A board slide converts minutes saved by meeting summaries and draft preparation into a wider managerial span. It divides an estimated time saving by current supervisory hours, removes a layer and books the result. The arithmetic is neat. The structure is fiction. Summaries may save calendar time while managers still assign work, resolve conflict, inspect evidence, coach judgement, coordinate boundaries and recover failed outcomes. Agents may also produce more drafts and exceptions than the organisation can safely absorb.

Span is not a ratio of people to manager. It is the amount and pattern of coordination a manager can carry while decisions, people and outcomes remain healthy. Digital agents are not equivalent to direct reports. They may work at high volume without employment needs, yet add model monitoring, vendor dependency, permission, evidence review and incident obligations. A unit with fewer people can create more supervisory load.

Layers are also services, not empty boxes. They process information, allocate authority, integrate specialised knowledge, resolve exceptions, represent a boundary and develop people. Chandler's account of structure addressed the administration of expanding enterprises, while Galbraith connected uncertainty with information-processing demands.13 The relevant question after AI enters work is which coordination service changed, who now performs it, and whether the outcome improved.

01 / Reject the flattening shortcutMinutes saved do not prove structural capacity

The simple productivity rival deserves a fair test. If an agent reliably assembles routine evidence, tracks commitments and reduces rework, measured coordination load can fall. A manager may then supervise a broader unit without sacrificing development or control. A layer whose integration service becomes redundant may be moved or removed. Time saving can contribute to that result.

The rival fails when saved task time is absorbed elsewhere. Meeting notes arrive faster, but decisions still wait for authority. Draft volume rises, so evidence review grows. Routine questions fall, but rare exceptions become harder. The manager spends less time collecting status and more time investigating opaque outputs. Demand expands to consume the saved time. Nothing in this pattern creates structural capacity.

Simon described complexity through hierarchical and near-decomposable systems, providing a richer account than a simple preference for fewer levels.2 Williamson and coordination research likewise direct attention to the costs of governing dependencies rather than to the visual shape of an organisation alone.45 A structure can be flatter and still more expensive if coordination migrates into informal work, repeated negotiation and unresolved exceptions.

02 / Six load componentsThe coordination-load ledger

The commitment load covers assigning, sequencing and tracking work. Agents can lower it when commitments are visible and routine follow-up is automated. They can raise it when agents create many drafts, partial actions or states that someone must reconcile. Count work arrival, number of open commitments, waiting time and failed hand-offs.

The exception load covers novelty, ambiguity, conflict and escalation. Average exception frequency is insufficient. Severity, clustering and tail behaviour matter. A low-frequency agent failure that affects many customers can dominate a manager's capacity even when ordinary work is quiet.

The evidence load covers sources, output quality, risk and downstream effect. Visibility can reduce investigation, but new obligations can increase review. A persuasive summary does not lower load if the manager must reopen every source. The people load covers coaching, judgement development, motivation, wellbeing and conflict resolution. It does not disappear when task support improves.

The boundary load covers other teams, platforms, control functions, vendors and external parties. A local agent may reduce internal effort while adding platform change, security, assurance and vendor dependencies. The recovery load covers incidents, rework, remediation and return to stable operation. Every proposed span change must reserve capacity for it.

For each managerial unit, record the decision population, human reports, digital agents, work and exception arrival, service time, information quality, authority, review obligations, dependencies, development needs, interruptions, coordination delay and outcome. The ledger is not a single weighted score. Separate components preserve the reason a span is feasible or unsafe.

Build the baseline around accountable decisions

Start by defining the unit of managerial accountability. It may be a customer resolution, a release decision, an underwriting judgement or an operational recovery. A manager often oversees several decision populations with different consequences. Separate them. A broad span may be feasible for routine service commitments and unsafe for the smaller consequential population. Headcount obscures that distinction because the same employee can contribute to both.

Measure arrivals as a pattern, not a monthly total. Record ordinary rate, within-day peaks, seasonality, clustered events and the proportion that becomes an exception. Digital agents can shorten one service time while increasing arrivals by making it cheaper to create requests. They can also batch work into bursts when a scheduled process completes. The resulting interruption pattern can matter more than average volume.

Service time includes active handling and elapsed waiting. A manager may spend only a few minutes approving a case, but the case can occupy attention repeatedly while evidence arrives, questions are answered and another function responds. Record touches, reopenings and switches. Agents that make status visible can reduce these interruptions. Agents that send partial work or regenerate drafts can multiply them.

Information quality should be measured at the supervisory boundary. Ask whether the manager receives source-linked evidence, the current policy, uncertainty, downstream status and the reason for escalation. A shorter summary is useful only if it supports the required decision. If managers must reconstruct context from logs or ask a specialist to validate every claim, nominal visibility has not reduced evidence load.

Authority belongs in the baseline because waiting can arise from a mismatch rather than insufficient labour. Record which decisions the manager may settle, which require another role and which are delegated to teams or agents. Count cases that bounce between levels. A proposed layer change that leaves the same authority bottleneck cannot create capacity merely by moving reporting lines.

People load needs direct evidence. Record planned and delivered coaching, performance conversations, judgement calibration, conflict resolution, onboarding, absence support and signs of overload. Include unscheduled time when a manager helps a worker interpret an unfamiliar situation. These services are easy to erase from a capacity model because they do not appear as workflow events. Their absence emerges later through errors, attrition and weak escalation.

Boundary load is mapped as dependencies with service conditions. Name the platform, data, risk, legal, finance, supplier and external-party interactions required for the decision population. Record their response time and failure mode. Introducing an agent may lower internal coordination while creating a new dependency on a central model platform, identity service or vendor. That movement should appear in both units' ledgers.

Recovery load begins with the last accepted state. Record how incidents are detected, which decisions must stop, who investigates, who communicates, what can be replayed and who accepts return to service. Include ordinary rework, not only major incidents. A manager whose team handles frequent corrections is already supplying recovery capacity even if the business calls it customer service or quality support.

Run the same ledger after the agent is introduced without changing structure. Compare normal and peak periods, and distinguish temporary implementation work from persistent obligations. Some initial training load should fall. Model review, vendor coordination or new exception duties may remain. Structural capacity exists only when persistent total load and outcome risk fall while people service and recovery remain viable.

03 / Span feasibilityVolume, variety, consequence and observability interact

A span can widen when work is relatively homogeneous, commitments are visible, ordinary variance is resolved near the work, consequence is bounded and managers retain time for people and tails. It narrows when case variety, opaque evidence, approval concentration, external dependencies or consequence increase. Digital agents can move a unit in either direction.

Garicano modelled hierarchies as a way to organise knowledge and route problems that workers cannot solve.6 The insight matters for agents because routine resolution can free expert knowledge, but novel cases can also become more concentrated. If front-line automation removes easy learning opportunities and sends only difficult exceptions upward, expert service time and development needs can rise even while total case count falls.

Empirical research has documented changes in corporate hierarchy and delayering, including associations with competition and organisational roles.78 Those patterns do not establish that AI causes flattening. They provide useful structural variables and caution against treating hierarchy as fixed. The present decision still requires unit-level load and outcome evidence.

04 / Layer service contractA box can disappear only after its service has a home

Before moving or removing a layer, name the service it supplies. The service might integrate commercial and operational priorities, translate policy, allocate scarce expertise, resolve cross-team conflict, coach first-line managers or accept a class of risk. Name its consumers, evidence inputs, service time, authority and failure boundary. This is the layer service contract.

The replacement test asks who will supply every service after the change. An agent may provide visibility and routine coordination. A platform may standardise requests. A senior specialist may accept technical exceptions if granted authority and capacity. A higher executive may retain a consequential decision. If the answer is simply that teams will self-coordinate, the layer has not been replaced. Its cost has been hidden.

Authority must travel with work. Meagher and Wait connect decision delegation with information and change, while Jensen and Meckling distinguish costly-to-transfer knowledge from general knowledge.910 If specialists inherit coordination without decision rights, they become an informal layer that can recommend but not resolve. Delay and frustration rise.

05 / Structural failure modesWhere apparent capacity turns into queues

Output flood and approval manager

Agents increase proposal and document volume. Every exception routes to one manager because only that role can accept consequence. The manager becomes an approval server with human latency. Widening the span intensifies the queue. The repair is to reduce low-value output, clarify permission envelopes, route exceptions by type and retain specialist capacity where consequence clusters.

Hidden layer

A manager is removed, but senior specialists inherit prioritisation, conflict resolution and cross-team translation. Their job descriptions, authority and capacity do not change. Coordination becomes interruption-based and invisible in financial reporting. The organisation has not delayered. It has created an unpaid layer embedded in expert work.

People-service loss

Managers redirect capacity from coaching to AI output review. New workers receive less judgement development, difficult feedback is postponed and experienced staff absorb more exceptions. Near-term throughput may hold while future capability and wellbeing deteriorate. People load is not slack available for a technology business case. It is a layer service with delayed outcomes.

Calendar-saving fallacy and no-change boundary

Summaries reduce meeting administration but do not change exception, authority, coaching or recovery. Elsewhere, local task time falls but demand growth and new controls consume the saving. Dillon and colleagues found individual time changes without detecting a corresponding task-composition shift in their setting.12 The lesson is not that structural change never follows. It is that individual time evidence cannot establish it.

06 / Observation and tail loadDesign for the week that breaks the average

Average service time hides clustered incidents, regulatory deadlines, model changes and employee absence. Simulate representative peaks and tail events before widening a span. Estimate how many exceptions arrive together, which require the same authority, what evidence must be reopened and which boundaries join the incident. Reserve recovery capacity rather than assuming ordinary workload continues.

An observation window should include normal volume, peak demand, at least one controlled recovery exercise and enough people-development cycles to detect displacement. Record coordination delay and outcomes before and after. Do not widen provisionally and treat manager overtime as proof that the design works. Overtime is load escaping the ledger.

Authority and control independence constrain structure. The NIST AI RMF Core links roles, monitoring and incident responsibilities, while the IIA distinguishes management accountability and independent assurance contributions.1314 Removing a layer must not collapse challenge into the operating role or leave exceptions without a decision owner. Organisational capability, not technology access alone, determines whether adoption can be absorbed.15

Simulate the proposed structure before moving it

A useful simulation does not need a false precision model. Build a scenario table for ordinary flow, peak flow, a clustered exception, a platform outage, a consequential error and a staff absence. For each scenario, route the work through the proposed spans and layers. Record which role receives it, available authority, evidence required, service time, queue growth, downstream dependency and recovery path.

Vary agent performance and demand separately. Productivity may increase completed output and therefore create more downstream work even when error is unchanged. A model change may reduce ordinary exceptions but increase one rare class. A new control may require sampling. Keeping these variables separate shows whether a proposed span is robust or works only when every assumption moves favourably together.

Correlated arrivals deserve special attention. A policy change, data defect or model incident can send many similar cases to the same expert at once. Average exception rate will understate this load. Test how long the queue can grow before consequence becomes unacceptable, whether another authorised role can assist and which routine commitments should pause to release recovery capacity.

Downstream simulation follows the completed work. If agents accelerate proposals, can execution, fulfilment, quality, complaints and remediation absorb them? If not, the constraint has moved. A span decision should use the end-to-end outcome rate and rework, not the output of the local team. Otherwise the organisation widens capacity upstream while paying for delay elsewhere.

Set observation thresholds before the structural trial. Name acceptable queue age, interruption, evidence failure, coaching delivery, rework, incident recovery and outcome. Define which threshold pauses widening, which restores capacity and which invalidates a layer replacement. Pre-commitment protects the review from treating every negative signal as temporary after leaders have announced a flatter structure.

Use a reversible trial where possible. Keep reporting and employment changes separate from the first operating test. Reassign a bounded decision population, run the replacement service and observe receipts before removing the old layer. A short overlap may cost more, but it reveals missing work while the former service owner can still explain and repair it.

07 / Worked structureA service operation widens one span and keeps expert depth

A service unit receives high-volume routine requests and a smaller population of consequential exceptions. Before AI, first-line managers coordinate evidence assembly, status, allocation and many avoidable clarifications. An expert layer resolves policy ambiguity, customer vulnerability and recovery. The proposal introduces agents for verified evidence assembly and commitment tracking.

The coordination ledger shows lower commitment load because work state and due conditions are visible. Routine evidence load falls because sources and transformations are linked. Exception volume does not disappear, but its entry conditions become clearer. Boundary and recovery load remain stable. People-development time is protected as an explicit service condition. Downstream capacity absorbs the extra completed work.

One first-line span widens under a measured load ceiling. The expert exception layer stays because its integration, authority and coaching services are not replaced. Its queue is monitored by severity and tail age, not only count. Agent permission does not extend to consequential resolution. The observation window includes a peak period, an agent outage and a coaching cycle.

The no-change counter-case uses the same local time saving but fails the absorption test. Demand expands, review obligations rise and the manager remains the authority for every exception. People and recovery load increase. The unit keeps its structure, redesigns the exception path and observes again. No structural change is a positive decision when claimed savings are not coordination capacity.

Write the structural change clause

A structural decision should name more than a target span. The clause records the managerial unit, decision populations, current and proposed span, agent services, layer services that remain or move, load ceilings, authority, specialist capacity, people-service floor, observation window and reversal conditions. It also names an owner for each transferred service. This is the operational counterpart to the chart.

When a span widens, teams need to know which commitments they can resolve without managerial contact and which exceptions still require review. Permissions, policy and escalation must match. A manager cannot carry a wider span if every ambiguous case still seeks informal approval. The exception redesign is part of the structural change, not a later process improvement.

When a layer moves or disappears, consumers need a service notice. It should state where integration, coaching, conflict resolution and authority now sit, expected response time, evidence required and the route when the replacement fails. Specialists receiving new work need explicit capacity and decision rights. Executives retaining an escalation need to reserve time rather than assume rare cases are costless.

Finance should distinguish realised structural capacity from forecast labour avoidance. Track whether work was removed, automated, moved to another unit or absorbed as extra demand. Include central platform, control, vendor, specialist and recovery cost. A cheaper local hierarchy is not an organisational saving when another function supplies its missing coordination without funding.

Workforce effects also belong in the observation window. Monitor role clarity, interruption, manager accessibility, expert workload, coaching, progression and wellbeing. A wider span can support autonomy when teams have information and authority. It can create abandonment when the manager is distant but still controls every exception. These are different designs even if the ratio is identical.

Review the clause after a model, policy, demand, product, staffing or dependency change. A span that was safe under one exception mix may narrow when the agent handles more routine work and leaves a harder residual. A layer replacement may become unnecessary as a platform matures or newly necessary when regulation changes. Structure is maintained through evidence, not frozen after a transformation programme.

08 / Interactive structure testSpan-and-Layer Reconfiguration Test

The test returns one structural clause. It does not create a delayering score. It asks how all six loads change, whether tails and downstream demand are absorbed, whether people service survives, and whether any removed layer has a named replacement owner.

How the negative controls route

Meeting summaries with unchanged load produce no structural change. Output flood narrows the span or adds specialist capacity. A manager acting as the only approval path keeps the span and redesigns exceptions. A hidden layer blocks removal until work, authority and capacity are named. Lost coaching blocks span expansion. Consequence-heavy tails narrow the span. A fully replaced integration service can support moving or removing its layer after observation.

09 / Structural ruleChange the load path, then change the chart

AI does not carry a universal organisational shape inside it. It changes work arrival, visibility, variance, evidence, authority and recovery. Those changes can make one unit broader and another narrower. They can remove a layer service, create a new specialist need or leave the current design untouched.

Start with the accountable decision population. Record all six coordination loads before and after AI. Price tails, interruptions and downstream absorption. Protect coaching and judgement development. Name every layer service and its consumers. Move authority with integration work. Observe peaks and recovery, not only an average week.

Then make the smallest structural change supported by evidence. Widen a span under a load ceiling, repair an exception path, add consequence-bearing capacity, replace a layer service or hold the chart steady. The proof is lower coordination load per accountable decision with preserved outcomes and people capability. Fewer boxes are not the result. A structure that carries its real load is.

The ledger should remain live after the change. Managers and teams can record a new interruption source, control duty, boundary dependency or recovery burden as it appears. Periodic review should compare those observations with the assumptions in the structural clause and distinguish a temporary transition from a persistent load shift. This prevents small pieces of invisible coordination from accumulating until the span fails suddenly. It also makes later simplification safer, because leaders can see which services have genuinely become routine, which still depend on scarce expertise and which have moved beyond the unit. A maintained load path turns organisation design from a one-time restructuring exercise into evidence-based operational stewardship. That is structural discipline.

SourcesReferences

1. Galbraith, J. R. (1974). “Organization Design: An Information Processing View.” Interfaces, 4(3), 28–36. DOI.
2. Simon, H. A. (1962). “The Architecture of Complexity.” Proceedings of the American Philosophical Society, 106(6), 467–482. DOI.
3. Chandler, A. D. (1962). Strategy and Structure. MIT Press. Publisher record.
4. Williamson, O. E. (1975). Markets and Hierarchies. Free Press. Author-index record.
5. Malone, T. W. and Crowston, K. (1994). “The Interdisciplinary Study of Coordination.” ACM Computing Surveys, 26(1), 87–119. DOI.
6. Garicano, L. (2000). “Hierarchies and the Organization of Knowledge in Production.” Journal of Political Economy, 108(5), 874–904. Verified publisher DOI.
7. Rajan, R. G. and Wulf, J. (2006). “The Flattening Firm.” Review of Economics and Statistics, 88(4), 759–773. DOI.
8. Guadalupe, M. and Wulf, J. (2010). “The Flattening Firm and Product Market Competition.” American Economic Journal: Applied Economics, 2(4), 105–127. DOI.
9. Meagher, K. J. and Wait, A. (2014). “Delegation of Decisions About Change in Organizations: The Roles of Competition, Trade, Uncertainty, and Scale.” Journal of Law, Economics, and Organization, 30(4), 709–733. Verified publisher DOI.
10. Jensen, M. C. and Meckling, W. H. (1995). “Specific and General Knowledge, and Organizational Structure.” Journal of Applied Corporate Finance, 8(2), 4–18. Verified publisher DOI.
11. Brynjolfsson, E., Li, D. and Raymond, L. R. (2025). “Generative AI at Work.” Quarterly Journal of Economics, 140(2), 889–942. DOI.
12. Dillon, E. W., Jaffe, S., Immorlica, N. and Stanton, C. T. (2025). “Shifting Work Patterns with Generative AI.” NBER Working Paper 33795. DOI.
13. NIST. AI Risk Management Framework Core. Official resource.
14. The Institute of Internal Auditors (2026). Statement of Position on the Three Lines Model. Official resource.
15. OECD, BCG and INSEAD (2025). The Adoption of Artificial Intelligence in Firms. Official publication.