Machine Bureaucracy

Abstract

This paper presents an agent-based simulation of a non-robotic e-commerce fulfilment centre, isolating the operational interface between the stow (inbound placement) and pick (outbound retrieval) stages. The simulation models how individual hourly throughput targets interact with spatial constraints, worker fatigue and bin capacity to produce systemic rule-breaking. Drawing on Goodhart’s Law and Campbell’s Law, the paper argues that non-compliant placement behaviour is not a disciplinary failure but a structural outcome of the measurement framework: when operators are evaluated on isolated speed metrics, rational adaptation dictates optimising for the local target while externalising friction to downstream processes. Four operational scenarios are parameterised, ranging from orderly flow to systemic collapse, demonstrating a compounding feedback loop in which stow-phase shortcuts degrade bin hygiene and directly increase pick-phase latency. The analysis suggests that addressing these failure modes requires shifting from individual surveillance to integrated value-stream measurement, and from enforcing compliance to redesigning the conditions under which compliance is rational.

1. Introduction

The term machine bureaucracy, introduced by Mintzberg (1979) ^[1], describes an organisational structure designed to maximise predictability through standardised work processes. In Mintzberg’s framework, these organisations depend on formal rules, repetitive tasks and a dominant technostructure whose role is to measure and standardise operational output. The model assumes that efficiency follows from treating work as decomposable and workers as interchangeable. That assumption holds under stable conditions. It breaks down when the metrics used to enforce standardisation begin to reshape the behaviour they were designed to measure.

This paper uses an agent-based simulation to examine that breakdown in a specific operational setting: a non-robotic fulfilment centre. The simulation models how strict individual performance metrics structurally alter worker behaviour at the interface between inbound stow and outbound pick processes. Rather than functioning as the predictable mechanism Mintzberg described, the warehouse operates as a complex adaptive system (Holland, 1995 ^[2]; Choi et al., 2001 ^[3]), where worker fatigue, spatial constraints and target pressures interact dynamically, producing emergent behavioural patterns that no single variable predicts in isolation.

The central theoretical lens is Goodhart’s Law. The original formulation, drawn from monetary policy, states that ‘any observed statistical regularity will tend to collapse once pressure is placed upon it for control purposes’ (Goodhart, 1975 ^[4]). The generalised version, attributed to Strathern (1997 ^[5]), is more widely cited:

“Any observed statistical regularity will tend to collapse once pressure is placed upon it for control purposes.”

A parallel formulation from social science, Campbell’s Law, makes the institutional consequences explicit: ‘the more any quantitative social indicator is used for social decision-making, the more subject it will be to corruption pressures and the more apt it will be to distort and corrupt the social processes it is intended to monitor’ (Campbell, 1979 ^[6]). Together, these two principles predict that when operators are evaluated on raw throughput (units processed per hour), achieving those quotas under spatial or temporal constraints will produce systematic circumvention of placement rules. The resulting degradation in bin hygiene is not indicative of individual failure but of a structural tension within the measurement framework itself (Mannion & Braithwaite, 2021 ^[7]; Kohn, 2018 ^[8]).

1.1 Process Mapping (SIPOC)

To contextualise these behavioural adaptations, the operational pipeline can be mapped using a SIPOC (Suppliers, Inputs, Process, Outputs, Customers) framework (Sunder M., 2016 ^[9]). Table 1 presents the end-to-end process for a manual e-commerce fulfilment centre.

This simulation isolates the boundary between Static Stow and Item Picking. In manual operations, stowers and pickers use the same physical storage bins at different times within the shift cycle. Both roles carry hourly rate targets. When spatial constraints tighten, stowers face pressure to maintain throughput by violating placement protocols: overfilling bins, ignoring dimensional constraints, or bypassing item-count verification. This transfers operational friction downstream. Pickers subsequently encounter disorganised bins, resulting in increased retrieval times and elevated error rates. The efficiency gained during the stow phase manifests as a direct latency penalty during the pick phase.

2. Literature Review

2.1 Performance Measurement and Perverse Incentives

The unintended consequences of target-based performance management have been documented across disciplines. Goodhart’s original observation (1975 ^[4]) concerned the collapse of monetary aggregates as policy instruments once central banks began targeting them. Strathern (1997 ^[5]) generalised the principle to audit culture in British universities, coining the phrasing most commonly cited today. Mannion and Braithwaite (2021 ^[7]) extended the framework into healthcare, documenting how clinical performance indicators distort the very processes they were designed to monitor. Treem et al. (2023 ^[10]) examined how digitisation accelerates these distortions by increasing the volume and granularity of behavioural data available for targeting.

Campbell’s Law (1979 ^[6]) operates from the same premise but emphasises institutional corruption: when quantitative indicators carry high stakes, the pressure to optimise the indicator eventually degrades the process it was meant to measure. Kohn (2018 ^[8]) argues more broadly that extrinsic reward systems produce predictable dysfunctions across educational, organisational and workplace settings, with compliance incentives crowding out intrinsic motivation.

The relevance to warehouse operations is direct. Fulfilment centres typically evaluate workers on a small number of throughput metrics (units per hour, scan rates), creating the conditions that both Goodhart and Campbell describe.

2.2 Warehouse Operations and Algorithmic Management

Recent scholarship has examined the labour conditions of large-scale fulfilment operations, particularly those driven by algorithmic management systems. Delfanti (2021 ^[11]) describes a regime of ‘augmented despotism’ in Amazon warehouses, where real-time tracking systems constrain worker autonomy and intensify output expectations beyond what the physical environment can sustain. Cheon and Erickson (2025 ^[12]) document the ‘work games’ warehouse employees develop to cope with algorithmic rate targets, showing how workers improvise around system constraints in ways that are individually rational but collectively costly.

De Koster et al. (2007 ^[13]) provide an operational baseline for warehouse process design, cataloguing the trade-offs between order-picking strategies, storage assignment policies and routing heuristics. Their review establishes that warehouse throughput depends on spatial layout and process sequencing, not solely on individual worker speed. Head (2014 ^[14]) situates contemporary warehouse management within the longer trajectory of ‘digital Taylorism’, arguing that modern tracking technologies extend rather than replace the logic of scientific management.

2.3 Complex Adaptive Systems in Operations

The CAS framework, as developed by Holland (1995 ^[2]), characterises systems in which many interacting agents produce emergent, non-linear behaviour that cannot be predicted from the properties of individual components. Choi et al. (2001 ^[3]) applied this framework to supply chain networks, demonstrating that the tension between managerial control and emergent adaptation is a defining feature of operational systems. In this framing, a fulfilment centre is not a machine executing a fixed programme but a network of agents (workers, bins, carts, scanners) whose interactions produce system-level patterns that no single agent controls.

Deming’s System of Profound Knowledge (1993 ^[15]) offers a complementary perspective. His distinction between common-cause and special-cause variation implies that the majority of performance problems are attributable to the system, not to individual workers (Deming, 2000 ^[16]). This principle underpins the argument developed in Section 5 and the conclusion of this paper: that the analytical focus should be on system architecture, not individual compliance.

3. Simulation Architecture

The simulation environment integrates several discrete modules to model the propagation of systemic pressure across the facility floor:

• Control Parameters: Independent variables include stow and pick target rates.
• Spatial Canvas: The simulation abstracts the workflow into a continuous cycle across four stages (Buffer → Stow → Bin → Pick). The stow cycle models inbound placement, while the pick cycle models outbound retrieval.
• Complexity Heatmap: A real-time spatial grid displays bin capacity saturation and messiness.
• Systemic Metrics: The dashboard tracks throughput (stow versus pick rates), rule adherence distribution, physical slot availability and aggregate worker fatigue. As target pressures increase, the data show a quantifiable shift from compliant placements to shortcut behaviours, which subsequently depresses the downstream pick rate.

3.1 Methodology and Design Rationale

The simulation is an agent-based model (ABM) (Bonabeau, 2002 ^[17]). ABM is appropriate here because the phenomenon under investigation, the emergent shift from compliance to shortcutting, arises from the interaction of individual agents operating under local rules, not from any single variable in isolation.

The simulation’s deterministic engine computes throughput from a mathematical model that is independent of stochastic fatigue or spatial penalties unless these are explicitly enabled. Parameter ranges (target rates, bin capacities, compliance coefficients) were calibrated for face validity against published descriptions of fulfilment centre operations (Amazon, 2019 ^[18]; Amazon Technologies, 2013 ^[19]). The model is a demonstration, not an empirical study: it is designed to make the feedback loop between target pressure and rule-breaking visible and manipulable, not to replicate the output of a specific facility.

4. Operational Mechanics and Rate Drivers

4.1 Rule Adherence Mechanics

• Velocity Penalty: A fully compliant stower (expected compliance rate = 1.0) incurs a 20% duration penalty (speed multiplier = 1.2) to account for alignment, verification and spatial distribution of items. A non-compliant stower operates at the baseline velocity (speed multiplier = 1.0).
• Spatial Distribution: Compliant stowers restrict placements to a maximum of 4 items per bin before relocating. Non-compliant stowers bypass this constraint, depositing entire cart volumes into single bins.
• Theoretical Maximum Throughput:
  • Non-Compliant: Up to 210 units/hour per operator (17.14s baseline).
  • Compliant: Up to 175 units/hour per operator (210 / 1.2).
• Compliance Probability: Rule adherence is calculated dynamically as a function of several weighted inputs. The dominant terms are Stow Target Pressure (negative contribution) and Storage Complexity (negative contribution). Individual Operator Bias (±0.25) introduces heterogeneity across the workforce. The full model includes additional terms for friction pressure, picker target bias and slot scarcity; the coefficients presented here represent the primary drivers.

4.2 Target Pressure and Systemic Constraints

Stow Target Pressure quantifies the operational tension that compels operators to sacrifice rule adherence for raw throughput.

Pressure is calculated as a weighted sum bounded between [0.04, 0.98]. For clarity, the equation below presents the dominant terms; the full model includes additional inputs for non-neat bin share, counting difficulty and friction pressure:

As pressure increases, operators transition from compliant stows (1.2x duration) to shortcut stows (1.0x duration). The relationship is not binary: the compliance probability decreases continuously as pressure rises, producing a gradual shift in the distribution of behaviours across the workforce.

4.3 Bin Saturation and Complexity

Both spatial saturation (fill ratio) and structural complexity (messiness) directly affect throughput:

• Saturation Penalties: As a bin reaches physical capacity, the engine applies a non-linear penalty: $\text{fillPenalty} = (\text{fillRatio})^{1.3} \times 0.25$. This subtracts up to 0.25 from the rule-following probability, forcing a decision between missing targets or non-compliant placement.
• Complexity Penalties: Bin messiness subtracts up to 0.18 ($-\text{messiness} \times 0.18$) from the compliance probability. Compliant placements into messy bins also require additional physical time for reorganisation ($+0.18\text{s} + (\text{messiness} \times 0.32\text{s})$) and recounting.

5. Systemic Degradation and Feedback Loops

The operational shift follows a compounding feedback loop:

1. Target Establishment: Hourly throughput quotas are fixed. Under conditions of high targets, operators experience immediate pressure to meet the number above all else.
2. Spatial Constraint Escalation: As physical bins reach capacity, compliant placement requires increased search and manipulation time (Delfanti, 2021 ^[11]).
3. Behavioural Adaptation: Faced with the immediate penalty of failing a quota versus the delayed, externalised cost of a disorganised bin, operators select non-compliant shortcuts. This choice is locally rational: the penalty for a missed target is personal and immediate, while the cost of a messy bin is deferred and lands on another worker (Cheon & Erickson, 2025 ^[12]; Kohn, 2018 ^[8]).
4. Downstream Latency: Shortcut stows generate structural messiness. Pickers subsequently encounter elevated search times and error rates. The velocity gained during stow translates directly into latency during pick.
5. Systemic Failure: If problem-resolution capacity is exceeded, defect rates compound and the overall operational flow degrades.

6. Scenario Analysis

The simulation parameterises four distinct operational states:

• Orderly Flow: Bins have sufficient available capacity. Operators experience minimal conflict between accuracy and velocity, permitting target achievement without rule violation.
• Emergent Congestion: Bin capacity approaches functional limits. Search times increase, forcing operators to negotiate the trade-off between throughput and compliance with increasing frequency.
• High-Pressure Degradation: Elevated target pressure combined with high spatial saturation drives structural disorganisation. Operators prioritise immediate throughput over bin hygiene, producing a rapid proliferation of shortcut behaviours (Cheon & Erickson, 2025 ^[12]).
• Systemic Collapse: Storage capacity is severely restricted. Every placement inherits the structural debt of previous shortcuts, creating compounding downstream latency. Psychological safety deteriorates as operators recognise systemic failure but suppress reporting due to surveillance pressures (Edmondson, 1999 ^[20]; Edmondson & Lei, 2014 ^[21]).

7. Core Modelling Equations

The simulation relies on the following simplified proportionalities to model floor pressure and compliance. These represent the dominant relationships in the model; the full implementation includes additional terms and interaction effects (Mannion & Braithwaite, 2021 ^[7]; Treem et al., 2023 ^[10]).

1. Storage Complexity

Complexity quantifies the friction of locating compliant placement slots. It scales non-linearly with scarcity and messiness:

2. Rule Adherence Probability

The probability that an operator follows placement protocols, bounded between 0.05 and 0.98:

3. Process Efficiency

Efficiency calculates the proportion of value-adding operational work versus non-value-adding rework generated by shortcuts and search latency:

8. Limitations and Future Work

Several limitations constrain the scope of this analysis.

Abstraction trade-offs. The simulation simplifies spatial layout to a grid of uniform bins, omits worker heterogeneity beyond a single bias term, and does not model social dynamics between workers (peer pressure, informal communication, collective resistance). Real fulfilment centres contain aisle geometry, variable bin sizes, elevation changes and walking distances that affect both fatigue and compliance decisions in ways the current model does not capture.

No empirical validation. The model has not been calibrated against production data from an operating fulfilment centre. Parameter values were set for face validity against published operational descriptions, not fitted to observed throughput or error-rate data. The simulation demonstrates a plausible mechanism, not a measured effect size.

Scope. The paper isolates the stow-pick interface. In practice, degradation cascades through all seven stages of the fulfilment pipeline (Table 1). Upstream bottlenecks at decant, or downstream congestion at rebin and pack, introduce additional feedback loops not represented here.

Future work. Three extensions would strengthen the contribution. First, field validation through partnership with an operating facility would allow calibration against real throughput and error-rate distributions. Second, extending the model to automated or semi-automated fulfilment centres (where robotic stow and pick introduce different failure modes) would test whether the feedback loop holds under reduced human agency. Third, integrating real warehouse management system (WMS) data would permit the model to operate as a diagnostic tool rather than a standalone demonstration.

9. Conclusion

The dynamics demonstrated in this simulation are not confined to physical warehousing. They appear wherever isolated velocity metrics generate unmeasured downstream rework: in software development pipelines where deployment speed creates production incidents, in healthcare settings where throughput targets compromise diagnostic accuracy, and in customer service operations where call-handling time metrics reward premature resolution.

When distinct operational teams are evaluated on isolated throughput metrics, rational adaptation dictates optimising for the local metric while externalising friction (Mannion & Braithwaite, 2021 ^[7]; Campbell, 1979 ^[6]). High psychological safety permits operators to identify systemic bottlenecks; conversely, surveillance-driven fear ensures defects compound silently (Edmondson, 1999 ^[20]; Edmondson & Lei, 2014 ^[21]).

Deming (2000 ^[16]) estimated that approximately 94% of performance problems are attributable to the system rather than to the individual worker. Whether or not the precise figure holds, the principle applies here: the simulation shows that non-compliance is not a deficiency in the workforce but a predictable output of the system’s design. Addressing the failure modes documented in this paper requires a shift from individual surveillance to integrated value-stream measurement (Kaplan & Norton, 1996 ^[22]; Qian et al., 2019 ^[23]). The question for operational leaders is not ‘how do we force workers to follow rules?’ but ‘why does our system make rule-following irrational?’ (Jones, 1991 ^[24]).

10. References

• [1] H. Mintzberg, The Structuring of Organizations: A Synthesis of the Research. Englewood Cliffs, NJ: Prentice-Hall, 1979.

• [2] J. H. Holland, Hidden Order: How Adaptation Builds Complexity. Reading, MA: Addison-Wesley, 1995.

• [3] T. Y. Choi, K. J. Dooley, and M. Rungtusanatham, “Supply networks and complex adaptive systems: Control versus emergence,” Journal of Operations Management, vol. 19, no. 3, pp. 351–366, 2001.

• [4] C. A. E. Goodhart, “Monetary relationships: A view from Threadneedle Street,” in Papers in Monetary Economics, vol. 1. Sydney, Australia: Reserve Bank of Australia, 1975.

• [5] M. Strathern, “‘Improving ratings’: Audit in the British university system,” European Review, vol. 5, no. 3, pp. 305–321, 1997.

• [6] D. T. Campbell, “Assessing the impact of planned social change,” Evaluation and Program Planning, vol. 2, no. 1, pp. 67–90, 1979.

• [7] R. Mannion and J. Braithwaite, “When a measure becomes a target, it ceases to be a good measure,” BMJ Quality & Safety, vol. 30, no. 3, pp. 263–267, 2021.

• [8] A. Kohn, Punished by Rewards: The Trouble with Gold Stars, Incentive Plans, A’s, Praise, and Other Bribes, 25th anniversary ed. Boston, MA: Houghton Mifflin Harcourt, 2018.

• [9] V. Sunder M., “A hybrid framework based on SIPOC and Six Sigma DMAIC for improving process dimensions in supply chain network,” International Journal of Quality & Reliability Management, vol. 33, no. 9, pp. 1269–1290, 2016.

• [10] J. W. Treem, P. M. Leonardi, and B. van den Hooff, “Coping with Goodhart’s law in an era of digitisation and datafication,” Journal of Computer-Mediated Communication, vol. 28, no. 4, 2023.

• [11] A. Delfanti, “Machinic dispossession and augmented despotism: Digital work in an Amazon warehouse,” New Media & Society, vol. 23, no. 1, pp. 39–55, 2021.

• [12] E. J. Cheon and I. Erickson, “Fulfillment of the work games: Warehouse workers’ experiences with algorithmic management,” Proc. ACM Hum.-Comput. Interact., vol. 9, no. CSCW, Art. no. 228, 2025.

• [13] R. de Koster, T. Le-Duc, and K. J. Roodbergen, “Design and control of warehouse order picking: A literature review,” European Journal of Operational Research, vol. 182, no. 2, pp. 481–501, 2007.

• [14] S. Head, Mindless: Why Smarter Machines Are Making Dumber Humans. New York, NY: Basic Books, 2014.

• [15] W. E. Deming, The New Economics for Industry, Government, Education. Cambridge, MA: MIT Press, 1993.

• [16] W. E. Deming, Out of the Crisis, 2nd ed. Cambridge, MA: MIT Press, 2000. (Original work published 1986).

• [17] E. Bonabeau, “Agent-based modeling: Methods and techniques for simulating human systems,” Proceedings of the National Academy of Sciences, vol. 99, Suppl. 3, pp. 7280–7287, 2002.

• [18] Amazon, “Inside Amazon’s fulfillment centers: What you can expect to see on a warehouse tour,” About Amazon, 11 Mar. 2019.

• [19] Amazon Technologies, Inc., “System and method for stow management of similar items,” U.S. Patent US8341040B1, 1 Jan. 2013.

• [20] A. C. Edmondson, “Psychological safety and learning behavior in work teams,” Administrative Science Quarterly, vol. 44, no. 2, pp. 350–383, 1999.

• [21] A. C. Edmondson and Z. Lei, “Psychological safety: The history, renaissance, and future of an interpersonal construct,” Annual Review of Organizational Psychology and Organizational Behavior, vol. 1, pp. 23–43, 2014.

• [22] R. S. Kaplan and D. P. Norton, The Balanced Scorecard: Translating Strategy into Action. Boston, MA: Harvard Business School Press, 1996.

• [23] W. Qian, J. Hörisch, and S. Schaltegger, “Using a balanced scorecard to manage corporate social responsibility,” International Journal of Management Reviews, vol. 22, no. 2, pp. 185–208, 2019.

• [24] T. M. Jones, “Ethical decision making by individuals in organizations: An issue-contingent model,” Academy of Management Review, vol. 16, no. 2, pp. 366–395, 1991.

• [25] S. Kim, S. B. Choi, and K. Kim, “The influence of corporate social responsibility on safety behavior: The mediating role of psychological safety and organizational commitment,” Frontiers in Public Health, vol. 10, 2022.