Four ex-Amazon workers put warehouse attendance math on trial under the new pregnancy law

An accommodation approval lives in one record. The absence it creates lives in another, and the second record does not know why the hours are missing.

That gap is the substance of the case filed in Brooklyn. The Next Web's reading of the complaint puts it plainly: the plaintiffs argue that Amazon said yes to accommodations, then counted the resulting absences anyway, producing the same outcome as a denial ^[11]. Amazon's answer is an approval rate, a figure drawn from the intake queue ^[8][10].

Work the two numbers against each other. Take the low end of "tens of thousands" as 20,000 requests; a 99.9% approval rate leaves 0.1%, or roughly 20 denials in a year ^[19]. The complaint describes thousands of pregnant employees ^[4]. Even at a conservative 2,000, that is about 100 times the denial count ^[20]. Both figures can be accurate, because they count different objects. One counts forms that were processed ^[8]. The other counts people who worked pregnant in a fulfilment centre ^[4].

Teams tell themselves the accommodation programme is the request queue, because the queue is the part that emits a number. What workers do is negotiate with a supervisor about a chair or a bathroom break, and then watch what the attendance record does to them for the prenatal appointment ^[5][6].

The forcing function here is a two-by-two. One axis is whether the accommodation was approved. The other is whether the systems that count attendance were updated to reflect it. Three of the four cells end with the worker worse off, and only the approved-and-reconciled cell is compliance. It is also the only cell an approval-rate dashboard cannot tell apart from the rest ^[10].

The complaint describes a reconciliation failure, not an algorithm. It details medical documentation Amazon was allegedly not entitled to request, threats, and dismissals over excessive absences ^[6]; no automated decision system appears in it ^[21]. Reconciliation failures do not need automation to scale - they need a points ledger and a supervisor with a target.

The reason this reaches past one employer is the statute. The Pregnant Workers Fairness Act took effect in 2023 and requires reasonable accommodation unless it would impose an undue hardship ^[12], and it removed the older requirement that a pregnant worker show she was treated worse than a comparable colleague ^[13]. "We treat everyone in the building the same way" is no longer a defence. What remains is whether the accommodation happened, which is a question about operational records rather than policy documents.

The Next Web reported a similar suit against Amazon over pregnant warehouse workers in 2019, which is context and not evidence for this complaint ^[16]. None of the current allegations has been tested, and Amazon has not been found to have done anything wrong ^[9].

For anyone who owns an accommodations workflow, the test that settles it internally is short. For each accommodation granted, name the downstream record it has to change, and name the person who verifies that it changed. If nobody can name either, the approval rate measures the form, not the floor. That is also the number a court would be reading if the class is certified, which is the stage where this stops being four people's accounts ^[18].