Span of Accountable Exposure

Abstract

Organizations are removing their middle management layer at scale, and artificial intelligence is the stated justification. Survey and layoff data through 2026 show management layers being trimmed across industries, with middle managers absorbing a disproportionate share of white-collar reductions and a widely cited forecast that a fifth of organizations will use AI to eliminate more than half of their middle management positions. The prevailing argument for this is that agents absorb coordination, and that coordination was what the layer did.

This paper argues that the argument is incomplete in a specific and consequential way. The middle management layer held at least six distinct functions, and agents absorb some of them well, some partially, and one not at all. That one is accountability: the capacity to bear a consequence, which requires assets, standing, licensure, or a career to lose, and which an agent does not possess in any form. Removing the layer therefore does not remove accountability; it relocates it, either upward onto executives who cannot inspect what they are now answerable for, or downward into units small enough to hold it.

Haier's RenDanHeYi transformation is the closest available natural experiment, having removed roughly ten thousand middle management positions and reorganized into several thousand micro-enterprises with rights over decisions, personnel, and profit distribution. We read that transformation not as a template but as evidence about the residual, and we identify its enforcement mechanism as market exposure borne by the delegate rather than as any check on the decision itself. From this we derive the paper's central prediction: the binding constraint on organizational unit size shifts from how many people a manager can supervise to how much exposure a human principal can credibly answer for, which we name span of accountable exposure. We then argue that the resulting topology resembles the micro-enterprise while being unable to run on the micro-enterprise's ex post discipline, state the recentralization counter-hypothesis in its strongest form, and specify the conditions under which the argument fails.

1. The Flattening Is Real and the Argument for It Is Incomplete

The removal of the middle management layer is no longer a forecast. Korn Ferry's global workforce research reports that a substantial share of professionals say their organization eliminated management levels in the preceding year, analysis of layoff data reported by Bloomberg found middle managers accounting for close to a third of white-collar layoffs in 2023 against roughly a fifth in 2018, and Gartner has projected that through 2026 approximately one in five organizations will use AI to flatten structure and eliminate more than half of current middle management positions (Korn Ferry, 2025; Live Data Technologies, as reported in Bloomberg, 2024; Gartner, 2024). Individual firms supply the texture. Amazon set and met a target of raising the ratio of individual contributors to managers, Meta restructured around what its internal communications called AI-native design principles while flattening layers, and Oracle and Coinbase both reduced management populations while citing minimal-management efficiency as a goal.

Two cautions belong here before the argument proceeds. First, delayering is a recurring cycle rather than a novelty, with a comparable wave running through the late 1980s and 1990s for the same stated reasons of cost and decision speed, which means the current episode should be read against that base rate rather than as a unique event. Second, the causal attribution to AI is largely self-reported by the firms doing the cutting, and cost reduction has an obvious interest in acquiring a technological justification. The evidence establishes that the layer is being removed. It does not establish that agents are the reason, and this paper does not require that they are.

What matters for the argument is the justification being offered, because it is the justification that is incomplete. The reasoning runs that middle managers coordinated, reported, scheduled, and routed, that AI now performs those functions, and that the layer is therefore redundant. Each step in that chain is defensible. The conclusion does not follow, because coordination was not the only thing the layer held, and the functions it held were bundled together for reasons that had nothing to do with anyone's preference.

2. What the Middle Manager Actually Held

Unbundling the role is the analytical move this paper depends on. A middle manager held at least six distinct functions, and treating them as one thing is what produces the incomplete argument.

Routing. Information moved up and down through the layer, aggregating on the way up and decomposing on the way down. This is the function most often cited and the one agents absorb most completely, since routing is a mechanical operation on structured and unstructured information and machine systems now perform it faster and with less distortion than a human chain.

Coordination. Work spanning units was reconciled through managers who negotiated sequencing, resource contention, and interface definitions. Agents absorb a large portion of this where the coordination problem can be expressed in shared state, and they absorb almost none of it where coordination depends on relationships, obligations, and informal reciprocity across organizational boundaries.

Exception judgment. When a rule did not obviously apply, a manager decided what it meant in this case. Agents perform this competently inside a defined boundary and cannot perform it at all outside one, since judging whether an exception is admissible requires knowing what the institution is trying to protect, which is precisely the knowledge that was never written down.

Tacit context. The layer held an understanding of what a policy means here, which customers matter for reasons not in the record, and which past decisions constrain the present one. Agents can absorb this only to the extent it is made explicit, and the act of making it explicit is expensive, incomplete, and frequently reveals that the organization did not agree with itself.

Authority relay and attenuation. Delegated authority passed through the layer, and a competent manager narrowed it as it descended, granting a subordinate less than they themselves held and taking back the difference when conditions changed. Agents can carry authority and can be constrained in how much they carry, but they cannot originate it, since institutional authority originates only with a principal capable of bearing responsibility for the grant.

Accountability. Someone with a name was answerable for what the unit did. This is the function that cannot be transferred to an agent under any architecture, and the rest of this paper is an argument about what follows from that.

3. Accountability Is the Residual

An agent cannot be accountable in the institutional sense, and the reason is not philosophical squeamishness about machine moral status. Accountability requires the capacity to bear a consequence, and bearing a consequence requires something to lose. A human manager holds assets, compensation, professional standing, licensure in regulated fields, a career with a future, and in extreme cases legal liability. An agent holds none of these. It can be deprecated, which costs its operator and costs the agent nothing, because there is no subject there to experience a cost.

This produces an asymmetry that most agent deployment plans do not confront. Authority can be delegated to an agent, and evidence can size that delegation responsibly, and enforcement can hold its boundary at execution. Accountability cannot make the same journey. It stops at the last human principal in the chain, and every action the agent takes downstream of that point remains that principal's to answer for.

Removing the middle management layer therefore does not remove accountability from the system. It relocates it, and there are only two destinations. It can move upward, concentrating on executives who become answerable for decisions produced at machine speed across a surface they cannot inspect, which is a formal accountability that will not survive contact with a serious failure and which the Convergent Enterprise thesis names institutional incoherence. Or it can settle into units small enough that a named human can credibly answer for what happens inside them.

The second destination is a micro-enterprise in everything but name, which is why the Haier experiment is worth examining closely rather than admiring from a distance.

4. The Haier Natural Experiment

Haier is the largest and longest-running attempt to operate a major industrial firm without a middle management layer, and it predates the current wave by roughly two decades. Beginning in 2005 and implemented substantially by 2012, Zhang Ruimin dismantled the layer and reorganized the company into several thousand micro-enterprises. Zhang's own account puts the number of eliminated middle management positions at approximately ten thousand, while secondary sources cite figures ranging to thirteen thousand, and the resulting structure is generally described as roughly four thousand micro-enterprises operating as autonomous units with their own profit and loss (Knowledge at Wharton, 2018; Thinkers50, 2023; Denning, 2022). Group headcount fell from more than eighty thousand in 2012 to approximately sixty thousand by 2016 during the transition to the model's second phase, a reduction facilitated by provisions of Chinese employment law permitting restructuring layoffs where a firm materially revises its business method (Mol and colleagues, 2018).

Three structural features carry the model. Micro-enterprises hold what Haier describes as three rights, covering decision-making, personnel selection, and profit distribution. Central functions including human resources, finance, and information technology operate as service platforms that micro-enterprises may use or decline in favor of external providers. And compensation rests on value created for users rather than on position, with base salary minimized and profit sharing at the unit level, such that a unit which cannot sustain itself dissolves and its people move to other units.

The phrase zero distance belongs to this model and to Zhang, who used it to describe the elimination of distance between the employee and the needs of the customer. That is a relational claim about intermediation rather than a temporal claim about decision latency, and the distinction matters for anyone importing the phrase into an argument about machine speed.

What is genuinely instructive here is the enforcement mechanism, which is easy to miss because it does not look like governance. Haier did not make its delegations enforceable in any architectural sense. The three rights are grants from the platform, revocable by the center, and no party outside the firm would learn if they were narrowed. Under the taxonomy developed in the Convergent Enterprise thesis, the model is a house rule regime, and an unusually well-run one. Its constraint comes from somewhere else entirely: the unit bears the economic consequence of its own decisions, and that consequence is witnessed by users who pay or do not. Zhang solved the problem of the grantor who can rewrite the rule not by binding the grantor but by outsourcing the witness to the market.

That is a third mechanism, and it is absent from the two-state model of house rules and externally witnessed clauses. Section 9 restates the taxonomy to include it.

5. Why RenDanHeYi Did Not Travel

The model's diffusion record matters because it tells us which parts were Haier and which parts were general. Haier-affiliated research centers report very large numbers of firms studying or adopting the model, and those figures should be treated with caution, since they originate from institutions with an interest in the count and use adoption criteria that are not independently defined. The stronger evidence is what happened in Haier's own acquisitions and in the academic record.

Four obstacles recur. The first is legal and contractual. Senior Haier managers in United States and European subsidiaries reported that local employment law and entrenched employment relationships were incompatible with a compensation model that minimizes base salary and pays from unit performance, and the Chinese restructuring provisions that enabled the original transition have no equivalent in those jurisdictions (Mol and colleagues, 2018). A model that requires the ability to convert employment into something closer to internal entrepreneurship does not survive a legal regime built to prevent exactly that conversion.

The second is the founder paradox. Dismantling hierarchy at Haier required a founder with sufficient authority to impose an anti-hierarchical model against internal resistance, and to keep imposing it through repeated restructurings. The instrument used to remove hierarchical authority was hierarchical authority, which is available in a founder-led firm and largely unavailable to a professional chief executive with a board, a quarterly reporting obligation, and a five-year tenure.

The third is transaction cost. Internal contracting between units reintroduces negotiation, measurement, and dispute costs that hierarchy exists to suppress. Coase's account of the firm holds that activity is brought inside the boundary precisely when internal direction is cheaper than market transaction, so a model that converts internal direction into internal transaction should be expected to raise coordination cost, and the question is only whether the gain in responsiveness exceeds it (Coase, 1937).

The fourth is the function without a customer. Units optimizing their own profit and loss produce cross-unit externalities, and activities with no external buyer sit badly in a model where everything must be sold to someone. Safety, compliance, long-horizon research, and platform integrity are precisely the functions that cannot be organized as micro-enterprises without either subsidizing them, which reintroduces central allocation, or starving them, which reintroduces the failures they exist to prevent.

Each of these four is a coordination-without-hierarchy problem. Each requires shared state that units can act against without a manager to reconcile it, and that shared state did not exist in 2012 in any form a firm of eighty thousand people could operate. This suggests a reading of the diffusion failure that is more interesting than concluding the model was wrong: Haier attempted a structure whose enabling substrate had not been built. Continuously reconciled enterprise state is what the convergence argument says is now becoming available, which makes the obstacle that stopped the model the thing that AI supplies.

That reading is a hypothesis rather than a finding, and it is one of the places this paper is most exposed.

6. Span of Control Becomes Span of Accountable Exposure

The classical limit on organizational unit size was cognitive and relational. Graicunas modeled the combinatorial growth of relationships a superior must manage as subordinates are added, and Urwick and the administrative management tradition converted this into prescriptions clustering around five to seven direct reports (Graicunas, 1933; Urwick, 1956). Whatever the exact number, the binding constraint was a manager's capacity to attend to people.

Agents dissolve that constraint. A human principal directing agent capacity is not limited by attention to subordinates in the way a manager was, since agents require no supervision in the relational sense, no development, no morale, and no arbitration among themselves that cannot be specified. If attention to people were the only limit, the agent-driven firm would have no reason to organize into units at all, and the correct structure would be a very small number of principals directing arbitrarily large machine capacity.

A different limit binds instead, and it is the residual identified in Section 3. A named human principal can only be answerable for so much, and the measure is not headcount but exposure: the financial commitment, regulatory obligation, safety consequence, and reputational liability that the principal will have to defend when something goes wrong. That capacity is bounded by what the principal can inspect, what they can explain afterward, what they can afford to be wrong about, and in regulated domains what their licence or certification permits them to stand behind.

We name this constraint span of accountable exposure, and the claim is that it succeeds span of control as the sizing rule for organizational units.

Several consequences follow directly. Unit size becomes uncorrelated with headcount, so a cell may be one person or forty depending on the exposure envelope rather than the workload. Units become heterogeneous in a way traditional org charts are not, since a cell handling reversible low-value decisions may run enormous volume under one principal while a cell touching regulated safety systems may be small and tightly bounded. The organizational chart begins to be drawn in units of consequence rather than units of reporting. And the executive question when forming a unit changes from what work does this team own to what exposure can this principal answer for, which is a materially different design conversation.

This is the paper's central contribution and the claim most worth attacking.

7. Where the Agent Firm Diverges From Haier

The topology that follows from span of accountable exposure resembles the micro-enterprise closely enough that the resemblance is worth taking seriously. Small units, a named principal bearing consequence, central functions operating as services rather than as authorities, and boundaries drawn by exposure rather than by function. If the argument holds, the agent-driven firm converges on a structure Haier reached by another route.

The mechanism cannot be carried across, and this is where importing RenDanHeYi as a template would fail.

Market discipline is ex post. A micro-enterprise learns it was wrong when its profit and loss says so, and that works because a human unit's mistakes accumulate at human speed, leaving an interval in which the signal arrives before the damage compounds beyond the unit's capacity to absorb it. A single agent decision inside an exposure-bounded cell can exceed that capacity before any market signal exists, and a cell that discovers its exposure was breached by observing its own quarterly result has discovered it too late to matter.

The exposure bound therefore has to be enforced rather than settled. It has to be evaluated in the execution path, before the action lands, against the cell's current standing and the exposure at stake, by an enforcement point that holds no write path and exists to refuse. That is the architecture developed in the Earned Authority Kernel, and this paper's contribution is to show what organizational form it implies rather than what technical form it takes.

The two arguments are complementary in a specific way. Haier supplies the topology, meaning small units with a named principal bearing consequence. Earned authority supplies the binding, meaning an exposure limit checked before action and sized by verified evidence rather than by market outcome. Neither is sufficient alone. A firm that adopts the topology without the binding has built cells that cannot hold their own exposure, and a firm that adopts the binding without the topology has built enforcement with no accountable principal behind the grant.

8. The Recentralization Counter-Hypothesis

The strongest objection to this paper's prediction runs in the opposite direction, and it deserves stating at full strength rather than dismissed.

Coase's boundary logic holds that a firm expands while internal coordination is cheaper than market transaction and contracts when the relation reverses (Coase, 1937). Agents reduce both costs simultaneously. Internal coordination becomes cheap because machine systems route, reconcile, and schedule at near-zero marginal cost, and external transaction becomes cheap because search, contracting, and monitoring across firm boundaries are themselves coordination problems that agents attack. The direction of the boundary is therefore not determined by the technology alone, and a plausible reading is that internal coordination falls faster, which would make hierarchy cheap again.

On that reading the agent-driven firm looks nothing like a network of cells. It looks like a very thin center directing enormous machine capacity, with no need for intermediate units because there is nothing for them to coordinate. Jensen and Meckling's analysis of decision rights holds that authority should be colocated with the specific knowledge required to exercise it, and that colocation can be achieved either by moving knowledge to the decision maker or by moving decision rights to the knowledge holder (Jensen and Meckling, 1992). Agents move knowledge, which favors centralization, and the entire history of enterprise software has been an argument for that direction.

The reason to expect the cellular outcome anyway is narrow and rests on one asymmetry. Coordination capacity now scales with compute and can be purchased. Accountability capacity does not scale with anything, cannot be purchased, and is bounded by the number of humans willing and able to answer for consequences of a given size. If exposure grows faster than the supply of principals who can hold it, then units must multiply even as coordination centralizes.

The prediction that follows is specific and testable: coordination recentralizes while accountability decentralizes, producing a thin coordinating center surrounded by many exposure-bounded cells. That is neither the traditional pyramid nor the Haier network, and if the argument is right it should be visible in organizations that adopt agents heavily while remaining answerable to regulators.

9. Four States of a Rule

The Convergent Enterprise thesis distinguishes a house rule, which the grantor may silently rewrite, from an externally witnessed clause, which cannot be amended in private because an outside party would notice, and adds a third state created by machine execution, the enforced constraint that no actor can bypass because the check sits in the execution path. The Haier case shows that the taxonomy was missing a state.

The house rule. Bound only by the grantor's continued willingness to constrain itself. This is the self-binding failure, and it is the default condition of most internal policy.

The externally witnessed clause. Bound because an outside party would learn that it changed. The audit opinion, the officer certification, the lender covenant, the regulatory filing. Scarce, expensive, and slow.

The market-exposed unit. Bound because the delegate directly bears the consequence of its own decisions, and that consequence is witnessed by parties outside the firm who transact or decline. This is Haier's mechanism, it requires no external governance apparatus, and it fails whenever the consequence arrives more slowly than the decisions that caused it.

The enforced constraint. Bound because a violation cannot become a consequence, since the check sits between the actor and the action and changing the rule is itself a governed act that leaves a record. This is the state machine execution makes newly available, and it is the only one of the four that operates at machine speed.

The four are not a maturity ladder and an institution will hold rules in all four states simultaneously. What matters is knowing which state a given rule occupies, because the failure mode of each is different and mistaking one for another is how governance becomes an intention.

10. What Would Prove This Wrong

The central prediction fails if organizations that adopt agents heavily do not develop exposure-bounded cells, and instead operate at genuinely large scale with a small executive group and no intermediate accountable units, while sustaining acceptable loss rates and defensible regulatory positions across several years. That would establish that accountability capacity scales further than this paper assumes, and that the residual identified in Section 3 is smaller than claimed.

It fails if accountability turns out to be transferable after all through instruments this paper does not anticipate. Insurance, bonding, capital reserves held against agent action, or a legal regime that assigns a form of standing to autonomous systems would each weaken the residual, and any of them would break the chain from accountability to unit size. The insurance route seems most plausible, and evidence that agent liability markets can price and absorb the exposure would materially damage the argument.

It fails if span of accountable exposure proves unmeasurable in practice. If firms cannot state the exposure envelope of a unit in terms specific enough to bound, then the sizing rule is a metaphor and the operative constraint remains headcount, budget, or reporting line by default.

It fails if the diffusion reading in Section 5 is wrong. If firms adopt RenDanHeYi-style structures at scale in the coming years without any shared-state substrate underneath them, then the obstacle was never coordination and the argument that AI supplies the missing enabler is a post hoc rationalization of a coincidence.

And it fails if the recentralization hypothesis simply wins. Evidence that heavily agent-driven firms are flattening toward a thin center without forming accountable cells, particularly in regulated domains where the exposure argument should bite hardest, would be direct disconfirmation rather than a matter of interpretation.

Acknowledgements

The line of argument that produced this paper began in public exchange with Brad Wolfe, whose convergence thesis and whose distinction between technical entitlement and institutional authority shaped the treatment of delegation throughout. Funsho A. supplied the self-binding formulation underneath the house rule, Tim Zlomke the establishment test that separates an auditable rule from an executable constraint, Andrii Gertner the observation that changing a governing rule invalidates authority already delegated under it, Ertan Aygordu the reformulation of the override record as a test of whether authority remains relevant, and Moohit Mehrotra the observation that segregation of duties applied to the policy layer already exists and is almost never used that way. The Haier material entered this argument through a reader's correction of the attribution of zero distance, which is the sort of correction that improves a paper by embarrassing it first.

A Note on AI-Assisted Research and Intelligence Provenance

Research and drafting for this paper were supported by AI-assisted tooling under the 451 Labs Writing Covenant. Astra, the adaptive strategic thesis reasoning agent, and Clearstory, the knowledge and intelligence provenance system that is the primary realization of the Astra thesis, are theses under active development rather than shipped products. All analytical judgments remain the author's own. Claims about Haier rest entirely on published secondary sources, and no primary access to the firm informed this analysis.

References

Coase, R. H. (1937). The nature of the firm. Economica, 4(16), 386 to 405.

Denning, S. (2022, January 30). Can firms succeed without managers? The case of Haier. Forbes.

Gartner. (2024). Predicts 2025: The future of work. [Prediction that approximately 20 percent of organizations will use AI to flatten organizational structure through 2026.]

Graicunas, V. A. (1933). Relationship in organization. In L. Gulick and L. Urwick (Eds.), Papers on the Science of Administration. Institute of Public Administration.

Hamel, G., and Zanini, M. (2018, November). The end of bureaucracy. Harvard Business Review.

Jensen, M. C., and Meckling, W. H. (1992). Specific and general knowledge, and organizational structure. In L. Werin and H. Wijkander (Eds.), Contract Economics. Blackwell.

Knowledge at Wharton. (2018, April 20). For Haier's Zhang Ruimin, success means creating the future. University of Pennsylvania.

Korn Ferry. (2025). Workforce 2025. [Survey of approximately 15,000 professionals reporting elimination of management layers.]

Live Data Technologies. (2024). Analysis of white-collar layoff composition, as reported in Bloomberg.

Mol, M. J., and colleagues. (2018). Management innovation made in China: Haier's Rendanheyi. California Management Review, 61(1).

Reddy, S. ADAM: A Governed Operating Machine for AI-Native Software and Multi-Agent Systems. SSRN Abstract 7245658, posted August 7, 2026.

Reddy, S. ADAM Earned Authority Kernel A Constitutional Execution Architecture for Governed Human and Machine Agency. SSRN Abstract 7268938, posted August 12, 2026.

Reddy, S. Earned Authority Agent Competency Intelligence and the Agent Earned Authority Score for Governed AI Agents. SSRN Abstract 7268001, posted August 12, 2026.

Thinkers50. (2023). Zhang Ruimin biography and Lifetime Achievement Award citation.

Urwick, L. F. (1956, May). The manager's span of control. Harvard Business Review.

Zohar, D. (2022). Zero Distance: Management in the Quantum Age. Palgrave Macmillan. [Treated here as advocacy rather than independent evaluation.]

Titles of SSRN papers are reproduced as registered. SSRN does not display DOIs for these papers; citation is by abstract number and posting date.