Domain Atlas / Hiring & employment screening AI
The rule with no number to disclose
Explore this deployment in the PAN Lab ↗
In the PAN Lab, the readouts of this case's model organization carry a shaded evidence band whose width follows the least-established class among the modeling inputs the readings rest on.
The least-established input behind this case's model organization's readings is an assumption, not a measurement. Evidence base: 2 assumed · 14 published baseline.
Amazon's fulfilment centres time every task an associate performs and feed the result into structured discipline processes that generate written warnings, final warnings and terminations. A letter from an attorney for Amazon to the National Labor Relations Board dated 4 September 2018, obtained and published in April 2019, described a system in which 'Amazon's system tracks the rates of each individual associate's productivity, and automatically generates any warnings or terminations regarding quality or productivity without input from supervisors', and stated that hundreds of workers had been terminated at a single facility between August 2017 and September 2018: about 300 full-time workers at the Baltimore fulfilment centre over that roughly thirteen-month window, representing roughly ten percent of that site's workforce. Amazon disputed the characterisation on the record the day the documents were published — 'It is absolutely not true that employees are terminated through an automatic system' — adding that it would not dismiss an employee without ensuring they received coaching, that managers can intervene in the process, and that terminations can be appealed; a spokesperson described the same period as about 300 employees of productivity-related turnover at that site. Amazon contested the adjective and not the count. Six years later a U.S. Senate committee majority, working from internal documents produced to it and from two letters by Amazon's outside counsel in 2024, found in its own voice that 'When workers cannot keep up, Amazon uses automated systems to initiate disciplinary procedures. These disciplinary procedures progress in severity and eventually result in termination.' That is the finding of a committee majority, disputed by Amazon to the committee, and no court or regulator has adjudicated whether any individual termination issued without a human decision-maker. The tracker behind the 2019 documents was reported as ADAPT, the Associate Development and Performance Tracker; Amazon has confirmed that name in no source verified for this record, and the process names it put on the record with Congress are Structured Productivity Performance Review and Structured Quality Performance Review.[5]
What happened
Amazon's fulfilment centres measure work continuously. Handheld scanners and workstation terminals time each task, producing a per-worker rate and a takt time, and the gaps between scans accumulate as 'unknown idle time' or time off task — tracked, in internal documents provided to a U.S. Senate committee, in some instances down to the second, with one warning citing 97.68 minutes of it on a named date. The accumulator fills from conveyor breakdowns, station queues, pallet problems, manager conversations and restroom trips alike, and it renders all of them as the same number.
What happens to that number is the case. Amazon's counsel told the committee that its speed-related discipline process — Structured Productivity Performance Review, with a parallel Structured Quality Performance Review running on counted defects — compares 'each eligible [worker's] performance in a given week to the performance of other employees doing the same work at the same facility', and that the slowest five percent may be disciplined; as of 2020 the eligible set was the bottom five percent whose actual rate was 50 percent or less of expectation, and the committee recorded that it does not know whether that threshold remains policy. Eligibility is gated and quantified: Tier 1 associates with at least five hours in an eligible process path that week and at least 160 hours of tenure, about sixteen shifts, which Amazon described as a minority of the workers at its fulfilment centres. Crossing the threshold generates a document — written warning, then final warning, then termination.
The figure that first put this on the record belongs to the employer's own counsel and is disputed by the employer. A letter from an attorney for Amazon to the National Labor Relations Board dated 4 September 2018, obtained and published in April 2019, described a system in which 'Amazon's system tracks the rates of each individual associate's productivity, and automatically generates any warnings or terminations regarding quality or productivity without input from supervisors', and stated that hundreds of workers had been terminated at a single facility between August 2017 and September 2018 — about 300 full-time workers at the Baltimore fulfilment centre, roughly ten percent of that site's workforce. Amazon answered the same day: 'It is absolutely not true that employees are terminated through an automatic system.' It added that no employee is dismissed without coaching, that managers can intervene, and that terminations can be appealed; a spokesperson described the same period as about 300 employees of productivity-related turnover at that site. Amazon contested the adjective and not the count. The tracker was reported in that coverage as ADAPT, the Associate Development and Performance Tracker; Amazon has confirmed that name in no source verified for this file, and the process names it later put on the record with Congress are the two above. The Baltimore figures are site-scoped, are now eight years old, and remain the only per-facility termination base rate anywhere in this record; Amazon says the rate of termination is very low and has published no number.
Six years later a Senate committee majority made the finding in its own voice, from internal documents produced to it and two letters by Amazon's outside counsel in 2024: 'When workers cannot keep up, Amazon uses automated systems to initiate disciplinary procedures. These disciplinary procedures progress in severity and eventually result in termination.' The same record supplies the human step, and its position in the flow is the point. A manager holds a 'seek to understand conversation' about accrued idle time and may exempt part of it. In the warning document the committee published, the manager exempted 14 of 48 minutes for the worker's travel to and from a restroom in a one-million-square-foot warehouse and issued a first written warning for the remaining 34. The discretion is real; it operates on the accrual rather than on the decision; and the committee found instances of more than ten days between an alleged infraction and the delivery of its consequence, which it identified as making it difficult for workers to defend themselves. Above the manager sits Human Resources: a low-level manager reported that management had discretion over whether to terminate after a given number of write-ups and would not terminate when headcount was low, but only with that permission to deviate from protocol. Discipline volume moves with labour demand as well as with performance — Amazon told the committee that during the peak holiday period 'non-automated warnings, reprimands, write-ups, and improvement plans are paused', and an internal chart shows write-ups dipping through the 2019 peak and rising sharply in the first week of January. When speed write-ups were paused at the start of the pandemic, an internal Amazon team observed managers increasing behavioural, attendance and safety write-ups instead.
A second automated path runs off time rather than pace, and it is where the record shows a reconciliation step being removed. The committee found terminations for 'job abandonment' generated when automated time-tracking failed to account for workers on approved medical leave and registered large negative unpaid-time-off balances; before automation, a Human Resources employee reviewed a daily report of negative balances by hand — sometimes finding workers on leave carrying hundreds of thousands of hours of negative time — and removed them so they would not be flagged. One worker recovering from a foot injury was terminated by email a week before her scheduled return. A federal class action filed on 12 November 2025 in New York pleads this path directly, alleging an automated absence-control system that tracks attendance and then automatically imposes discipline up to termination. It is a pleading, nothing in it is adjudicated, and Amazon says claims that it does not follow federal and state law are simply not true.
What the workers do in the gap is a governance fact in its own right. They told the committee they run their own parallel timekeeping — 'I keep a timer on my watch to keep track of everything', another carrying a notebook — because they cannot see or contest the system's own ledger in time. Amazon's account of that same channel, given in response to the California citation, is that 'Employees can - and are encouraged to - review their performance whenever they wish. They can always talk to a manager if they're having trouble finding the information.' In a weighted opt-in national survey of 1,484 frontline Amazon warehouse workers fielded in 2023, 58 percent said their pace is ranked and compared with their coworkers' always or most of the time, and 45 percent said the monitoring is used mainly to control or discipline workers, against 36 percent who said it is used mainly to develop their skills.
Enforcement, when it came, landed on the notice and not on the threshold. The California Labor Commissioner's Office cited Amazon.com Services LLC $5,901,700 for 59,017 violations of the state Warehouse Quotas law at the Moreno Valley and Redlands fulfilment centres, for violations between 20 October 2023 and 9 March 2024 at $100 per violation, on an inspection opened 22 September 2022 with the Warehouse Worker Resource Center assisting. Labor Commissioner Lilia Garcia-Brower named the mechanism directly: 'The peer-to-peer system that Amazon was using in these two warehouses is exactly the kind of system that the Warehouse Quotas law was put in place to prevent.' Amazon appealed. Minnesota OSHA, after an October 2023 inspection at Shakopee, issued two serious citations in April 2024 totalling $10,500, one of them for the failure to provide workers a written copy of the quota before they were expected to meet it; Amazon contested them. Both are agency determinations rather than judicial findings and no source located for this file reports a resolution of either. Amazon's answer to the California citation states the design as plainly as anything in the record: 'The truth is, we don't have fixed quotas. At Amazon, individual performance is evaluated over a long period of time, in relation to how the entire site's team is performing.'
The statutes those citations enforce are notice-and-data statutes. California Labor Code sections 2100 to 2112, effective 1 January 2022, define a quota, require a written description of each one on hire, bar adverse employment action for failing to meet an undisclosed quota, give the worker a right to request the description and the most recent 90 days of their own personal work-speed data, and create a rebuttable presumption of retaliation for adverse action within 90 days of such a request. New York's Warehouse Worker Protection Act adds a fourteen-calendar-day no-cost response deadline, notice of changes within two business days, provision in the worker's primary language, and a right to aggregate speed data for similar workers at the same site. Six states now carry such a statute — California, New York, Minnesota, Washington, Oregon and, from July 2026, Connecticut. The only merits ruling on the quota machinery is a January 2023 dismissal in the Northern District of California of a proposed class action alleging the quotas discriminate against older workers, decided on pleading specificity: the allegations were too vague, and the court's observation that a decline in physical strength with age does not automatically mean older workers are more likely to fail the quotas is reasoning about a discrimination theory rather than a finding about the tracker.
One more instrument belongs in this file for what it does not contain. On 19 December 2024 the U.S. Department of Labor announced a corporate-wide settlement with Amazon creating Site Ergonomics Leads, annual site-level ergonomic risk assessments and employee channels for raising concerns across fulfilment centres, sortation centres and delivery stations in federal OSHA jurisdiction, for a $145,000 penalty. It contains no quota provision, no pace-setting provision and no discipline provision. The largest governance instrument attached to these buildings does not reach this decision surface at all.
The sociotechnical reading
Read this as an employment-decision system whose decision rule is made of the people it decides about. The comparison that produces discipline is computed from a store every eligible associate writes into, every shift, by doing the work: a scan is an entry. So the store is not an input to the rule — it regenerates the rule weekly, per site, out of the output of the class the rule governs. Two consequences follow that no amount of accuracy fixes. A site that speeds up raises its own bar. And a share-of-the-site rule names an eligible cohort every week whatever the site's absolute pace, which is why the employer can say truthfully that it has no fixed quota and a labour commissioner can say truthfully that it is an undisclosed one. The same fact is the defence and the violation, depending on whether you are describing the rule or the worker's ability to know it.
The second reading is about where the human sits. There is a person in this flow and the record shows them changing an outcome — a manager who exempted fourteen minutes of forty-eight. But the exemption happens before the paperwork, on the accrual, and there is no documented review of the decision itself; the escalation authority above it is documented being exercised in the direction of keeping headcount rather than of correcting error; and no exemption, override, appeal or reversal rate has ever been published by anyone. So the correction capacity here is real, thin, and pointed slightly away from the question the decision needs answered — which is not whether the worker was slow but why, and the metric that measures the first cannot represent the second. Idle time from a broken conveyor and idle time from a restroom trip are the same number.
The third reading is about what the governed party does when the ledger is unreadable to them: they build a second one. The private timer and the notebook are a record store held against the system by its subjects, kept because the official account arrives after the consequence does. Nothing on the diagram is more instructive about the state of the correction channel than the fact that this store exists, and its route into the official record — a conversation somebody else schedules — is the thinnest check on the board.
The fourth reading is about what governance actually reached. Two independent regulators inspected this design and cited it, and neither cited the threshold. Six legislatures wrote statutes and what they wrote was a read right: tell the worker the rule and the discipline attached to it, and give them ninety days of their own speed data on a fourteen-day clock. That is the cleanest case in the atlas for asking whether disclosure alone governs an automated adverse action — whether knowing the rule is worth anything when the rule is a percentile you cannot see the distribution of, and whether a right to your own data is a correction channel or a receipt.
What is deliberately absent from this reading: any injury, pace or ergonomics claim. That evidence attaches to the same buildings and to a different deployment, whose governed object is the pace parameter and whose outcome variable is the body. This deployment's governed object is the adverse employment action, and the December 2024 ergonomics settlement — the largest instrument attached to these facilities — contains no quota, pace or discipline provision at all. Served people are outside the boundary entirely: the measured associates here are the operator network, and the package recipient appears nowhere.
The concepts used in this reading are defined in the Field Guide; the governance responses live in the Practice Library. The model organization for this case can be stress-tested in the PAN Lab.