In November 2022, Jake Moffatt’s grandmother died. He went to Air Canada’s website to ask about bereavement fares, and the airline’s support chatbot told him he could book a normal-price ticket and claim the discount afterward. That was false. Air Canada’s real policy, published on a page the chatbot itself linked to, does not allow a retroactive claim. The bot stated a policy that did not exist, with total confidence, and Moffatt believed it. Air Canada then argued in front of a tribunal that the chatbot was “a separate legal entity” responsible for its own words. In February 2024, the British Columbia Civil Resolution Tribunal rejected that argument outright and held the airline liable [1].

That case became famous because it reached a tribunal. Most versions of it never do. A customer hears a confident, wrong answer from an AI, believes it over the human standing in front of them, and the human, not the company that deployed the tool, is the one left holding the resulting anger.

That dynamic is small in any one interaction. It is also a preview of a shift that is not small at all.

The tasks that go first are the tasks that gave you room to breathe

The promise of workplace AI was straightforward: automate the tedious parts, free people for the parts that need a person. What is actually happening, according to a 2026 Ipsos poll conducted for the Groundwork Collaborative, is closer to the opposite of relief. The workforce it describes is “bracing for AI rather than embracing it,” and the divide is sharpest by income: more than one in five workers earning under $50,000 believe AI will make their jobs harder or less fulfilling, compared with just 5% of high earners [2].

Read alongside frontline accounts, that finding has a specific shape. When a system automates the easy, routine interactions first, in a call center it might be the account balance check; in retail, the price lookup; in a pharmacy, the routine refill, what is left for a human to handle is whatever the automation could not resolve. That is disproportionately the complicated case, the edge case, and the furious customer. The easy interactions used to be breathing room between the hard ones. Remove them, and a shift that used to alternate between simple and difficult becomes wall-to-wall difficult.

The Resilience Connection: This directly supports our Mental Resilience pillar. Grounded thinking under sustained pressure requires recognizing the structure that is producing the pressure, not just absorbing it as personal failure to cope.

The AI doesn’t just fail to help. Sometimes it actively makes the job harder

The Moffatt case above is not an isolated legal curiosity. It describes a specific and newer failure mode, one that plays out far more often outside a courtroom than inside one: an AI assistant states a company policy that is false, a customer believes it because the AI stated it with total confidence, and a human employee is then required to either violate real policy to satisfy a hallucination or absorb the anger of someone who believes they are being lied to. The employee did nothing wrong. The AI did. The employee pays for it anyway.

This is a distinct problem from “AI takes over easy tasks.” It is AI actively manufacturing new hard ones, then routing the emotional cost of its own error to the person least equipped to correct it in the moment.

The savings do not go to the worker

Here is the part that turns a hard week into a structural trap. When a company’s AI adoption genuinely does save time, that time is not reliably returned to the people doing the work.

In mid-2026, Meta employees asked whether AI-driven productivity gains could justify reviving “Meta Days,” a discontinued paid-time-off program. Meta’s CTO, Andrew Bosworth, answered plainly: “I hope that what we do with our extra time is do even more and cooler stuff for the users who use our products every day… I get an extra hour. You know what I do with it? I put it into that.” Reporting on the exchange found Meta employees clocking well beyond a standard 40-hour week, in some cases past 70, even as leadership framed AI as an efficiency win [3].

That is the shape of the trap. AI removes the easy tasks, which were the breathing room. It creates new hard tasks, correcting for its own errors. And where it genuinely saves time, that time is claimed by the organization as more output, not less strain. Three separate mechanisms, one result: the person doing the work ends every day more depleted, for the same pay, doing work that used to include some slack and now does not.

What agency looks like inside a trap you didn’t design

HRP exists on the premise that the answer to “AI will make this worse” is engagement, not resignation, and not blind optimism either. Three concrete things are within a frontline worker’s control here, even when the structure above is not.

Document the hallucination, specifically. When a customer cites an AI-stated policy that is wrong, that is not just a bad interaction to survive; it is evidence. Note what the customer said the AI told them, when, and what platform if known. A pattern of these reports is the concrete artifact that gets a manager’s or a company’s attention in a way that “customers have been rude lately” never does.

Name the doubled-output demand out loud, in writing. If a role’s expectations have quietly increased since a tool was introduced, that increase deserves to be stated as a fact, not absorbed as a personal shortfall. A sentence like “since this tool was introduced, my caseload has gone from X to Y” is something a manager or an HR file can act on. Silence is the one input that guarantees nothing changes.

Practical Takeaway: Keep a running, dated log of two things: AI-caused customer conflicts, and any measurable increase in workload since an AI tool was introduced. A log converts a feeling of being ground down into a record someone with authority can act on.

Refuse the isolation. The Groundwork Collaborative polling above exists because workers organized to ask the question at scale. What one person experiences as a personal bad week is, at scale, a documented and worsening pattern across an entire income bracket. Recognizing that difference, this is not just happening to me, is itself a form of resilience: it is the difference between carrying a structural problem alone and carrying it as part of something larger than any one shift.

None of this makes the shift easier tomorrow. It is the difference between absorbing a trap silently and building the record that eventually forces someone with the power to change it to look at it directly.

How this was made: This post was drafted by an AI model from a topic and direction chosen by Kenneth G. Hartman, checked against current sources, then reviewed and approved by him before publication. He is responsible for what it says. A project about staying human alongside these tools should be plain about using them.

References

[1] CBC News, “How can I mislead you? Air Canada found liable for chatbot’s bad advice on bereavement rates,” Feb. 2024. [Online]. Available: https://www.cbc.ca/news/canada/british-columbia/air-canada-chatbot-lawsuit-1.7116416

[2] Groundwork Collaborative, “AI at Work: Gains at the Top, Pressure for Everyone Else,” poll conducted by Ipsos, June 2026. [Online]. Available: https://groundworkcollaborative.org/work/ai-at-work-gains-at-the-top-pressure-for-everyone-else/. Also reported at Common Dreams.

[3] Futurism, “Meta Exec Rages Against Employees Asking for More Time Off Because AI Made Them More Efficient,” 2026. [Online]. Available: https://futurism.com/artificial-intelligence/meta-exec-rages-against-employees-more-time-off-ai