Amazon's internal AI forecasting tool Atlas assigns each warehouse a union-risk score between 100 and 900, aggregating hundreds of data points — Reddit posts, internal survey responses, local labour history — to predict the likelihood of unionisation, sit-ins, or other 'disruptive labour events.' Two UK warehouses are classified Tier 1, the highest-risk category: a £500m robotic-fulfilment centre in Sutton Coldfield (EMA4) and the Coventry warehouse (BHX4) where GMB organiser Garfield Hylton works. The documents, surfaced through a Canadian legal challenge by union Unifor, are framed as manager training resources. Lower scores mean higher union risk. For Amazon leadership, higher scores are 'better.' The system's architecture is straightforward: scrape signals of collective discontent from social media and internal channels, feed them into a predictive model, and output a risk tier that 'informs staffing decisions, resource planning, standard actions, and ownership to mitigate risks to Amazon labour structure.' That last phrase is doing all the work. 'Mitigate risks to Amazon labour structure' is corporate euphemism for preemptive union suppression — resource allocation guided by surveillance to neutralize organizing before it gains critical mass. Amazon's spokesperson drew a careful line: Atlas monitors 'site/building level activity only,' not individual employees. The company says it does not track 'individual employees' opinions, affiliations or union membership.' But the distinction between individual and site-level surveillance collapses when a warehouse has identifiable organisers, when internal survey responses feed the model, and when the output guides staffing decisions that directly affect the workers producing those signals. Amazon also told openDemocracy that 'a framework called Atlas is not currently used in the UK' — but did not respond to the question of whether a functionally identical system operates under a different name. The Coventry warehouse is ground zero. GMB lost a formal union recognition ballot there in July 2024 by just 28 votes, after workers staged the first-ever strike action against Amazon in the UK in January 2023. GMB and tech-justice nonprofit Foxglove have accused Amazon of mass-hiring non-unionised workers before the vote and plastering 'click-to-quit' QR codes — allowing workers to cancel GMB membership — across screens, cafeteria tables, and toilet doors. Foxglove and Leigh Day are pursuing a legal challenge against the QR code tactic as unlawful inducement. The Canadian proceedings that produced these documents ended with the British Columbia Labour Board ruling Amazon violated multiple local labour laws, including an unlawful wage freeze on unionised workers near Vancouver. The €32m GDPR fine levied by France's CNIL in 2022 — for tracking worker inactivity down to one-minute intervals — established precedent that Amazon's surveillance appetite routinely overshoots legal boundaries. Atlas extends the same logic from individual productivity tracking to collective action prediction. The UK government's response was notably passive: employers and workers have 'a shared interest in workplace monitoring technologies being used transparently and in ways that are fair, accountable and reliable.' This frames the power asymmetry as a collaboration problem rather than a structural one. Amazon has the capital to build bespoke AI union-detection systems, the data infrastructure to feed them, and the operational flexibility to act on their outputs through staffing and resource decisions. Workers have, as Hylton put it, 'only their own experiences to defend themselves.' The fundamental question is not whether Atlas tracks individuals or sites — it is whether an employer should be permitted to build predictive infrastructure whose explicit purpose is to identify and suppress the conditions under which workers exercise a legal right. UK law protects the right to organise. A system designed to detect organising before it succeeds, and to allocate resources to prevent it, is functionally an anti-union weapon regardless of whether it logs a single name.
