← All articles

Mikhail Polyakov

Amazon used AI to rate the risk of unionisation at its warehouses: what is known about the Atlas system and what an employer can take from the story

Level: BeginnerReading time: 9 min

Topics: AI, Small business

An empty warehouse with shelves and boxes, a brass dial with a needle hanging from the ceiling and chairs arranged in a circle on the floor

The short answer: internal Amazon documents describe a system called Atlas. According to these documents, it collects "hundreds of data points" about warehouses, including Reddit posts and answers in internal surveys. Based on them, each warehouse gets a score from 100 to 900: the lower the score, the more likely, in the model's estimate, a union or other worker action. The documents became public because of Amazon's dispute with the Canadian union Unifor. Amazon replies that this is not surveillance, that the system looks at a warehouse as a whole rather than at individuals, and that Atlas is not currently used in the UK. Below, in order: what is in the documents, who says what and what an employer can take from it.

Where the news comes from

I read two publications. The first appeared on the American site Truthout on 30 September 2026, written by Jonathan Rosenblum. The author is a union organiser and, according to the byline, works part-time as an Amazon delivery driver, so this is the view of one of the sides. The documents were passed to Rosenblum by the union Unifor. The second publication appeared on the British site openDemocracy on 16 October 2026, written by Evie Breese. It adds British details and replies from Amazon, the GMB union, lawyers and the government. I did not open the Amazon documents myself: everything said about them below is as reported by these two outlets.

How the documents became public

Unifor is in a dispute with Amazon before the labour relations board of the Canadian province of British Columbia. In the course of these proceedings, according to Truthout, Amazon was required to hand documents over to the provincial labour board and the union, and the union shared them with the outlet. In openDemocracy's assessment, these appear to be training materials for Amazon managers in Canada. According to Truthout, the documents date from 2024.

What Atlas is, according to the documents

In the documents, as quoted by Truthout, Atlas is described as a "labour risk assessment framework": how Amazon assesses emerging workforce issues, union objectives and internal or external "labour threats".

What data is taken into account. According to Truthout, the model takes in:

  • training completed by employees;
  • health and safety incidents, from first aid to serious injuries;
  • employee feedback about management;
  • tenure and turnover;
  • a warehouse's proximity to another workplace with union activity;
  • the local minimum wage and the area's economy;
  • the temperature inside the warehouse and outside.

Separately, Truthout writes about messages from the workers themselves: the internal electronic board Voice of the Associate, answers to daily questions and external forums, including Reddit. According to a quote from the documents, this is to "understand the tone of the conversation at the site level". AI, according to Truthout, picks out phrases and frequent keywords and looks at how many reactions the messages get and of what kind: positive, negative or neutral.

What the result looks like. According to openDemocracy, each warehouse gets a score from 100 to 900. The lower the score, the more likely, in Atlas's estimate, that workers will form a union or take actions that affect productivity or profit. The document says that for management a higher score is "better". Warehouses, judging by the document, are divided into five tiers: tier five means "no immediate risk", tier one means that future activity, in the model's estimate, is "imminent". According to openDemocracy, the document proposes using this scale for staffing decisions, resource planning and standard actions to reduce risks to "Amazon's labour structure".

How many warehouses are in tier one. According to the document, 16 warehouses in the US, Canada and the UK. Two of them are British: the £500 million robotic warehouse EMA4 in Sutton Coldfield and the BHX4 warehouse near Coventry.

What else is in the tables. According to Truthout:

  • for the LDJ5 sortation centre on Staten Island in New York, the main complaint topics recorded are bathroom breaks, equipment, rest breaks, management and temperature, with a note of "increasing risk" next to them;
  • at four tier-one warehouses the risk was considered high enough to send in a "rapid response" team;
  • "high severity" events include signing union cards, a delegation of workers arriving with demands, protests, mass sick-outs and work stoppages.

What Amazon replies

According to openDemocracy, an Amazon spokesperson said:

  • Atlas does not monitor individual employees: "this is site or building level activity only";
  • the company does not track the opinions, affiliations or union membership of individuals or their personal data;
  • Atlas is there to understand where workers need support and to address the issues they raise;
  • a system called Atlas is not currently used in the UK. To the question of whether there is a similar system there under a different name, Amazon had not replied by the time of publication;
  • in response to GMB's accusations: "This is not surveillance", and employees choose for themselves whether to join a union.

What unions, lawyers and the government say

All of this is according to openDemocracy.

  • Lawyer Michael Newman of the London firm Leigh Day believes that a scale where "higher is better" makes it clear that the company considers a union something bad. He intends to ask what data is taken into account and how the scores are calculated.
  • Rachel Fagan, a GMB organiser, says the documents make it clear that these sites are being watched for union activity.
  • Garfield Hylton, 62, works at the warehouse near Coventry and is a GMB member. He believes the employer should not interfere: "workers should decide for themselves whether they want to join a union".
  • Rosa Curling of the campaign group Foxglove called the surveillance described "extremely intrusive". Foxglove, together with Leigh Day, represents two workers in a legal challenge against Amazon over QR codes that employees could use to leave the union. Foxglove is now examining the documents to understand whether to add new arguments to the case.
  • The UK government, through a spokesperson, replied that employers and workers have an equal interest in workplace monitoring technology being used transparently, fairly, accountably and reliably, and that all employers are expected to respect the right to lawful trade union activity.

Context of the dispute. The warehouse near Coventry saw the first strike against Amazon in the UK in January 2023. About six months later, around 2,600 workers took part in a vote on recognising the union, and GMB lost by 28 votes. Earlier still, in 2022, the French data protection regulator CNIL fined Amazon 32 million euros for overly detailed tracking of workers: according to openDemocracy, the system raised an alert if a person was idle for between one and ten minutes.

What is fact here and what is interpretation

This is my breakdown, not words from the sources.

  1. Fact according to the documents, as the outlets retell them. There is an Atlas system, there is a score from 100 to 900 and five tiers, there is a list of data including Reddit and internal surveys, there are 16 tier-one warehouses in the 2024 sample.
  2. Amazon's position. The system works at the level of the warehouse, not the person, helps understand where workers need support, it is not surveillance, and it is not currently in the UK.
  3. The other side's interpretation. Unions, workers' lawyers and the Truthout author see Atlas as an anti-union tool. This is their assessment, not an established fact.

What we do not know: how exactly the model calculates the score, how accurately it predicts, which decisions were made on its basis at specific warehouses and which of the things described are in use now. The publications do not say.

What an employer can see in this

These are my observations and conclusions from the two publications, not advice from the sources and not legal advice.

What the story shows.

  1. Internal documents can become public. Materials that openDemocracy considers training materials for managers reached the labour board and the union in the course of proceedings, and through the union reached journalists. Now the dispute is partly about the words in them: for example, that a higher score is "better" for management.
  2. Scoring by warehouse rather than by person did not end the accusations. Amazon stresses that Atlas works at the site level, yet the GMB union and the campaign group Foxglove still speak of surveillance.
  3. Regulators assess the monitoring of workers. This is a different case, unrelated to Atlas: in 2022 CNIL fined Amazon 32 million euros for overly detailed tracking.

If you are thinking of analysing employee surveys and feedback with AI yourself.

  1. Before launch. The manager who commissions the analysis writes down the goal (for example, to find areas where people complain about equipment or breaks) and the data sources. My advice: use only answers to your own surveys, without employees' posts on social media and forums. State plainly in the survey text that the answers are processed by a program and why. Show this record to a lawyer who knows employment and personal data law (for a company in Russia, that means Russian law), and do not start the analysis without their approval.
  2. How to check whether AI is useful. The same batch of answers is analysed in two ways: by AI and by an employee who reads the answers themselves. Then compare which problems each found and what each piece of work cost. If AI found nothing new, or found something but costs more than a manual review with the same result, there is no point paying for it. Whether the changes themselves helped the workers (fixing equipment, a new break schedule) will show in the next survey, but that is a test of your decisions, not of AI.
  3. If the analysis does not fit (the lawyer is against it, or AI found nothing beyond the manual review). What remains is an ordinary survey whose answers are read by a responsible employee, and conversations between managers and workers. AI is not required for this.

Summary

According to internal Amazon documents that became public because of a dispute with a Canadian union, the Atlas system gives warehouses a union activity risk score, taking into account, among other things, discussions on Reddit and in internal surveys. Amazon insists that this is analysis at the warehouse level, not surveillance, and that Atlas is not currently used in the UK. A Leigh Day lawyer intends to find out what data Atlas takes into account, and Foxglove is examining the documents to decide whether to add new arguments to its case. How the model calculates and how it was used in practice is not known from the publications.

See also another piece in this section on people's data and AI training: Closed startups' correspondence is being sold for AI training.

I work on AI agents and automation. If you would like to see my projects or discuss your own task, take a look at my portfolio.

Sources