Amazon scores its warehouses on a scale of 100 to 900, and a lower number means the people inside are more likely to organise. Internal documents released during a court battle with a Canadian union describe Atlas, an in-house forecasting tool that pulls “hundreds of data points” (Reddit posts, internal feedback survey responses, prior industrial action) to rank sites across five tiers. Sixteen sites sit in Tier 1, where “future activity is imminent”. Two of them are in the UK, including the Coventry warehouse where 62-year-old Garfield Hylton works.

Atlas is a functioning example of predictive workforce analytics built to manage employees rather than support them. If you are building models on operator data, the Atlas case shows exactly where the line sits, and what crossing it costs you.

A Coventry Warehouse Worker Found Out He Had a Risk Score

Hylton already knew the basics of the deal. Every product scanned, every second between scans, every pause while his scanner sat idle, all of it logged and turned into productivity targets. That part was never hidden. What he did not know was that his site was also being read for “dissent”, with social media posts and the internal staff bulletin board feeding the model.

Amazon’s spokesperson insists this is “site/building level activity only” and that the company does “not track individual employees’ opinions, affiliations or union membership”. Fine. The scoring still happened, and the worker on the receiving end found out from a court filing.

Here is what should concern operations leaders: the same forecasting stack being sold to you for throughput and defect prediction works just as well pointed at your own people. Almost nobody has written down where that line sits.

What Atlas Actually Does: Hundreds of Data Points, One Score

The 100-900 score and the five-tier risk ladder

The scale runs backwards from what most people expect. A high number is a quiet site. A low number means the model believes organising, a sit-in, or other “disruptive labour events” are coming. The training documents for Canadian managers put it bluntly: for leadership, higher scores are “better”.

Those scores then collapse into five bands. Tier 5 is “no immediate perceived risk”. Tier 1 means “future activity is imminent”. Sixteen buildings across the US, Canada and the UK sit in that bottom tier, including EMA4, the £500m robotic fulfilment centre at Sutton Coldfield in north-east Birmingham, and BHX4 near Coventry.

Tiering matters more than the raw number, because tiers trigger action. The documents say the scale can be used to:

“inform staffing decisions, resource planning, standard actions, and ownership to mitigate risks to Amazon labour structure”

That is a scoring model wired directly into headcount and budget decisions. Any operations leader who has built a risk index knows what happens next: once a tier drives resourcing, the tier becomes the decision.

Which signals feed the model, and which are scraped from outside work

Some of the inputs are ordinary operational data. Workplace characteristics and a site’s history of industrial action are the kind of context any labour relations team would already hold in a spreadsheet. Feeding them to a model is a mechanical upgrade, not a moral one.

The rest is where it gets uncomfortable. Atlas also ingests responses to internal feedback surveys, activity on the company’s internal staff bulletin board, and posts on public Reddit forums. The first two turn an employee listening channel into a detection channel. The third reaches outside the building entirely.

Amazon says Atlas is “site/building level activity only” and that it does not track individual opinions, affiliations or union membership. Aggregation is a real distinction and worth taking seriously. It also does nothing to change the source of the signal: people speaking candidly in channels they believed were for raising problems, not for scoring them.

Site-Level or Individual? Why Amazon’s Defence Doesn’t Settle the Question

The gap between ‘no personal data’ and ‘no personal consequences’

Take Amazon at its word. Assume Atlas holds no names, no union memberships, no individual opinions. The defence still leaves the hard part untouched, because the documents describe scores being used to “inform staffing decisions, resource planning, standard actions, and ownership to mitigate risks to Amazon labour structure”.

Staffing decisions land on people. A building-level score that changes headcount, shift patterns, or which manager gets parachuted in has individual consequences whether or not a single employee ID appears in the model. Aggregation is a privacy argument, not a fairness argument, and the two get confused constantly.

If you are building warehouse workforce analytics, test your own models the same way. Ask what decision the output triggers, who feels that decision, and whether anyone affected could see the reasoning. “No personal data” answers none of those.

Model intent is revealed by what counts as a good score

Every predictive model encodes a preference in its scoring direction. Atlas treats a high number as a quiet site, which means the thing being minimised is collective action by employees. Michael Newman, a partner at Leigh Day, read it exactly that way.

This “makes it clear that they think the union is bad”

Newman said it tells “a very different story” from Amazon’s previous assertions that it does not seek to prevent workers from joining a union. That gap between stated scope and scoring logic is the useful lesson here. A company can describe a tool as support-oriented while its objective function rewards the absence of worker organising.

So when your team specifies any employee monitoring AI, write down the target variable first and read it out loud in a room with HR, legal, and a shop-floor representative present. If the sentence describes suppressing something employees are legally entitled to do, the model is not a measurement tool. The scope language will not save you later.

The Same Stack, Pointed Somewhere Useful

Early-warning scoring for quality and safety, not for dissent

Strip the politics out of Atlas and what remains is a fusion model: dozens of weak signals, combined into a single site-level score that moves before the event does. That capability is genuinely useful. Most plants already hold the inputs, just scattered across systems that never talk to each other.

Engagement survey free text sits in HR. Overtime and absence sit in the shift system. Near-miss reports sit in EHS, scrap and rework sit in the MES, and supplier deviations sit in a spreadsheet someone maintains by hand. Fused, those predict things worth predicting: scrap rate drift on a specific line, a near-miss cluster three weeks before an injury, an audit finding that was visible in deviation patterns all quarter, an absenteeism spike that reliably precedes a quality escape.

The three questions to answer before you build a workforce model

Before anyone writes code, answer three things in plain language and write them down.

• What decision does the score drive? “Dispatch a process engineer to Line 4” is a decision. “Inform staffing decisions” is not, it is cover for whatever someone decides later.
• Who sees it? If the operators generating the signal cannot see their own site’s score, you have built a monitoring tool and called it analytics.
• What remedy does it unlock? Name the fix in advance. Extra training hours, a tooling change, a maintenance window, a headcount add.

That third question is the whole test. Amazon’s own documents describe scores used to “mitigate risks to Amazon labour structure”, which is a remedy aimed at the signal rather than the cause. A model whose output is “reduce the chance people complain” is pointed backwards. A model whose output is “fix what they complained about” earns the data it consumes, and it keeps earning it, because people feed honest inputs into systems that visibly act on them.

What UK Manufacturers Should Expect Next on Workforce AI

Atlas did not become public through a journalist’s FOI request or a whistleblower memo. It surfaced because documents were disclosed in litigation. That is now the most common route by which internal tooling gets read aloud, line by line, by people who did not build it and owe it no charity.

Expect regulators and worker representatives to converge on two questions through 2027: what purpose was this data collected for, and does the model make or materially shape decisions about employment. UK GDPR already treats automated decision-making in employment as high risk. Works councils across the EU will ask the same thing in blunter language.

Build models you would be comfortable disclosing

The test is simple. If the training deck for your managers were handed to a union lawyer tomorrow, would the wording embarrass you? Amazon’s documents told on themselves by stating that higher scores are “better”, which is how a risk model reveals whose side it is on.

Take this checklist into the next steering meeting:

• Written purpose statement: one sentence per model, naming the problem it solves and the decisions it is barred from informing.
• Input register: every data source listed, with the lawful basis and the system it came from.
• Human decision owner: a named person who acts on each score, with the authority to overrule it.
• Worker-facing explanation: what the model measures, in plain English, published internally before go-live.
• Retention and review dates: models drift, and an unreviewed one becomes a liability.

Trust is the input your workforce data depends on

Engagement surveys, near-miss reports, and quality escapes all depend on people choosing to report honestly. Once staff suspect that free-text answers feed a risk score, the free text stops being useful. You degrade your own dataset.

Predictive workforce analytics earn their keep when they surface problems leadership can actually fix: understaffed shifts, a supervisor with a pattern, a line where rework keeps climbing. The moment the same stack is used to forecast and pre-empt people, it stops measuring anything real.

Source: opendemocracy.net