What’s trending in AI on 18 October 2026: Most workplace AI stories are about tools that help staff do their jobs. This one is about a tool that watched how staff felt about their jobs. Internal Amazon documents, disclosed in a Canadian labour dispute and reported this week by openDemocracy, Computer Weekly and others, describe Atlas, an AI forecasting system that scored warehouses on how likely their workers were to unionise, stage a sit-in or take other action. It reportedly drew on Reddit threads, an internal staff feedback board, safety incidents, staff turnover and even local minimum wages, then sorted sites into five tiers. Two UK warehouses were rated in the highest-risk tier just weeks before new UK union access rights take effect. Amazon says Atlas looks at buildings, not individuals, and that it is not surveillance. Below: what the documents describe, what Amazon disputes, why the legal exposure is real in the UK and EU, and a 10-question test you can run on any people-analytics tool in your own business.

1. What the Atlas documents describe

The documents came out of a dispute between Amazon and Unifor, the Canadian union, over organising at a warehouse near Vancouver. Amazon handed them to the British Columbia Labour Board and the union as part of that case; the board has already ruled that Amazon broke several local labour laws. The material appears to be training content for Canadian managers, including a management guide on union activity that runs to more than 100 pages, and it suggests the system covered the US, Canada and the UK.

At its core, Atlas is described as a forecasting model. It combined what the documents call “hundreds of data points” to estimate where organising might emerge next. Each site got a score between 100 and 900, where a lower number meant a higher predicted likelihood of union or other disruptive activity, and leadership reportedly treated a higher score as “better”. Sites were then placed in one of five tiers, from Tier 5, “no immediate perceived risk”, to Tier 1, where future activity is “imminent”. According to openDemocracy, the tiers could inform staffing, resource planning and steps to protect what the document calls Amazon’s labour structure.

Organising itself was logged as “employee relations events” and given severity scores. Computer Weekly reports that signing union cards, worker delegations, protests, sickouts and work stoppages were rated high severity. Those are, in most of the countries involved, legally protected activities.

2. What went into the score

The most striking thing about Atlas is how ordinary its inputs were. None of them needed a camera or a badge reader. They were the kind of data most large employers already hold, plus public social media. According to Computer Weekly’s reading of the documents, AI was used to scan for keywords and to “deep dive” into the workers driving particular discussions, while TechRadar Pro reports that the system tracked keywords, upvotes and replies on forums. That detail sits uneasily with Amazon’s statement that nothing is tracked at the individual level, and it is one of the points unions and lawyers now want explained.

Reported input	Where it comes from	Why it matters for privacy
Reddit posts (keywords, upvotes, replies)	Public forums used by warehouse staff	“Public” posts can still be personal data once collected, linked and analysed by an employer
Voice of the Associate (VOA) board	Amazon’s internal staff feedback board	Staff may not expect a feedback channel to feed a union-risk model
Feedback on managers and surveys	Internal HR surveys	Survey answers given for one purpose reused for another
Health and safety incidents, appeals, investigations	Operational and HR records	Safety complaints become a risk signal, which could discourage reporting
Tenure, turnover, training completion	HR systems	Routine workforce data, but combined into a new kind of inference
Local minimum wage, nearby unionised sites	External and geographic data	Low individual impact, but shows the model’s purpose is predicting organising
Past industrial action	Historical records	Protected activity recorded as a risk factor

The manager guidance in the same documents is more careful than the scoring. Managers were told to spot and escalate labour activity quickly, watching for signs such as unusual absences, flyers or pointed questions about benefits. They were also trained to avoid “TIPS” behaviours (threatening, interrogating, promising or spying), told not to use security cameras to identify VOA commenters, and told not to discipline anyone for VOA comments without approval from Legal. In other words, the company’s own lawyers knew where the legal lines were. The open question is whether an AI model that does the watching at scale crosses them in a different way.

3. Amazon’s answer, and the gaps in it

Amazon’s position is consistent across outlets. It says Atlas operated at “site/building level” only, was not used to monitor individual employees, did not track anyone’s opinions, affiliations or union membership, and used no personal data. It frames the tool as a way to see where staff need support, says workers are free to choose whether to join a union, and told openDemocracy: “This isn’t surveillance.” It also said Atlas is not currently used in the UK and that the system is no longer called Atlas.

Each of those statements leaves a question open. Amazon declined to name the replacement system, and did not answer whether something that works the same way runs in the UK under another name. “No personal data” is hard to square with scanning named staff forums and feedback boards for keywords; aggregation at the output stage does not mean no personal data was processed at the input stage. And a site-level label can still land on people: if a building is flagged Tier 1 and managers are trained to escalate warning signs, the practical effect is closer scrutiny of whoever works there.

None of this has been tested by a UK regulator or court yet. Computer Weekly is careful to note that the documents do not establish a breach of UK law; those are claims made by the GMB union and the campaign group Foxglove. What the leak has done is move the debate from whether such tools exist to how they should be governed. We saw the same dynamic with Flock’s licence-plate cameras: once the public learns what a system infers, the business case has to survive daylight.

4. Why the UK timing is awkward

The two UK sites named in the documents, BHX4 near Coventry and EMA4 at Minworth near Birmingham, are where GMB members have been organising. Coventry was part of the August 2022 wildcat strikes, and in 2024 the GMB lost a recognition ballot there by just 28 votes. GMB organiser Rachel Fagan told Compare the Cloud: “This is a shocking revelation.” The union has called the practices “Orwellian” and written to UK minister Kate Dearden and to Liam Byrne, who chairs the Commons Business and Trade Committee, asking the government to get answers from Amazon about any UK use.

The timing matters because the Employment Rights Act 2025 gives independent unions a statutory right to access workplaces from October 2026. Employers with more than 21 staff will have to negotiate access, and if they cannot agree, the Central Arbitration Committee can impose terms. Repeat breaches can bring fines of up to £75,000 for a first penalty, £150,000 for a second and £500,000 after that. GMB’s argument is that predictive AI could be used to blunt those rights by spotting organising before it reaches any formal threshold.

There is also a live legal front. Foxglove and the law firm Leigh Day are already challenging Amazon’s “click-to-quit” QR codes, which they say generated emails that ended GMB membership. Both say they are reviewing whether the Atlas documents support those claims. Michael Newman of Leigh Day said he will seek more detail on the data and scoring.

5. The privacy law most employers forget

Under UK and EU GDPR, trade union membership is “special category” data under Article 9, alongside health and religion. Processing it needs both a lawful basis and a specific extra condition, and the bar is high. Data protection lawyer Jonathan Armstrong, quoted by TechRadar Pro, pointed to that article and to a useful precedent: in 2020, H&M was fined €35.3 million in Germany for excessive processing of employee data. And in 2022, France’s regulator CNIL fined Amazon itself €32 million under GDPR for tracking warehouse workers’ activity too precisely.

The lesson from those cases is that regulators look at proportionality and purpose, not just whether data was “public” or “aggregated”. A model whose stated aim is to predict union activity is, by design, making inferences about a special category of data, even if it never stores a membership list. Purpose limitation also bites: staff who answer a survey about their manager, or post on an internal feedback board, did so for a reason that probably did not include feeding a labour-risk forecast.

Then there is the EU AI Act. The Digital Omnibus on AI, published in the EU’s Official Journal on 24 July 2026, pushed the start date for Annex III high-risk obligations from 2 August 2026 to 2 December 2027. Employment is one of those areas, and law firm Fisher Phillips summarises it as AI used to hire, monitor or manage workers. A system like Atlas, if used in an EU warehouse, would be a strong candidate, which would mean risk management, documentation, human oversight and logging duties. As we noted when the EU took its first AI Act enforcement steps, the delay is a window to prepare, not a reprieve.

6. This is not just an Amazon problem

Atlas is unusual mainly in being documented. Computer Weekly reported back in 2020 that Whole Foods, which Amazon owns, used a heat map to track unionisation risk across its 510 US stores. Many employee-experience platforms now offer “sentiment”, “flight risk” or “engagement” scores built from surveys, chat tools and HR data. The same techniques that predict who might leave can, with a different label, predict who might organise, complain or blow the whistle.

That makes this a governance issue for any business adopting AI tools across the workforce, not only for large retailers. It also connects to the wider trend of AI quietly turning ordinary data into sensitive inferences, as with the police tool that can match a single room photo to a home listing. In the warehouse itself, AI-driven management sits alongside heavy automation; our look at Walmart’s robot rollout showed how much of that still depends on the people on the floor. And as Anthropic’s own employment data showed this week, trust between staff and the AI systems around them is becoming a productivity question as much as an ethical one.

7. The 10-question workplace analytics test

Use this on any tool that scores, ranks or predicts anything about your staff, whether you built it or bought it as a feature in an HR, collaboration or security platform. Answer each question yes or no, give yourself one point per yes, and use the score bands at the end.

1. Purpose on paper: Is the tool’s purpose written down in one sentence that you would be comfortable reading aloud to staff?
2. No protected targets: Have you confirmed it does not predict or flag union activity, health, religion, whistleblowing or other protected characteristics, directly or by proxy?
3. Staff told: Does your employee privacy notice name the tool, its data sources and what decisions it can influence?
4. Inputs match purpose: Has someone checked that each data source (surveys, chat, forums, safety reports) was collected for a purpose compatible with this use?
5. Impact assessed: Is there a current data protection impact assessment, and does it cover AI inferences, not just raw data?
6. No drill-down: Is it technically impossible, not just against policy, for managers to go from a team or site score to named individuals?
7. Safety stays safe: Are health and safety reports and grievances excluded from any score that could count against a team?
8. Vendor answers: Has your vendor stated in writing what models it runs on your workforce data and whether it can switch on new scoring features without your approval?
9. Human decision: Is there a named person who must approve any action based on the score, with the reasoning logged?
10. EU and UK ready: If you have staff in the EU, have you checked whether the tool is a high-risk AI system under the AI Act before 2 December 2027, and if in the UK, how it interacts with the new union access rights?

9 to 10: defensible, keep reviewing yearly. 6 to 8: fix the gaps this quarter, starting with questions 2, 3 and 6. 5 or below: pause the tool until you can answer the “no” questions, and bring in your data protection officer and employment counsel. For tools that act rather than just score, add the controls in our guide to AI agent security.

8. The bottom line

Atlas shows how easily routine workplace data can be turned into predictions about the things employees are most entitled to keep from their employer. Amazon may be right that the tool never named an individual, and no regulator has yet ruled on it. But the direction of travel is clear: union data is special category data, UK access rights start this month, the AI Act will regulate workforce AI from December 2027, and European regulators have already fined employers, Amazon included, for watching too closely. For every business, the cheapest moment to ask what your people analytics actually infer is before a leak, a tribunal or a regulator asks for you.

Frequently asked questions

What is Amazon’s Atlas AI tool?

According to internal documents disclosed in a Canadian labour dispute, Atlas is an AI forecasting system that scored Amazon warehouses from 100 to 900 and sorted them into five tiers based on how likely staff were to unionise or take other collective action.

What data did Atlas reportedly use?

Reported inputs include Reddit posts, Amazon’s internal Voice of the Associate feedback board, staff surveys and feedback on managers, health and safety incidents, tenure and turnover, training completion, local minimum wage rates and proximity to unionised sites.

Which UK warehouses were flagged?

Two UK sites were rated Tier 1, the highest-risk tier: BHX4 near Coventry and EMA4 at Minworth near Birmingham. openDemocracy reports 16 Tier 1 warehouses in total across the US, Canada and the UK.

What does Amazon say about Atlas?

Amazon says Atlas works at site level only, does not monitor individuals or collect personal data, is not currently used in the UK and is no longer called Atlas. It declined to name the replacement and disputes the GMB union’s characterisation.

Is it legal to use AI to predict union activity?

It depends on the country and how the tool works, and no UK ruling on Atlas exists. Under UK and EU GDPR, union membership is special category data with strict processing rules, and regulators have fined employers for excessive worker monitoring. This is general information, not legal advice.

When do EU AI Act rules for workplace AI apply?

After the Digital Omnibus on AI, published on 24 July 2026, high-risk obligations for Annex III systems, which include AI used in employment such as monitoring or managing workers, apply from 2 December 2027.

---

Sources

• openDemocracy: Amazon uses AI to monitor union risk in UK warehouses
• Computer Weekly: Amazon uses AI-powered ‘risk’ system to track union activity globally (14 Oct 2026)
• TechRadar Pro: Amazon accused of deploying AI-powered union activity surveillance system
• Compare the Cloud: GMB calls for government inquiry after reports of Amazon AI tool (12 Oct 2026)
• Chambers: Trade unions’ right of access from October, what you need to know
• Future of Privacy Forum: The AI Act implementation timeline under the AI Omnibus
• Fisher Phillips: EU overhauls AI Act just before key deadline