Agents Arrive in Thinner Structures

Large employers have spent two years removing layers of management, and they are now deploying AI agents into the structures that remain. No delayering business case has yet had to answer the question that follows: once the agents are running, who still approves what they do, and how much of it? Korn Ferry’s 2026 workforce survey reports that 42% of organizations cut management roles over the past year, and that more than half of the managers who remain say they are exhausted. In May, Cloudflare’s chief executive wrote in The Wall Street Journal that the vast majority of the people his company had just laid off were “measurers,” a group that, in Fortune’s account, spans middle management, finance, legal, internal auditing, and revenue recognition. It is an unusually candid description of a cut, since the people it took out were the ones whose job was to check.

Some of those layers should go, and the familiar warning about the rest is already on record: Deloitte’s guidance on delayering warns of “shadow approvals and undocumented escalation paths” when the work managers did reappears informally. A narrower and more mechanical effect arrives first. A firm that removes a layer and puts agents into the thinner structure squeezes the human review it can afford from three directions at once, and what gives way is the share of agent actions that ever reaches a person. That share falls by increments that each look like housekeeping, a raised approval threshold here and a thinner sample there, while the reviews that remain keep their slot on the calendar and lose their bite.

What a Span Ratio Prices

Delayering cases rest on spans and payroll. Amazon set out in 2024 to raise its ratio of individual contributors to managers by at least 15%, and Gartner’s prediction that a fifth of organizations would use AI to remove more than half of their middle-management positions led with reduced labor costs.

Part of what that payroll math removes deserves to go. A layer that spent its week collecting status and carrying information up and down the line is doing what telemetry now does better, and Deloitte’s guidance is right that “telemetry replaces manual reporting.” Studies of what managers actually do, from Mintzberg’s observation of chief executives to Hales’s later review of the evidence across levels, found much of the week going elsewhere: allocating resources, handling disturbances, and keeping lateral contact with peers. In an operation that runs through an approval matrix, the written list of who may approve what and up to which amount, that work takes concrete forms. A management level holds approval limits in that matrix and owns the thresholds and sampling rates that the analysts below it work within. In many operations its team leads also clear escalations themselves. Its managers settle exceptions that cross team boundaries, prevent others through the contact they keep with their peers, and develop the people whose expertise every future review depends on.

A span ratio prices what is legible about management, which is how many people report to whom, and it has no denominator for the integrative part, which is who holds judgment when an exception arrives that no rule anticipated. The guidance itself describes what remains after a delayering as “the work that resists automation.” The business case never measured that work, so it cannot report what the cut did to it.

The Arithmetic of Human Review

An earlier article on agent decision rights, The Approval Chain Was Built for Humans, set out the constraint every escalation design has to satisfy. Stated in hours and extended to cover samples, it says that the items routed to people, multiplied by the minutes a credible review takes, cannot exceed the review hours of the people accountable for them. A design that breaks the constraint keeps producing approvals, at a pace that rules out judgment, which is the condition that article called review theater. The constraint is also a ceiling, and a queue run close to it stalls well before reaching it.

Read across an org chart after a delayering, the constraint describes a squeeze with two parts. The first is capacity. Where team leads cleared escalations, removing their layer removes reviewers outright, and the managers and analysts who remain absorb the exception follow-through and coordination the layer carried, so each has less time to review; Gallup already puts the median manager at 40% of working time spent as an individual contributor. Agents meanwhile multiply the actions that could need a person, and exceptions climb with volume even when the rate at which agents produce them holds steady. The second part is ownership. The removed level owned the thresholds that decide how much reaches anyone at all, and its departure leaves those thresholds without an owner at the moment the volume behind them rises.

Galbraith described this pressure half a century ago. When exceptions arrive faster than a hierarchy can process them, he argued, the organization has to reduce the flow or add capacity, and if it does not consciously choose, lowered performance standards arrive on their own: “Not to decide is to decide.” A delayering paired with an agent rollout raises the flow and removes capacity in the same quarter, which leaves that default to do the choosing.

Thresholds Rise Through Routine Changes

The drop comes in increments, each reasonable on the day it happens. Queues lengthen once the review hours available no longer cover the items arriving, and someone raises the auto-approve limit to clear them. The change goes through as a configuration update, and in operational queues outside financial-reporting controls, such as know-your-customer refreshes, the team that owns the backlog can often approve it. The escalations that still come through then pull analysts off the quality sample, which shrinks. The Approval Chain warned that a threshold with no owner gets “widened by whoever is closest to a deadline,” and delayering removes the owner. Each increment has a change ticket and someone who approved it; the total has neither, and nobody has the job of reading what the tickets add up to.

Approvals that the removed level signed within its limit move up the approval matrix to fewer and more senior people, a pattern resembling what Wulf found when large firms flattened their top ranks, where delayering meant to push decisions down “in the end looks more like centralization.” Rasmussen’s account of how work systems drift fits the whole sequence. Under steady pressure toward efficiency and toward least effort, he argued, practice migrates toward the boundary of acceptable performance, through steps that look rational to people who cannot see the whole picture. Threshold creep is that migration applied to approval limits.

The arithmetic of a single queue shows how far the drift has to go, using illustrative figures. An exceptions queue in a payments operation runs agents on 40,000 items a week, routes 5% of them to a person, and samples 2% of what clears automatically. At six minutes for a credible review of an escalation, a conservative figure for payments work, and three for a sampled item, the queue needs 238 review hours a week. Twelve reviewers, three of them working team leads, give 25 hours each and supply 300, so the queue starts at about 80% of capacity. Then the team-lead layer goes. The nine remaining analysts absorb the follow-through the leads carried and keep 20 review hours each, the agents’ scope widens to 60,000 items, and the queue needs 357 hours against 180. It still clears, either when the escalation share falls from 5% to about 2% or when each escalation gets about two and a half minutes instead of six. With the agents’ scope held at 40,000, the share still has to fall to about 3.5%. Restoring the original standard takes around 22 reviewers at the utilization the queue started with, a cost that belongs in the same business case that counted three team-lead salaries.

The share may deserve to fall. A low share set deliberately, against risk tiers that someone owns and defect data that supports the line, is sound design, and The Approval Chain argued for exactly that. In this queue the share falls because the hours ran out.

Faster Approvals on the Delayering Scorecard

The scorecards that judge delayering read most of this as progress, at least at first. Deloitte’s guidance, among the more careful treatments of the subject, measures success partly through decision latency, defined as “less time from issue raised to decision” and “less waiting for approvals,” and it pairs that measure with quality signals: “No spike in defects/reversals” and “Monitoring catches drift early.” The pairing is right, and it still misses this failure. Defects surface in what someone samples or what a customer reports, so a thinner sample keeps the quality row green for the same reason that a raised threshold improves the latency row, and the shortfall shows up only when the defects arrive by another route. A program office reading that dashboard a year after the cut sees the improvement the business case promised.

The two routes differ in kind. Where thresholds rise, cases leave human review entirely, the same selection an earlier article on decision migration traced in governance forums, where contested cases settle upstream, in conversations that leave no record, before the forum built to examine them ever sees them. Where review minutes shrink, the cases stay and people still approve them, on the record and faster, which is review theater produced by the org chart itself.

What the dashboard leaves out is whether anyone still says no. Anthropic reports that users of its own coding agent approve 93% of permission prompts, and it describes the result as approval fatigue, “where people stop paying close attention to what they’re approving,” which is the vendor’s own evidence that attention fades with volume. Singapore’s framework for agentic AI governance warns that requiring people to oversee agents continuously “can also lead to automation bias and alert fatigue,” and lists among its measures how often people override or reject agent actions and how long they take to review them. A falling override rate and shrinking review times can also mean the agents improved or the tooling got better, so neither number settles anything alone. Read against a log of every change to an approval threshold or sample rate, they separate a faster decision from an unexamined one, and neither they nor the log appear among Deloitte’s delayering measures.

Control Coverage Under Threshold Sign-Off

Risk and controls teams will object that the checker stays in place, and in Deloitte’s worked example it does. Its model of a monthly close fuses two preparer roles, the accountant and the analyst, into a single “performance narrative owner,” keeps the manager as the reviewer who “handles exceptions,” and turns separation of duties from something informal into something explicit, with “threshold-based sign-off” and an audit trail. On paper the control environment improves.

What changes is coverage. Under threshold-based sign-off, items below the line get no mandatory review, and whatever still covers them, a sample or a downstream reconciliation, needs its own design and its own staff. Deloitte assigns “governing uncertainty thresholds” to managers, the layer delayering thins. PCAOB staff guidance on auditing management review controls ties their precision to the threshold for investigating deviations, so a review keeps running while its precision falls, and when the person who moves the line is the one whose backlog it relieves, separation of duties survives on the org chart and fails in operation.

Agents add a less visible fusion. Deloitte’s principle is “fused build, separated approve,” and with agents the build includes the instructions and the thresholds, so a person who configures an agent and then approves its output sets the scope of the check and performs it, which auditors call self-review. In a US public company the chain ends at the Section 302 certifications the chief executive and chief financial officer sign each quarter, commonly backed by sub-certifications from the layers below. Unless those sub-certifications move to the people who now hold the thresholds, delayering removes signers while the officers go on signing the same statement.

The rules written for agents set a floor without a size. The EU AI Act requires deployers of high-risk systems to “assign human oversight to natural persons who have the necessary competence, training and authority, as well as the necessary support,” obligations now deferred to December 2027 and limited to uses most back-office agents will not reach, and the model-risk guidance US bank regulators revised in April 2026 leaves generative and agentic models outside its scope. The older anti-money-laundering expectation that a program have adequate resources does test staffing, and TD Bank shows how far a budget rule can override it.

The Justice Department’s 2024 resolution with the bank describes senior executives enforcing a budget mandate known internally as a “flat cost paradigm,” which held the budget flat year over year while the bank postponed and cancelled necessary anti-money-laundering projects. Its monitoring program “remained effectively static” from 2014 through 2022, and 92% of transaction volume, about $18.3 trillion, went unmonitored between 2018 and 2024. The plea record calls the monitoring gaps intentional, and the case involved far more than capacity. Its narrow lesson for delayering is the condition that made those gaps cheap to choose, a budget held flat while volume and risk grew, and agents shorten the time such a gap takes to open, since a threshold can move in an afternoon.

Sizing Human Review Before the Cut

None of this is an argument for keeping managers. Layers that mainly relay information should go, and McKinsey’s survey of organizations concludes that structural redesigns, cost cuts, and flatter hierarchies are “achieving diminishing returns.” Gartner now advises HR leaders to assess the density and autonomy of agents in each team before changing manager headcount, its “span of complexity,” and the capacity check that compliance operations already run for alert queues turns that instinct into a number. Pooled approvers or monitoring agents feeding staffed queues can supply the capacity, and in some units an honest sizing will show that more layers can go. The case is for sizing review before the cut, while the shortfall can still change the plan.

The sizing starts with an inventory taken one decision type at a time, in which everything the level held goes somewhere explicit. Relay and reporting go to automation, approval limits, threshold ownership, and exception judgment go to a named role with the authority to exercise them, and human review becomes a capacity the firm sizes against agent volume after risk tiering. Development moves into a funded program, since reviewers can judge agent output only while they keep the expertise to do the work, and Bainbridge warned more than four decades ago that automated systems watched by former manual operators ride on skills that later generations will not have. In practice the inventory is one table per unit, a row per decision type showing volume, escalation share, sample rate, review minutes, the accountable role, the hours that role keeps, and the gap. The operating executive owns it with the head of HR, and the risk function challenges it before any announcement.

After the cut, a change to an auto-approve threshold or a sample rate becomes a controlled change, logged with its reason and approved by someone other than the owner of the queue it relieves, and changes beyond a set coverage tolerance go to the risk function. Agents get their own lines in the approval matrix and action logs complete enough to recompute the override rate, escalations go to a queue with a service level and one accountable owner, and someone holds standing authority to pause the agents.

The first year’s reporting shows whether the sizing held. The threshold log and the actual sample rate show whether coverage moved, and the defect rate among sampled auto-approved items, which anti-money-laundering teams already run as below-the-line testing, shows what the movement cost. Override rates and review times show whether people still examine what reaches them, and the share of approvals signed two or more levels above the removed layer shows whether its decisions moved up. The operating owner reports those figures, the risk function challenges them, and internal audit tests at six and twelve months whether the sizing held. A board can ask for all of it in one question when a delayering plan arrives: once the agents are running, how many consequential approvals and exceptions will each remaining accountable person handle in a week, and what is the basis for that number? An adequate answer names the risk tiers, the thresholds, the sample rate, the review minutes assumed, and who may change any of them. A plan that cannot answer has not removed a cost. It has moved review out of the budget and into the thresholds, where a latency measure with no coverage measure beside it will read every lost review as a faster decision.