AI usage is climbing. Trust is falling. And almost nobody is treating this as the emergency it is.
ManpowerGroup's 2026 Global Talent Barometer found that regular AI usage among workers jumped 13% year-over-year — to 45% of the global workforce. In the same period, confidence in using the technology fell by 18%. Not a dip. A collapse running in the opposite direction from adoption.
This is not the story anyone expected to tell in 2026.
The inversion
The assumption was linear: deploy tools, people learn them, confidence grows, value compounds. The data says something else entirely. The more workers interact with AI, the less they trust their organisation's ability to govern it. Employee confidence in their company's AI strategy fell from 47% in 2025 to 31% in 2026 — a sixteen-point crater that opened while those same companies were scaling their AI investments.
Gallup's February 2026 study of 23,717 U.S. employees adds texture. Among workers who have AI tools available but choose not to use them, 46% say they prefer to keep doing their work the way they do it now. Four in ten cite ethical concerns, data privacy worries, or simply don't believe AI can help them. These aren't Luddites. They're people making a rational calculation: if nobody has explained how this fits my workflow, why would I risk it?
The training desert
The ManpowerGroup data names the mechanism. More than half the global workforce — 56% — reported receiving no recent training. 57% have no access to mentorship. Workers report feeling capable in the jobs they hold today but increasingly uncertain about what comes next. 43% fear automation may replace their job within two years.
Deploy without explaining. Scale without supporting. Then measure the "adoption rate" and call it progress. The trust inversion is not a mystery. It's the predictable result of treating people as the last variable in a technology equation.
The managerial multiplier
One Gallup finding cuts through: employees are 8.7 times more likely to view AI as transformative when their manager actively supports its use. Not when the CEO announces it. Not when the training deck lands in their inbox. When their direct manager — the person who understands their workflow — actively engages with how the tool fits their day.
This is a workflow-level signal, not an enterprise-level signal. The trust inversion cannot be solved at the town hall. It resolves one team at a time, one workflow at a time, in the space between the tool and the person who decides whether to open it.
The governance question nobody asked
Organisations spent 2025 asking: "How do we get people to use AI?" The trust inversion suggests they were asking the wrong question. The right question was always: "How do we make it safe enough that people want to?"
Not safe in the compliance sense — safe in the workflow sense. Does this tool come with clear rules about what it does and doesn't do? Does it arrive with a person (or a system) that explains how it fits here, not everywhere? Does the worker get to shape how it's used, or just receive it?
The Bentley-Gallup Business in Society survey found that only 31% of Americans trust businesses to use AI responsibly. Seven in ten don't. Even after climbing from 21% in 2023, the number remains stubbornly below a third — two years of AI investment have barely dented the trust deficit. The organisations deploying AI fastest are deploying it into a gap between adoption speed and earned confidence that keeps widening.
What this means
The trust inversion is not a comms problem. You cannot memo your way past it. It's a design problem — specifically, a workflow design problem. Somewhere between "we bought the tool" and "people trust the tool," there's a gap that only closes when someone designs the rules, the guardrails, and the governance at the level where work actually happens.
Nobody hired that person. That's the inversion.
Sources: ManpowerGroup Global Talent Barometer 2026 (January 2026); Gallup AI in the Workplace Study (February 2026, n=23,717); Bentley University-Gallup Business in Society Survey (2025, n=3,007).