Internal Amazon documents suggest the company runs an AI forecasting tool called Atlas that gives each warehouse a score from 100 to 900 for how likely its workers are to organize. openDemocracy reported on October 16, 2026 that two UK sites, one near Coventry and one in Sutton Coldfield, appear in the highest-risk tier. Amazon rejects the "surveillance" label and says Atlas is not currently used in the UK.
This is a claimed-versus-verified story. The documents were disclosed in a Canadian labour dispute and read like manager training material, so what they prove about British warehouses is limited. Below is what they say, what Amazon says, what the unions and lawyers argue, and what the episode means for anyone building or buying AI that scores people at work.
TL;DR: what is reported and what is disputed
| Question | Answer |
|---|---|
| What is Atlas? | An in-house AI forecasting tool that, per the documents, predicts union drives, sit-ins and other "disruptive labour events" |
| What does it output? | A risk score from 100 to 900 per warehouse, mapped to five tiers |
| Which way is "good"? | Higher scores are "better" for leadership; a low score means workers are more likely to organize |
| Which UK sites are Tier 1? | The Sutton Coldfield robotic fulfilment centre (EMA4) and the Coventry warehouse (BHX4) |
| What data goes in? | "Hundreds of data points", including Reddit posts and answers to internal feedback surveys |
| What does Amazon say? | Site or building level only; no tracking of individuals' opinions or union membership; not currently used in the UK |
| Is it verified? | The documents are real court-disclosed material per the reporting; how Atlas is used in the UK is not established |
What do the documents say Atlas does?
Warehouse shelves and boxes representing the sites Amazon Atlas AI scores for union risk
According to openDemocracy, the documents were made public as part of a court battle between Amazon and the Canadian union Unifor. They appear to be training resources for Amazon managers in Canada, with an appendix stating that Atlas gathers "hundreds of data points" to rate warehouses in the US, Canada and the UK.
The inputs named in the report include posts on Reddit forums and employees' responses to internal feedback surveys, along with wider information about the workplace and any previous industrial action. The output is a risk score between 100 and 900. The report adds that warehouses are sorted into five tiers: Tier 5 means "no immediate perceived risk" and Tier 1 means "future activity is imminent". The document suggests the scale can be used to "inform staffing decisions, resource planning, standard actions, and ownership to mitigate risks to Amazon labour structure".
Sixteen warehouses across the US, Canada and the UK are Tier 1, according to the document. Two are British: the roughly £500m robotic fulfilment centre in Sutton Coldfield, referred to as EMA4, and the Coventry-area warehouse, BHX4.
The same documents also reached a US audience: a republished Truthout report on Resilience.org dated October 13 frames them as an AI system to spy on workers before unions form. This is a forecasting tool aimed at groups of workplaces, not a camera pointed at one worker, and that distinction is exactly where Amazon and its critics part ways.
What does Amazon say?
An Amazon spokesperson disputed the characterisation of Atlas as "individual employee monitoring or surveillance", saying: "This is site/building level activity only." The company added that it does "not track individual employees' opinions, affiliations or union membership", or any "personal data", and that Atlas is not used to monitor dissent but "to help us understand where associates need more support and to ensure we're addressing issues that they raise".
On the UK specifically, Amazon told openDemocracy that a framework called Atlas is not currently used in the UK. The outlet says the company had not answered whether a similar system under a different name is used there. That gap is the most important open question for British workers, and the reporting does not close it.
Why unions and lawyers read it differently
The reaction in the report is pointed. Michael Newman, a partner at the London law firm Leigh Day, said that scoring where higher is "better" makes clear the company thinks the union is bad, and that it tells "a very different story" from Amazon's earlier assertions that it does not seek to prevent workers from joining a union. Rachel Fagan, a GMB organiser, said the documents show sites being monitored for union activity. Rosa Curling of the tech-justice non-profit Foxglove called the described surveillance "highly intrusive" and said the firm is reviewing the documents to see whether to add arguments to an existing legal claim.
The GMB has long tried to win recognition at Coventry. Staff there were the first in Britain to strike against Amazon, in January 2023, and about 2,600 workers took part in an official recognition ballot that the GMB lost by 28 votes, according to openDemocracy (the article's own dating of that ballot is inconsistent, so we give no date). The union and Foxglove have accused Amazon of "dirty tricks" around that vote, including mass hiring and QR codes that let workers cancel their membership. Foxglove and Leigh Day represent two workers in a legal challenge to those "click-to-quit" codes. None of those allegations has been established as fact by the article, and Amazon's denials should be read alongside them.
The documents' Canadian origin also matters. openDemocracy says the British Columbia Labour Board found Amazon had violated multiple local labour laws in the dispute, after an earlier ruling that it unlawfully imposed a wage freeze on unionised workers at a warehouse outside Vancouver. That is Canadian law, not British.
Is this legal in the UK?
Nobody has tested Atlas in a UK court or before a regulator. Two things are on the record. First, a UK government spokesperson told openDemocracy that employers and workers share an interest in monitoring technologies being used "transparently and in ways that are fair, accountable and reliable", and that the government expects all employers to respect workers' rights to take part in lawful trade union activity. Second, Amazon has been penalised for worker monitoring before: openDemocracy notes France's data watchdog CNIL fined it €32m in 2022 under GDPR for tracking worker activity so precisely that alerts fired after one to ten minutes of inactivity.
Lawyers say the next step is questions rather than conclusions. Newman said that now it is known Atlas is used by Amazon, his firm will ask what data points are used and how scores are reached. That is the same transparency problem that runs through every AI scoring system: the person scored rarely sees the inputs.
What this means for people building or buying workplace AI
A scrolling log stream illustrating audit trails for workplace AI monitoring tools like Atlas
If you build or deploy AI that rates people or sites, this story is a case study in three design failures to avoid.
- Scoring direction encodes a stance. A scale where a higher number is "better" when workers are less likely to organize tells everyone what the system is for, whatever the marketing says.
- "Aggregated" is not a shield. A site-level score built from individual posts and survey answers still depends on individual data. Regulators and courts tend to ask what went in, not only what came out.
- Silent rollouts invite legal discovery. The documents surfaced because of litigation. Anything an AI tool says about workers may later be read aloud by opposing counsel and a union.
For employers, the practical takeaway is to document purpose, inputs and retention, run a data protection impact assessment before deploying any tool that touches employee communications, and tell staff. For workers, ask your employer in writing what monitoring tools score you or your site, and use data subject access requests where your law allows. If you are deploying AI agents that read internal chat, tickets or surveys, AgentBeam is the agent security platform from the explainx.ai team that stops agents before they take dangerous actions, which includes pulling data they should not touch.
How does this fit the wider pattern of AI at work?
It is one more entry in a run of workplace-AI controversies on this blog. California moved to ban workplace emotion tracking and sole-AI firing, as we covered in California's ban on emotion tracking and robo-bosses. Surveillance vendors have faced their own backlash, such as Flock Safety cutting 18 percent of staff amid surveillance backlash, and government use of voice analysis appeared in ICE's AI detainee call monitoring tender. On the labour-market side, readers can compare the claims about AI and Indian entry-level hiring. Amazon's AI footprint also extends well beyond warehouses, from an AI waiver dispute over a TV production to protests over its planned data centre in Thane.
The common thread is accountability. A model that scores a workforce is a policy decision wearing a technical costume, and it should face the same scrutiny as any other policy that affects how people can organize.
What to watch next
- Whether Amazon clarifies the UK position. The question of a similarly functioning tool under another name is still open.
- Legal steps. Foxglove says it is reviewing the documents for its existing claim, and Leigh Day intends to press for the data points and scoring method.
- Regulators. Watch for any statement from the UK Information Commissioner's Office or labour regulators; none is reported yet.
- More disclosures. The documents surfaced through Canadian litigation and a US publication. Further filings could show how widely the scoring is used.
Bottom line
The documents describe a tool that scores warehouses on how likely workers are to organize and tiers them for managers. Amazon says it is a site-level support tool, not surveillance, and not in UK use. Both can be true as stated and still leave the central questions open: what data actually feeds the score, and what managers do with a low one. Until a regulator, court or further disclosure answers those, treat the strongest claims on either side as unproven.
Details reflect the openDemocracy report of October 16, 2026 and Amazon's statements quoted there, and may change as the story develops.
Related reading
- California bans workplace emotion tracking and sole-AI firing
- Flock Safety cuts 18% of staff amid surveillance backlash
- ICE AI detainee call monitoring and voice recognition
- AI and entry-level hiring in India: claimed versus verified
- Amazon data centre protests in Thane
- AI agent safety monitoring
- openDemocracy: Amazon uses AI surveillance to monitor union risk in UK warehouses