At some point, your employer has asked how you feel about work, in a pulse survey or on a suggestion board: how stressed, how satisfied, whether you feel included. You answered, or skipped it, or wrote something blunter than you'd say to a manager's face. And at least once, you've probably wondered who reads the answers.
At Amazon, the question comes every day, on "computers, workstation devices, or hand scanners," and the company credits it with "surfacing issues before they become acute problems." Beside it sits a board, the Voice of the Associate. Amazon calls this listening; in October, its spokesman told The Globe and Mail that the system at the center of this piece "helps us listen to our teams' suggestions."
An internal Amazon report, handed over in a Canadian labor case, lists those two channels among that system's inputs. It calls the system Atlas, "a labor risk framework" for evaluating "emerging workforce issues, labor union targets, and internal or external labor threats." Its factors include "sentiment measures from the Voice of the Associate tool" and "Select survey data from Connections" — the daily question — down to the "Rather Not Answer rates." Even declining to answer is in there.
One worker posted a suspicion on the board: that "this platform is design so that leaders will know what your concerns are," so they can say they dealt with them "before you the associate take the real issues a step higher." The report reproduces the post, unsigned, as an example.
Listening tool or early-warning system? The pages alone can't settle it, but they reward a close read.
Where 900 means "Events not likely to happen"
Where do these pages come from? A case before British Columbia's labor board forced Amazon to hand over documents, and Unifor, the Canadian union on the other side, shared the documents with Truthout, which published them on September 30, 2026 (its author is a union organizer and part-time Amazon delivery driver). So I went to the 153-page file itself. The Atlas pages, from page 19, are a spreadsheet report bound in with Canadian manager-training slides, and they never say who they were written for.
Atlas sorts sites into five tiers, and Tier 1 involves no forecast; it "is assigned to sites with confirmed disruptive labor events." Tiers 2 through 5 come from "a mathematical forecasting model" that gives each site "a score between 900 (Events not likely to happen) and 100 (Events highly likely)." A footnote on the same page puts it plainly: "Higher Labor Score is better."
Which events? The high-severity list reads like a glossary of organizing: "Active Card Signing," an "Associate List of Demands," protests, sick-outs, walkouts and sit-ins, at a risk level that "theoretically refers to having more than 15 associates involved." A tier graphic runs from "No immediate perceived risk" to "Future activity is imminent," where "ER's Rapid Response Team will deploy within 24 hours in consultation with" legal, operations, PR and other teams on "de-escalation strategies." (ER is Employee Relations; bear with the acronyms.)
What feeds the forecast? "The hundreds of data points used in creating the Atlas Advisor forecast have been grouped into a taxonomy of factors," the appendix says, and they read like a portrait of a workplace: training, survey-based safety perceptions, investigations, pay defects, attrition, "(hidden)" diversity data, distance to a site "experiencing the forecasted event type," the "local minimum wage rate" — and, yes, the board and the daily question.
On Reddit, the coverage and the pages part ways. Truthout wrote that Amazon was "scraping data from worker communications channels to plug into the Atlas model," Reddit among them; in the pages, Reddit turns up only in a companion "ER Trend Detection model" that tracks trending phrases in board comments and "posts on Amazon Ops focused forums on Reddit." Whether that model feeds the score, they don't say. (Is any of this AI? The Globe calls Atlas "an AI tool"; Amazon says teams, "not AI," make the decisions.)
How big, and how old? The Tier 1 table lists 16 sites: 12 in the U.S., two near Vancouver and two in the U.K. One row — a sort center on Staten Island — pairs the themes in workers' feedback ("Bathroom Breaks, Equipment, Rest Breaks, Leadership/Management" and temperature) with the notes "Internal Activists" and "Risk Increase." The Globe reports Atlas in use "in recent years across more than 50 warehouses in the United States, Canada and Europe." The report's charts run from December 2023 to February 2024. And its sentiment section asks, "What anecdotes can we pull?" That is where the post above sits.
What the pages leave out
What don't they show? The data itself, for one: "The documents do not indicate what exactly the data points are, or how Amazon collects them," the Globe noted. Accuracy, for another: a key-terms page defines precision without reporting the model's own, and warns that unsupervised results "can be 'noisy'" and that "it is valuable for humans to conduct a layer of interpretation." Consequences: the Tier 1 table marks a Rapid Response deployment at four sites, and the documents don't say what happened there.
And the present. Amazon told Computer Weekly that the system is no longer called Atlas, and declined to say what it's called now. Asked by the Globe whether Atlas was still used in Canada, the company did not respond specifically.
The case for calling it listening
Amazon's answer deserves a full hearing. "This is site/building level activity only," a spokesperson told openDemocracy; the company does "not track individual employees' opinions, affiliations or union membership," or any "personal data," and Atlas exists to "help us understand where associates need more support and to ensure we're addressing issues that they raise." A framework called Atlas, Amazon added, is not currently used in the U.K. To the Globe, spokesman Sam Stephenson gave the full line: "This framework helps us listen to our teams' suggestions to drive continuous improvement, and how we act on their feedback is determined by the teams responsible for driving that improvement, not AI." And: "Interpreting this feedback requires human review, not automated decision-making." Amazon also told Computer Weekly that talk of minimizing disruption in the documents refers to industrial action, should it happen.
The documents back some of that. In the report, an "ER Solutions team manually reviews" events for misclassification. The Canadian training in the same file tells managers who find a troubling post: "Never try to find their name if they did not share it," and "do not attempt to review security video or spy in any way." It reminds them that non-supervisory employees "have a right to write VOA messages complaining about working conditions at Amazon," and bars discipline for a comment "without approval from Legal."
Surveys themselves are ordinary; a union-avoidance guide from the management-side law firm Conn Maciel recommends them because they help "uncover workplace concerns before they escalate." And remember that these documents reached the public through one side of a labor dispute.
So far, so fair. My difficulty is with "site/building level activity only." Organizing is never one person. Amazon's own training says the easiest way to recognize legally protected group activity "is to look for anything that involves more than one person," and lists "Posting on the VOA board or MyVoice" under "Where might you find evidence of organizing?" A forecast that never names anyone is still a forecast of people acting together — the very thing American labor law protects.
Is a building-level forecast "spying"? U.S. law hasn't said
Federal law protects workers' right "to engage in other concerted activities for the purpose of collective bargaining or other mutual aid or protection," and bars employers from acting to "interfere with, restrain, or coerce" workers who use it. The NLRB's guidance forbids spying on union activity, defined as "doing something out of the ordinary to observe the activity," and creating the impression of it. Does a model that forecasts organizing building by building, from data the company already collects, count? As far as I could find, the Board has never said.
It might have. In 2022, the agency's then-general counsel, Jennifer Abruzzo (its chief prosecutor, not the Board), wrote that she would urge the Board to find that an employer has "presumptively violated Section 8(a)(1)" where its surveillance and management practices, viewed as a whole, would tend to interfere with protected activity, and to require disclosure of "the technologies it uses to monitor and manage them." She allowed that "some employers may have legitimate business reasons" for monitoring, and noted settled law against using "technologies already in place for the purpose of discovering" protected activity. Acting General Counsel William B. Cowen rescinded the memo on February 14, 2025, citing the case backlog; as of October 24, 2026, nothing has replaced it.
In Congress, the Stop Spying Bosses Act (S. 4831, with a House companion, H.R. 9402) would bar using employee data to identify a worker "to determine if they have or intend to form, join, assist" a union. Its clauses are written about "any covered individual," so whether a building-level score is covered is left open; both versions have sat in committee since June. In California, AB 1883's April draft would have barred using a workplace surveillance tool to "infer information about workers engaging in a protected activity." As I reported on October 4, that clause is gone; the law signed September 30 bans AI emotion-reading and neural-data collection at work instead.
In France, Amazon's warehouse numbers went before a judge
Must it take a legal fight to find any of this out? France offers a partial answer, from Amazon's own warehouses.
On January 23, 2024, the CNIL, France's privacy regulator, announced a €32 million fine against Amazon France Logistique, the company's French warehouse business. AP reported that the watchdog faulted an "excessively intrusive system" for monitoring worker performance, saying "it was illegal to set up a system measuring work interruptions with such accuracy, potentially requiring employees to justify every break or interruption." Amazon called the conclusions "factually incorrect" and said warehouse management systems "are industry standard."
Amazon appealed, and on December 23, 2025, the Conseil d'État, France's highest administrative court, cut the fine to €15 million. Its decision traces the case to CNIL inspections in November 2019 at the company's offices and two of its warehouses. Amazon won on all three indicators: the court found the real-time "stow machine gun," "idle time" and latency indicators lawful, and said of the stow indicator that, given the "contraintes inhérentes à l'activité logistique" (the constraints inherent in logistics work, my translation), it asked no more than Amazon could reasonably expect of its employees. But it upheld the finding that keeping every indicator for 31 days broke the data-minimization rule, since Amazon hadn't shown it needed each one that long; nor had Amazon contested findings that it failed to tell temporary workers about its HR privacy policy.
To be clear, the French case was about productivity monitoring, not union forecasting, so it says nothing about whether Atlas would be lawful there, and the court sided with Amazon on the indicators themselves. But look at the process: a regulator walked in, examined how the numbers were built and kept, and a court weighed each indicator on the record. In the U.S., workers learned about Atlas because a Canadian labor case forced the documents out.
Imagine the version with no name
What follows is invention, built from parts that already exist.
Picture a big employer, not Amazon, a few years from now. Its feedback app has a friendly name; the forecasting layer underneath has none you've heard. Far-fetched? Business Insider reported in 2020 that Amazon-owned Whole Foods ran a heat map "designed to identify stores at risk of unionization," fed partly by internal surveys. (Whole Foods said it "recognizes the rights of our Team Members to decide whether union representation is right for them.")
Say it works the way Amazon's report describes. One hot week the board fills with posts about the heat, and a trend model flags the phrase before any manager has read a post. The building's score slips a band. Then twenty of you sign a letter asking for fans: a group complaint the NLRB counts as concerted activity, and an "Associate List of Demands" the report treats as high severity once it "theoretically" involves "more than 15 associates." If it climbs to the top tier, a response team deploys "within 24 hours." Your names never come up. They don't have to.
Would that be illegal? Possibly, if someone filed a charge and the Board took up the question. Would you even know? Under today's federal rules, probably not.
Where labor, business and the right come down
Is this a scandal or a survey? The lines don't fall quite where you'd expect.
A business coalition including TechNet and the California Grocers Association told legislators that AB 1883's protected-activity clause was "just not practical as written." Yet the management-side guide quoted above also says flatly: "Monitoring employees or their union-related discussions or activities is not allowed."
In a literature review for the U.S. Labor Department (the views are the authors' own), Penn State's Matthew Fischer-Daly and Mark Anner note that employer-run "non-union employee representation" has been found "self-defeating if motivated by union avoidance," and that surveyed workers "were less likely to quit than non-surveyed coworkers. However, it is unclear whether such surveys contributed to any substantive changes at work."
On the right, the think tank American Compass wants labor law amended to allow works councils and similar bodies "while protecting against employer domination of such organizations." It doesn't mention Amazon, but the principle travels: a voice the employer controls isn't quite a voice.
On labor's side, UC Berkeley Labor Center researchers Annette Bernhardt, Lisa Kresge and Reem Suleiman argue that employers should not use digital technologies "to identify, monitor, or punish workers for organizing" (a proposal, not current law). Unifor's organizing director, Justin Gniposky, told the Globe the documents show Amazon "views unionization as a threat and a risk to its business," while saying the union knows of no monitoring of Canadian employees without their knowledge or consent. Britain's GMB union warned that Amazon "could be deploying Orwellian AI-fuelled tools to monitor workers right here in the UK."
Strip out the adjectives and most of these voices draw the same line — between listening and watching — and these pages sit right on top of it.
What does this mean for you?
Talking with coworkers is protected. The NLRB says you have the right to "act with co-workers to address work-related issues," including pay and working conditions, and a group complaint to management counts.
Mind the deadline. If you think your rights were violated, you can file a charge, but quickly: no complaint can issue over anything "more than six months prior" to the charge. As of October 24, 2026, the Board has four of its five seats filled, enough to decide cases.
Ask, in writing, where the daily question and the board go. Amazon says Connections answers are aggregated and shared with managers at the team level. Ask whether survey or board data feeds any labor-risk forecast, who sees it and how long it's kept. (Retention is where France's top court ruled against Amazon.)
Know that declining counts too, at the site level, as "Rather Not Answer rates." Neither answering nor declining is invisible in the aggregate.
In California, use your right to know. The CCPA covers California residents who are employees or job applicants, with a right to know what personal information businesses have collected about you and how they use and share it. (It may not reveal a site score; Amazon says Atlas uses no personal data.)
How organizers handle it. Unifor's organizers, Gniposky told the Globe, contact Amazon workers only through their private phones and email addresses.
Tell your members of Congress whether you want S. 4831 and H.R. 9402 to move. As of October 24, 2026, both sit in committee; both die when the 119th Congress ends on January 3, 2027.
The lesson, as I see it
A listening tool and an early-warning system can be the same machine, and Amazon may well be right that this one never tracked a single person's opinion. That defense answers a privacy question, not an organizing one — organizing happens between people, at exactly the level Amazon says Atlas works.
So the question for any employer isn't whether it listens but what it listens for, and whether you're told. Federal labor law's nearest thing to a disclosure rule was a prosecutor's proposal, withdrawn after a little over two years.
My vote? Tell workers what their answers feed before they answer. If a daily question also feeds a labor-risk forecast, say so on the screen, next to the question. British Columbia already asks something like it of personal information: employers, Vancouver labor lawyer Menachem Freedman told the Globe, "need to tell employees beforehand what information they are collecting, why and how it will be used." Watch January 1, 2027, when California's new law takes effect without the clause that might have reached this, and January 3, when the federal bill dies unless someone moves it. A listening tool that can't survive that disclosure was never only listening.
Forward this to the coworker who answers every daily question, and to the one who declines them all. In Amazon's report, both choices count.