PAN Lab example

Amazon fulfillment-centre algorithmic management

Units per hour up and a cost measured in bodies

Amazon's warehouse software sets workers' tasks and pace. Throughput rose, and a federal regulator and a Senate committee tied that pace to worker injuries.

Amazon's fulfillment-centre production system is in-house software. It assigns each worker's pick, stow, and pack tasks, sets how many tasks an hour they must complete (their rate), and tracks it. In robotics facilities, robots bring shelves of items to the workers' stations.

The benefit

Amazon's own researchers published the redesigned picking algorithm's results in a peer-reviewed journal in 2023. It cut the distance the robots drive per item picked by 62 percent, and the number of robots needed by 31 percent. The authors report about $0.5 billion in direct savings from 2017 to 2020.

The case file calls the gain from robots and people working together genuine and documented, not a marketing claim. On units per hour, the measure the system is built to raise, it works.

How the pace is set and enforced

Amazon calls the number of tasks a worker must complete each hour their rate. Workers in robotics facilities told a Senate committee of expected rates of up to 350 items stowed and 450 items picked an hour. They are expected to keep the rate for the whole shift.

When workers cannot keep up, automated systems start disciplinary procedures that can end in termination, the committee's report says. In roles like Pick and Pack, Amazon ranks workers by items processed and sends automatic write-ups to the lowest five percent. Amazon told the committee this applies only to entry-level workers, and that the rate of termination is very low.

What the regulator found

OSHA, the U.S. Occupational Safety and Health Administration, is the federal workplace safety regulator, part of the Department of Labor. In January 2023, it announced inspections at three Amazon warehouses. They found that the company exposed workers to ergonomic hazards, risks from how the work loads the body. The sites were in Deltona (Florida), Waukegan (Illinois), and New Windsor (New York).

OSHA proposed $60,269 in penalties, in what was described as its largest ergonomics enforcement effort. On 1 February 2023, it announced findings at three more sites, in Aurora (Colorado), Nampa (Idaho), and Castleton (New York). The hazards it named include high-frequency lifting and the long hours needed to complete assigned tasks.

What the Senate committee found

On 15 December 2024, the majority staff of the U.S. Senate Committee on Health, Education, Labor, and Pensions published a report on an 18-month investigation. Its title is The Injury-Productivity Trade-off: How Amazon's Obsession with Speed Creates Uniquely Dangerous Warehouses.

The report found that Amazon warehouses recorded over 30 percent more injuries than the warehousing industry average in 2023. It found Amazon workers nearly twice as likely to be injured as other warehouse workers in each of the past seven years. It says Amazon claims its warehouses are nearly as safe as the industry average, and reaches that claim by cherry-picking data.

It found that the rates force workers to move at an extremely fast and often dangerous pace. Workers report bypassing safety procedures to keep up. The report concludes that they must choose between following those procedures and risking discipline for moving too slowly.

What Amazon's own studies found

The report says Amazon studied the link between speed and injury for years. In 2020, Amazon launched Project Soteria, a 2020 initiative of several Amazon teams that studied warehouse injuries and proposed ways to lower injury rates. It found that pausing discipline for missed rates and giving more time off both lowered injury risk. Amazon denied the team's request to keep those policies.

Leadership then directed the team to find ways to maximize rates without increasing injuries. The team called this the injury-productivity trade-off, the phrase in the report's title. In 2022 it again found a link and proposed slowing the pace. Amazon rejected that proposal.

In 2021, another team, Project Elderwand, found how many times per shift a worker in one role could repeat the same tasks before the risk of harm rose. Amazon decided not to put that limit in place.

The settlement

On 19 December 2024, the U.S. Department of Labor, of which OSHA is part, announced a settlement with Amazon. A settlement is a legal agreement between a regulator and a company. This one requires ergonomic measures at facilities across the country. It requires ergonomic risk assessments, piloted engineering controls, Site Ergonomics Leads, and channels for workers to raise concerns. The sources read for this case do not say how well these roles work.

Why the measure misses the cost

The case file says units per hour cannot show the cost of the pace, because that cost is not counted in units. It is borne by workers' bodies. It shows up in safety inspections, a legislative report, and worker testimony, not on the operations dashboard. So throughput can rise while the injury cost grows off that number.

The robots genuinely help. The case file says the question is whether the pace-setting treats a sustainable rate as part of what optimal means, or pushes the cost out as injury.

Who owns the pace

The case file says the pace is chosen and enforced by the system, not discovered. Field research by Cheon and Erickson, from two years of fieldwork in Amazon fulfillment centres, describes the work as a game whose rules the worker cannot change.

The targets, the pace, and the measures are set by the system and the organization behind it. So the case file concludes that the injury tied to that pace is a consequence the organization owns. "The algorithm set the rate" is no defense.

What the available tools can and cannot address

A failure pathway is a link between two parts of the network where a mistake made by one part can be passed on to the other. A pathway counts as closed once it passes on only a few mistakes. It need not stop them all. Closing one means mistakes stop passing along it, not that the work along it stops.

This case's budget is 12 units. Each tool costs units from it. The tool named Understand the system is not offered here. Six failure pathways here start open.

Explore (No Targets) sets no targets. There, two tools costing 4 units are enough to stop mistakes building on one another across the network. One pair is Peer sharing rules with Train the staff.

Under Service Targets Only, the targets are to stop mistakes building on one another and to keep the system clearly helping the work. Service means the benefit the deployment delivers. Five pairs of tools meet the targets for 4 units. One is Assign a challenger with Escalate checks. More than 240 different sets of tools meet them within the budget.

Under Service and Safety Targets, the targets include closing every failure pathway. Two of them can close. Escalate checks closes Tasks and pace set for the worker. Gate record entries closes Rates and task times recorded.

No tool offered here acts on the other four. They are Work done at the set rate, Rate data used for pacing, Rates shown to the worker, and Low rate turned into a write-up. So at this level the case is not fully addressable with the available tools. Every combination was checked.

Under All Governance Targets, it is not fully addressable either, for the same reason. That level also asks for a larger, steadier gain in service and a target on the deployment's environmental footprint.

Stylized model of a documented deploymentLogistics dispatch & scheduling AI

Open this example in PAN Lab v0.1 to apply pressures and levers and watch what the system does.

What this models

This example runs on the Algorithmic-management-class whose pace is coupled to injury network: 7 components and 14 pathways between them. Every context in the Lab is a stylized model, never a reconstruction of any actual deployment, and each assumption behind it carries a provenance label.

Evidence base: 3 assumed · 2 published baseline. In the Lab, the shaded evidence band behind each headline readout draws its width from the least-established class below.

Show all 5 assumptions

assumed
Three documented parts are drawn beside the system, the workers, and the record. The first is the pick and stow task sequence. It turns a throughput decision into what a worker reaches for, and how often. No pathway connects it, so it does not change how mistakes move in this example. The second is the automatic write-ups, which follow from the rate record. The step that would weigh a missed rate against the day's conditions is drawn as a check on them. The same rate count sets the pace and triggers the discipline, which makes the rate a verdict as well as a measurement. The third is the Site Ergonomics Leads and concern channels, drawn as a second group of staff. A December 2024 OSHA settlement required them. They are human authority over a workflow the system manages, put there by an enforcement action. Their pathway into the pace-setting is drawn too. Whether ergonomic findings change the rate or are set aside is the open question it poses. The example assumes a heavy workload, with very little staff capacity to absorb it.
baseline
This example follows the pattern the case file documents for Amazon's fulfillment centres. It is not a copy of Amazon's actual system. The system pairs a peer-reviewed benefit, robots and people picking together, with a documented injury cost. OSHA, the federal workplace safety regulator, cited Amazon for exposing workers to ergonomic hazards, risks from how the work loads the body. A Senate committee report described an injury-productivity trade-off and called the warehouses uniquely dangerous. The benefit is entered as the peer-reviewed research reports it. The cost is entered as the inspections and the report record it.
assumed
The pathway named Pace couples throughput to injury stands for one pace seen from two sides. Units per hour rise because the system sets a rate. That rate is what OSHA and the Senate committee tied to injuries from the strain of the work. So the gain and the injury risk are one fact, not two that happen to occur together. The workers are drawn as the people the system paces. The pathways that write each worker's rates to the record, and that use them for pacing, are marked as private. They carry monitoring data on individual workers.
baseline
Two checks mark where the cost could be made visible. The sustainable-pace check asks whether the pace-setting treats a sustainable rate as part of what optimal means. A throughput measure cannot see an injury cost, which is borne by workers' bodies. The injury-cost accountability review would read the injury the throughput measure leaves out, and treat the pace as a management decision. Field research in Amazon warehouses describes the work as a game whose rules the worker cannot change. So the organization chooses the pace, and the injury tied to it is a consequence it owns, not a fact of the work.
assumed
This example models no worker injury or safety outcome. It shows how mistakes can pass between parts of the operation, and it draws the workers as the people the system paces. The case file holds the benefit of robots and people picking together, the OSHA ergonomic-hazard citations, and the Senate committee's injury-productivity finding. It also holds the research describing the work as a game. None of them is computed from anything in this diagram.

What this example does not show

Show all 2 limitations

This example models no worker injury or safety outcome. It shows how mistakes can pass between parts of the operation, and it draws the workers as the people the system paces. The case file holds the benefit of robots and people picking together, the OSHA ergonomic-hazard citations, and the Senate committee's injury-productivity finding. It also holds the research describing the work as a game. None of them is computed from this diagram.
The throughput benefit is entered as the peer-reviewed research result. The injury cost is entered as the recorded findings of safety inspections and a Senate committee report, not as a harm this example computes. The sustainable-pace check and the injury-cost accountability review are drawn as checks. The sources record neither as adopted by Amazon.

Sources and evidence

What this example rests on, claim by claim. Every entry resolves to the same ledger the Evidence Registry publishes.

A warehouse operation's algorithmic management pairs a genuine, peer-reviewed human-robot picking benefit — robots and workers collaborating to raise throughput, documented in the operations-research literature — with a documented injury-productivity trade-off. When the algorithm sets the pace of the physical work, a federal safety regulator cited the operation for exposing workers to ergonomic hazards, and a legislative inquiry tied the speed the system demands to warehouses it described as uniquely dangerous. The productivity gain and the worker-injury risk are therefore coupled: the same pace that raises units per hour is the pace regulators and the inquiry connected to injury. The benefit is real and the injury cost is separately documented, one in the OR literature and one in safety-inspection findings and a legislative report.
empirical
- Academic Allgor, R., Cezik, T., & Chen, D. (2023). Algorithm for Robotic Picking in Amazon Fulfillment Centers Enables Humans and Robots to Work Together Effectively. INFORMS Journal on Applied Analytics, 53(4). https://doi.org/10.1287/inte.2022.1143
- Government U.S. Senate Committee on Health, Education, Labor, and Pensions (2024, December 15). The Injury-Productivity Trade-off: How Amazon's Obsession with Speed Creates Uniquely Dangerous Warehouses (Majority Staff Report) https://www.help.senate.gov/imo/media/doc/amazon_investigation.pdf
- Regulatory U.S. Department of Labor, OSHA (2023, January 18 and February 1). Federal safety inspections at Amazon warehouse facilities find company exposed workers to ergonomic, struck-by hazards (national news releases). https://www.osha.gov/news/newsreleases/osha-national-news-release/20230201
The lesson the case carries is that when an algorithm sets the pace of physical work, the productivity metric it optimizes — units per hour — cannot see the cost the pace imposes on the body executing it. The injury shows up in safety-inspection data and a legislative inquiry, not on the throughput dashboard, so a productivity number can rise while the cost accumulates unrecorded on the metric that reports success. The governable question is whether the pace-setting internalizes the worker's safety, treating a sustainable rate as part of what 'optimal' means, or externalizes it as an injury the metric never records. Ethnographic research describes this algorithmic management as a 'game' whose rules the worker cannot change, which is what makes the pace a management decision the organization owns rather than a fact of the work.
empirical
- Academic Cheon, E., & Erickson, I. (2025). Fulfillment of the Work Games: Warehouse Workers' Experiences with Algorithmic Management. Proceedings of the ACM on Human-Computer Interaction (CSCW). https://doi.org/10.1145/3757409 https://arxiv.org/abs/2508.09438
- Government U.S. Senate Committee on Health, Education, Labor, and Pensions (2024, December 15). The Injury-Productivity Trade-off: How Amazon's Obsession with Speed Creates Uniquely Dangerous Warehouses (Majority Staff Report) https://www.help.senate.gov/imo/media/doc/amazon_investigation.pdf

Where this connects

Institutional pressures in this domain

Austerity & recovery incentives — Cost-cutting and overpayment-recovery targets tilt the system toward denial and enforcement errors.
Reviewer bottleneck — One fixed-capacity checking stage sits between AI output and consequence; everything queues behind it.
Data & policy drift — The world, the intake process, and the rules change under a system trained on how things used to be — two mechanisms with different remedies: the statistical properties of what the system processes move (concept drift), or the mixture of inputs arriving in deployment differs from the mixture it was trained on (covariate shift).
Vendor opacity — The deploying institution cannot inspect the model, data, or update pipeline it is accountable for.
Compliance over substance — Paper controls (sign-offs, checklists) satisfy audits while the behavior they describe erodes.

All of them in context on the Logistics dispatch & scheduling AI domain page.

Levers available here and the patterns behind them

Peer sharing rules — Peer-edge governance
Assign a challenger — Structured dissent
Keep skills sharp — Deskilling-arrest mandate
Train the staff — AI literacy & boundary rules
Review on schedule — Oversight cadence & retrospectives
Gate record entries — Human-in-the-loop write gating
Check copied records — Reconcile copied records
Review the riskiest first — Risk-tiered oversight
Escalate checks — State-feedback vigilance
Upgrade model — Improve the model