopped, and recovered in time, and is the cost bearable? Irreversible, high-risk, and tightly regulated tasks, even when part of them can run automatically, often still need a person's review or approval.
6. Is accountability still clear. Who decides the standard, who supplies the context, who oversees the supplier, and who is responsible for the final result? Handing the work out is not the same as handing the responsibility out with it.

AI and outsourcing are also not the same road. AI depends more on machine-readable context, stable tool calls, and repeatable acceptance. Outsourcing also requires that the interface of the task can be written into a contract, that a suitable supplier exists in the market, and that the organization keeps the capacity to oversee, audit, and exit. OECD research on "task trade" shows that being digital and being separable is not the same as being worth outsourcing. Transaction costs, interdependence among tasks, solving problems on site, and incremental innovation can all lead an organization to keep a set of tasks inside (OECD, 2011; 2023).

Regulatory duty especially cannot be moved by a contract or by a model. The European Banking Authority's outsourcing guidelines state that the management of a financial institution cannot outsource its responsibility for the institution and its activities. The EU AI Act also requires that human oversight of high-risk AI be proportionate to the risk, the degree of autonomy, and the context of use, and that the overseer be able to ignore, override, or stop the system's output (EBA, 2019; EU, 2024). Whether the organization uses AI or an outside supplier, it has to keep a clear chain of accountability.

Trust has its own speed. A client will add constraints that are not on the form. A patient will complete uncertain symptoms step by step in a consultation. A safe reporting channel will let bad news enter the organization's record. Those interactions themselves produce context. AI can arrange material and propose a candidate. It cannot, for that reason, take over the promise and the responsibility in the relationship. The more mature the standard, the easier it is for repeated execution to migrate. The more important the relationship, the higher the risk, and the clearer the final accountability, the less a person can be left with only an approval button.

Ingka Group, IKEA's largest retailer, launched the chat tool Billie in 2021. Company statements and Reuters reporting show that Billie later handled about 47% of routine inquiries. The company said that from 2021 to 2023 it resolved about 3.2 million interactions and saved nearly 13 million euros. In the same period, 8,500 call-center staff received training in remote interior design, digital sales, relationship work, and handling of complex problems (Ingka Group, 2023; Reuters, 2023).

The mix of tasks then changed: routine queries went to the tool, and trained staff moved toward consulting that depends more on taste, relationship, and complex judgment. Ingka also disclosed that in fiscal 2022, sales from remote customer contact points were 1.3 billion euros, 3.3% of total sales. That revenue cannot be attributed to Billie or to the retraining. The public material also does not give a complete before-and-after of the 8,500 people's actual moves, retention, service quality, and income. The evidence on hand can only confirm that migration of tasks and retraining of people happened at the same time. It cannot prove that IKEA has already completed a company-wide transformation, and it cannot estimate a typical operating effect of this kind of rebuild.

4. Block: a drawing still under construction #

Block has published an organizational design whose results have not yet been verified.

On March 31, 2026, Dorsey and Sequoia partner and Block director Roelof Botha published From Hierarchy to Intelligence. The essay reads a traditional hierarchy as a way of routing and coordinating information: facts are rolled up from the front line, and decisions are passed down from above. The two authors argue that AI, for the first time, makes it possible for a company to keep a live model of its own operations, so that some of the routing no longer depends wholly on the management chain. In that design, people move step by step into three kinds of role (Dorsey & Botha, 2026).

Individual contributors (ICs) build and operate the company's capabilities, models, intelligence layer, and product interfaces. They are deep specialists in one layer. They should obtain the information an action requires from shared context, rather than wait for a manager to retell it.

Directly responsible individuals (DRIs) are accountable for a cross-functional problem, an opportunity, or a customer result. Inside an agreed time and permission, they can call on resources from different specialist teams. The example in the original is not "find a permanent boss for every matter." It is giving a concrete problem a named, changeable owner.

A player-coach does the work and, at the same time, develops the specialist capability of the people nearby. They still write code, build models, or design, and they also demonstrate, give feedback, and coach. This role replaces the traditional manager whose main work is forwarding information. It does not delete coaching, conflict, and the building of capability along with that manager.

The three roles themselves also show that "a hierarchy only hauls context" is too sweeping a claim. Block gives strategy and priorities to the DRI, professional growth to the player-coach, and deep building to the IC. The functions of management have not all disappeared. They have been taken out of one job package and given again to the owner of a problem, the specialist contributor, the person who develops people, and the system.

The hub of the design is what the company calls a "world model." The March 31 essay treats decisions, discussions, code, design, plans, problems, and progress as raw material. The system is used to keep answering what is being built, where it is blocked, how resources are allocated, and what is working. In a later text transcript published by Sequoia, Dorsey listed more concrete material: remote work leaves traces such as Slack messages, email, commits, documents, and recorded meetings, and the intelligence layer is built on those records that can be kept (Sequoia, 2026).

Connecting material to a model is not the same as everyone seeing all material. Unified context only requires that the people responsible for the same class of decision obtain, inside their permissions, the same version of the necessary facts. Pay, customer privacy, unannounced deals, security incidents, and regulated data cannot lose their access boundary because of the name "world model." Block's public essay does not set out staff notice, legal basis, least privilege, retention period, isolation of sensitive information, or a path of appeal. Those conditions decide whether the system can run safely.

The target for layers has to be kept apart from what has already happened. In the transcript, Dorsey said the greatest depth from himself to a member of the company was then about five layers, and that the aim within the year was two to three; "all about six thousand people reporting directly to me" was listed as an ideal. That is not a fact already achieved in the March 31 essay, and still less a structure already tested against operating results.

Block has already deployed some tools. In March 2026 the company said that the internal agent-orchestration tool goose was already in use in engineering, finance, legal, and design, and self-reported that from September 2025 production code delivered per engineer had risen 40%. Another risk model had been shortened from a planned quarter to two days. The company also said that some automated flows were close to finished, and others were still at an earlier stage (Block, 2026). Those numbers can show that the tools are in actual use. They cannot prove a causal effect on code quality, customer results, or the three-role design.

The essay explicitly leaves intuition, sense of direction, cultural context, relations of trust, ethical judgment, new situations, and high-risk moments on the human side. Filings to regulators also list risks of the restructuring and of AI, including operating disruption, cybersecurity, failure of controls, regulatory compliance, and expected gains that may not be realized.

The material presents a drawing still under construction. The layoff has already happened. Some tools have already been deployed. The roles and the layers are still aims. Operating results have not yet completed acceptance. Whether shared context can stably take over some of the routing still has to be tested later by customer results, risk, rework, and load on people, together.

5. How the people who remain work together #

When the pure-execution tasks inside a job drop away, the relation between a person and the organization changes. Over the long run, more members need to work in the manner of a partner. Partner-style work here means three things: taking part in judgment inside the tasks and permissions granted, being accountable for one's own trade-offs and commitments, and bringing results and exceptions back into the standards cycle.

It is not a legal status. That everyone works in the manner of a partner does not mean everyone holds the same title, equity, pay, vote, or strategic decision right. It also does not mean the company can shift operating risk, statutory duty, and labor protection onto the individual. A person can carry a large responsibility in product judgment and still have no permission to sign on finance. Partner-style work changes the relation of responsibility. It does not turn everyone into the same kind of person.

Four hiring standards check four conditions of match between a person and the present tasks.

First, whether values align. What has to align is mission, the floor of conduct, the rules of evidence, and the red lines that cannot be crossed—not opinions, history, manner of speech, or every plan being the same. A person can oppose the boss's product judgment, so long as they supply evidence as it is, respect the shared rules, and carry their own commitments. Reading "values align" as "thinks as I think" mistakes similar views for shared value, and also lowers the visibility of dissent.

Second, whether they can think. Thinking is not a more elaborate sentence. It is finding a gap in a standard, asking about the conditions of use, separating facts, assumptions, and preferences, stating a trade-off when standards conflict, and revising when new evidence appears. Finding the gap, borrowing a standard, judging migration, and verifying a revision can all be settled as observable work.

Third, whether they are self-driven. Self-drive is not acting without permission. It is pushing forward inside the permission, filling in necessary context on one's own, coming back with a plan when blocked, and escalating before a risk crosses a boundary. Waiting for an instruction at every step, and ignoring the boundary, both fail the test of responsibility.

Fourth, whether they can execute to the extreme. Execution has not gone out of date. What the organization opposes is only being able to obey. It does not oppose turning a correct judgment into an accepted result. Extreme execution requires a member to turn a decision into a delivery that can be accepted, to hold quality, the stop condition, and the promise, and to report a deviation in time. Speed and volume, if they are not joined to a customer result, to quality, and to financial value, are only worship of execution in a new wrapping.

Treat People with Honesty is the floor of conduct under the four lines. A claim has to be verifiable. Material information that would change someone else's decision cannot be hidden. A promise has to be deliverable. The other party's capacity for informed choice has to be kept. A person who thinks fast and moves hard still fails this standard if they advance the work by hiding bad news or by manufacturing an information gap.

Screening and development answer different questions. Screening judges whether a person currently accepts the floor of conduct, whether they have the minimum capability the task needs, and whether the job and their choice match. Development supplies context, demonstration, feedback, practice, and a responsibility that widens step by step, and is used to train judgment on a concrete task. Development cannot guarantee that every person fits every task. Screening cannot spare the organization the duty of developing people. When a mismatch continues, the organization can adjust the task, reconfigure the job, or end the collaboration. When the conditions of development are missing, a gap in capability cannot all be booked as a personal problem.

bioby.ai's internal record has not been audited. An early commercial job hired close to twenty people; none stayed as a lasting fit. That record can only show that the hiring standard and the job design of the time did not produce a lasting match. It cannot prove that "talent can only be screened." After the four standards were formed in April 2026, the company first used them to look back at jobs already in place, rather than only to screen new candidates. A standard that constrains only later arrivals, and does not check present power and the present definition of performance, will make two sets of rules between old and new members.

The same internal record also includes a rebuild of the R&D unit: some people moved toward writing standards, splitting tasks, and reviewing results, and several interns took part in development with AI assistance. The company's own figures for cost, output, and defects will be written in full in Chapter 14. They have no independent audit, and cannot carry external proof. More important is this: quality fell in the first month, and the team reported improvement only after it had built the quality standard. That stretch of experience can support, at most, a narrower judgment: expanding execution first and filling in the standard later will enlarge the mess. Whether AI, senior members, and junior members can work together depends on whether the standard, the oversight, and the feedback arrive at the same time.

Interns cannot be listed with AI as "capacity" that scales with the work. When an intern takes part in real work, the organization also carries a learning aim, reasonable oversight, timely feedback, fair conditions, and a path of growth. Counting a learning period only as cheaper execution cost falls back into worship of execution.

6. Five levels of decision rights: rights travel with the task #

After the organization has given members standards and context, a harder question remains: on this task, how far may a given person decide?

Five levels of decision rights describe that path of migration.

1. Instruction. The manager gives the result, the standard, the steps, and the timing. The member executes and reports deviations. It applies to a new task, a new member, and any case whose cost of error cannot yet be borne.
2. Guidance. The manager states the result, the standard, and the boundary of risk, and demonstrates the key method. The member makes local choices, and they review together after the work.
3. Consultation. The member proposes a plan, the evidence, and the trade-off. Both decide together. Who holds the final signature has to be written in advance. Consultation cannot become a blur of accountability.
4. Authorization. The manager states the result, the resources, the permission, the cap on risk, and the conditions of escalation. The member decides inside the boundary and is accountable for the result.
5. Delegation. The member may not only decide the method, but also define the problem, allocate resources, and revise a local standard. Mission, law, ethics, and the organization's final accountability remain red lines that cannot be moved.

The five levels are not a person's permanent title. A senior engineer can be authorized on an architecture choice and take instruction on a payment approval. A sales lead can adjust an ordinary customer plan on their own, and must escalate when a regulated industry or an excess discount is involved. The same person, facing different tasks, different risks, and different context, can hold different decision rights.

That also dissolves the surface contradiction between S2 and S3. Once a task needs a material exception outside the standard, a weighing of conflicting values, or a revision of the standard, the classification boundary moves from S2 into S3. That boundary can be clear. A person's growth can still be gradual. A member first sees, under guidance, how a judgment is formed, then proposes a plan in consultation, then obtains authorization inside a boundary. Years of repeated execution will not grow judgment on their own. Practice of judgment, with feedback, can accumulate capability.

Migration of decision rights has to look at least at three things at the same time.

The first is the concrete task. Does the member understand the standard, the context, the exceptions, and the stop condition of this task? A result in one domain cannot automatically be exchanged for power in another.

The second is risk. Is the error recoverable? Does it involve regulation, privacy, safety, material money, or an irreversible effect? When risk rises, even if the member's capability has not changed, decision rights may be recovered for a time, or a dual review and a legal signature may be added.

The third is judgment already shown. Watch the result, and watch the process: whether the member stated the basis, escalated before the boundary, surfaced bad news on their own, and could revise a standard from a counterexample. A single success may come from luck. An oral account is also not sufficient evidence. What has to be watched is behavior and results that can be traced over a stretch of time.

Every authorization should therefore be written as a concrete task card, not as a vague grant. The card at least writes the range of the task, the expected result, information permissions, usable resources, the cap on risk, the term, the situations that must escalate, the time of review, and the conditions of withdrawal. Demanding that a member be accountable for the result when the standard is incomplete and the context is not given is treating a guess as authorization. Keeping a member at instruction for a long time when information and the standard are already in place will let judgment atrophy while they wait.

Migration is also not a one-way promotion. When the business environment changes, knowledge expires, a member's performance declines, or a run of results exposes a new risk, authorization can fall back from delegation to consultation, or from authorization to guidance. A move back is not a demotion of the person. It is a rematch of the task and the risk. After the conditions recover and the evidence accumulates again, the rights can still move forward.

High-risk work especially needs restraint. The EU's oversight requirement for high-risk AI emphasizes that the person responsible must understand the system's limits, watch for automation bias, and be able not to adopt, to override, or to stop the output. The same principle holds inside an organization. Delegation is not abandonment of governance. Outsourcing is not outsourcing of accountability. Automatic execution by AI also does not erase the person who signs at the end.

The aim of management is therefore to put the right of judgment with a person who has the necessary context, can carry the matching risk, and has already produced evidence of judgment—not to delegate as far as the ladder will go. The value of a player-coach sits here too. They do not decide forever in a member's place. Through demonstration, feedback, and shared review, they let more tasks move safely from instruction to guidance, consultation, and authorization.

7. ofo: when the right of judgment freezes #

Block's design takes information and decision rights out of a fixed hierarchy and gives them to a shared system and to the owner of a concrete problem. ofo's governance shows the opposite mechanism. When a key power is bound for a long time to an identity and to a set of clauses that hold one another in check, organizational action can still be blocked by a voting mechanism even after the window of financing and competition has already shortened.

Public channels do not hold the original ofo shareholder agreement, articles, or board rules. In 2018, Tencent Qianwang (《潜望》), citing people close to the company, said the ofo board then had nine seats, and that Dai Wei and four co-founders, as concert parties, held five votes. Routine matters could be decided by a majority. Decisions involving financing, however, required the signature of the whole board. On that basis the report called a block formed by any one side's refusal to sign a "one-vote veto" (Tencent Qianwang, 2018).

That same year, 蓝洞商业 interviewed investors and gave a close but not identical version: the relevant protective clauses might cover financing, new issuance, merger, and a qualified listing, but the concrete range still depended on the legal documents signed in each round. Strictly speaking, the event that actually entered a formal veto was mainly the merger with Mobike at the end of 2017; other financings often stopped, because of opposition from the parties, before a formal vote (蓝洞商业, 2018). Whether Alibaba held an independent veto, and whether each side could block on its own, also conflicts across public reports.

The merger plan itself can also only be cited from reporting. An anonymous investor who took part in the talks said the plan had considered Cheng Wei as chairman, Dai Wei and Wang Xiaofeng as co-CEOs, and Dai Wei as sole CEO after six months. The report said Dai Wei did not accept a plan that included giving up the relevant powers, and the merger did not continue. Another sentence that often appears—"even in a merger, we would be merging Mobike"—comes from staff recounting an internal speech. It is not Dai Wei's public original wording, and therefore cannot stand as an unqualified direct quotation.

Clauses that protect the founding team and the investors, once the parties cannot agree, make every material choice pass through the same set of signature gates. Founders can stop a merger. Investors can also let later financing stop in the talks. Power was not reallocated by cash, by the competitive window, and by the judgment actually shown. It stayed attached to identity and to historical agreements.

In February 2018, media reporting from the National Enterprise Credit Information Publicity System said ofo had taken two loans totaling 177 million yuan, with shared bikes as movable collateral. Investigative reporting described the mortgage loans as a substitute structure easier to pass than new equity financing. It can show that, after equity financing was blocked, the company took another road. It cannot, on its own, prove that a given shareholder formally exercised a veto, and it cannot be written as if that were then the only financing that could pass.

By December 2018, the online queue for deposit refunds had passed ten million, and by noon the next day it had already exceeded 10.70 million (CNR.cn / 央广网, 2018). That result shows a severe crisis. It does not assign the whole crisis to one clause. Capital burn in bike-sharing, speed of expansion, fleet operations, competition, the business model, the financing environment, and management judgment were all acting at the same time. The investigative reporting itself also says plainly that a one-vote veto cannot carry the whole responsibility for ofo's failure.

Once the right of judgment is frozen, a protective mechanism can, in a critical window, become an obstacle to action. Block and ofo are not an operating win-loss experiment in the same industry, the same stage, and the same definition. One is an organizational drawing not yet accepted. The other is a governance deadlock that already happened. They can be compared only as a mechanism: one side lets information and responsibility flow with the problem; the other shows the constraint on action after power has been locked to identity and to clauses.

8. Dynamic matching, not a one-time rank #

Dynamic lifecycle management, applied to the talent structure, is not a one-time organizational verdict. It is a round of matching that will be run again.

The first step is to locate the job and the person. Split the job into real tasks, and judge whether each is currently S1, S2, or S3. Then record the evidence of judgment the member has already shown, rather than pinning a total rank on the person. What has to be watched includes: on which tasks they understand the standard, what exceptions they have handled, whether they can surface risk, and whether the result can be traced.

The second step is to locate what the organization needs. What result does the company actually need now? Which capabilities belong to the long-term core, and which tasks are only a peak of one stage? Are the standards and the context already ready? What boundaries do physical execution, regulation, trust, privacy, safety, and final accountability set? If the organization has not stated those questions itself, it cannot honestly judge whether a person or a tool matches.

Only the third step is matching and choosing. For each task, choose AI, outsourcing, keeping the work with people, development, screening, or reconfiguration. For the judgment tasks that stay in a person's hands, then set instruction, guidance, consultation, authorization, or delegation. 10/10/80 here marks where the weight of resource goes: a small share of time to locate the tasks and the person, a small share to locate what the organization needs, and most of the energy to actual matching, trial, acceptance, and adjustment. It is not a precise ratio from a survey. High-risk tasks can take more judgment up front.

Figure: S1, S2, and S3 mark tasks. Five levels of decision rights mark permission on a concrete task. After the task, the risk, the capability, or what the organization needs changes, locate and match again.

When an AI trial produces a new error, a former S2 task may have to fall back to S3. When a member handles exceptions well over a run, a task can move from guidance to authorization. When a regulation changes, power already given out may also be recovered. The return in the figure means that new evidence will push a re-marking of the task, a rematch of people, and an adjustment of decision rights.

Reuters reported Meta's Project OT in 2026. The report rested on internal files, posts, and recorded material, and on interviews with more than twenty people with knowledge of the matter. Meta had tested shrinking traditional product teams of ten to twenty people into groups of three to five, reducing the middle layer, and using "agent-assisted analysis" to help set daily priorities. By June, at least eleven units had used the small-group structure (Reuters, 2026).

Internal material cited in the same report showed that code changes on internal platforms and infrastructure were up 220% year over year, while new features or upgrades that actually reached users were up only 36%. Material technical and security incidents were up 40%, and time staff spent fighting fires was up by as much as 70%. Meta later cancelled a planned second wave of the rebuild. The company confirmed Project OT, the two-wave plan, and a scenario in which some teams would shrink by as much as 60%. It did not confirm those internal metrics. Reuters also could not determine a single reason management changed the plan.

The material is not enough to say that "AI replacing the middle layer is bound to fail." The boundary that can be observed is more concrete. Volume of code is not customer value. Before reducing layers, permissions, reliability, oversight, and accountability have to be arranged first. After results diverge, the organization also needs a mechanism to stop and to move back.

IKEA and Meta are one sample each. Industry, organizational conditions, and the definition of results also differ. A median cannot be computed, and a typical effect cannot be estimated. What the two sides support is limited to mechanism: simple tasks can move to AI at the same time as people are retrained and complex tasks are recombined; if delayering runs ahead of permissions, reliability, and accountability, volume of output can come apart from user value and from risk. They cannot support "every firm should adopt the same organizational form."

After the talent structure is redrawn, the constraint turns to time. Judgment tasks need blank space, continuous attention, reflection, and revision of standards, and they are easily filled again by meetings and short-term execution. The question is no longer only "who does this." It also includes "where they spend a day." The mix of hours therefore becomes the next organizational choice after the talent structure.

What to Do Monday Morning (principal-leader view) #

Do not begin by pinning labels on the whole company, and do not begin by setting a layoff target. Pick a job or a workflow that repeats every week and whose result can be observed, and complete the first round.

1. Split the tasks. List, item by item, the tasks this job actually carries. Write the input, the result, and the present owner of each. Do not let a job title stand in for the work.
2. Mark the tasks. On the present standard, context, and risk, mark each task S1, S2, or S3. Then check the six conditions of migration: whether the work can travel, and whether the standard, context, acceptance, risk, and accountability are in place.
3. Look at what the organization needs. Write which results this stage actually needs, which capabilities must stay inside, and which physical, regulatory, trust, and accountability boundaries cannot be moved.
4. Make the match. For each task, choose a next step: a small AI trial, a small outsourcing trial, keeping the work with people, development, screening, or reconfiguration. For the judgment tasks that still sit with a person, write the level of decision rights, the range, the cap on risk, the term, and the conditions of escalation.
5. Let the results return. Agree a time of review. Check the match with quality, customer results, risk, and real cost. After an exception appears, revise the standard, the classification of the task, and the decision rights. Do not set a uniform threshold such as "thirty percent" that has no source.

The same can be done for a person. Do not ask "which rank am I." Ask "of the tasks I carry today, which judgments have already been finished by someone else, which exceptions still need me, and on which tasks can I produce evidence of reliable judgment." The answer decides whether the next step is to learn a tool, fill in context, practice judgment, or ask for a responsibility whose boundary is clear.

Chapter Acceptance Self-Check (against chapter acceptance standards) #

1. Claim restatable in one sentence ✓: collapse first in pure-execution tasks inside a job; S1/S2 that can travel and that meet standard, context, acceptance, and risk move to AI or outsourcing; jobs recombine; S3 judgment and accountability stay with people; decision rights migrate by task, risk, and judgment shown.
2. Eight-section spine and whiteboard figure ✓: four thousand people as tasks not jobs → S1/S2/S3 mark tasks → six conditions of migration / IKEA → Block as a drawing under construction → partner-style work and four hiring lines → five levels of decision rights → ofo freeze → dynamic matching / Meta. Figure is locate / locate / match (10/10/80) with a return; pyramid-plus-execution-belt is out.
3. Evidence caliber ✓: Block from SEC 10,205 / >40% / Q2 2026 end of plan; three roles and world model as design; goose and +40% code as company self-report; layer target as aim not achieved fact. ofo from Tencent Qianwang 2018 and 蓝洞商业 2018, originals of agreements missing, veto list in conflict, 177 million yuan mortgage, 10.70 million deposit queue; no $4 billion valuation; no single-cause of failure. IKEA/Ingka 47% / 3.2 million / ~13 million euros / 8,500 trained / FY22 remote 1.3 billion euros not attributed to Billie. Meta Project OT from Reuters 2026 internal material, company did not confirm the metrics. ILO & NASK 2025; OECD 2011/2023; EBA 2019; EU 2024. bioby.ai unaudited; R&D numbers wait for Chapter 14.
4. Case contract ✓: Block for mechanism design, not a verdict of transformation; ofo for freeze of judgment rights, not a matched operating experiment; IKEA and Meta as one sample each of mechanism, no median. Roman legion, S ranks on people, S4/S5, everyone a partner as legal status, personal ROI, 30% S3+ threshold, nine-intern execution belt as proof, and Didi–Kuaidi as an iron rule are out.
5. Classification and rights ✓: S levels mark tasks; five levels of decision rights mark permission on a concrete task and can move both ways; no permanent personal rank.
6. Chapter boundary ✓: daily three questions and Feynman wait for Chapter 11; full R&D ledger waits for Chapter 14; intern labor is not scalable "capacity."
7. Fluency ✓: rewritten in English voice from the Chinese authority; current prose-standard.