One in Five American Workers Now Hand Tasks Directly to AI, New Poll Finds

A new nationally representative poll finds that one in five employed Americans now hands off work tasks directly to AI systems rather than merely using them as a productivity assistant, a shift that survey authors frame as a meaningful crossing point in how automation actually lands inside real jobs. The Epoch AI/Ipsos August 2026 poll, fielded July 10–19 among 1,103 employed U.S. adults screened from a broader probability sample of 1,747, is one of the most granular looks yet at the mechanics of workplace AI adoption: not just whether people use it, but what they hand over, who pays for it, and how much they still trust it enough to skip checking its work.

The topline number is stark on its own. But the more revealing data sits underneath it, in the gap between which tasks get full delegation versus partial assistance, in the split between employer-funded and self-funded AI access, and in how much human review still stands between an AI output and its use in the real world. Taken together, the numbers describe a workforce quietly renegotiating its relationship with AI tools, task by task, without waiting for employers or policymakers to catch up.

From "AI as Tool" to "AI as Delegate"

The distinction the survey draws matters more than it might first appear: using AI to help write an email is a different act than deciding AI should own the email entirely. Most coverage of workplace AI adoption to date has measured usage in the aggregate — the share of workers who have tried ChatGPT, Copilot, or Gemini at least once. That framing has always undersold the real question, which is not whether AI touches a task but who is accountable for the outcome once it does.

By that measure, the poll's headline finding is the first hard, nationally representative evidence that delegation, not just assistance, has become common enough to measure reliably. One in five employed Americans says AI now handles at least one task that used to go to a colleague or contractor. That is not a hypothetical about future automation — it is a present-tense description of how paychecks are already being earned.

Which Tasks Get Handed Off First

The survey breaks delegation down by activity, and the pattern tracks closely with which tasks are the most structured and text-based to begin with. Among the general working population, the tasks most frequently reallocated to AI were:

• Analyzing data — reallocated to AI by 7.1% of all employed respondents
• Reading work documents — reallocated by 5.7%
• Maintaining records — reallocated by 5.3%

Those figures describe delegation across the entire workforce, including people who never perform the task at all. Narrowing to just the workers who actually do each task in their jobs produces a sharper picture of where AI has actually broken through:

Task Type	AI Adoption Among Workers Doing This Task	Full or Near-Full Reallocation
Designing computer systems or software	57%	10%
Analyzing data	46%	Below 7%
Reading work documents	39%	Below 7%
Maintaining records	25%	Below 7%

Software and systems design stands out on both counts: it has both the highest overall AI adoption rate and the highest share of workers letting AI handle the task almost entirely. That aligns with what large language models are structurally best at — producing and iterating on code, a domain with clear syntax, fast feedback loops, and abundant training data. Data analysis and document-heavy work follow a similar but less advanced trajectory, while record maintenance, the most process-bound and system-dependent of the four, lags furthest behind.

Who Is Actually Paying for This

Perhaps the most operationally telling number in the release is how little of this adoption is being underwritten by employers. Among workers using AI at work, roughly half rely solely on personal subscriptions or free-tier access, while only about a third use a tool their employer has formally provided. In other words, a meaningful share of the workforce is not waiting for IT departments, procurement cycles, or corporate AI policies to catch up — they are expensing their own productivity gains, quietly, out of their own pockets or off the free tier.

Why the Employer Gap Matters

The payment source is not a trivial detail. Related Epoch AI/Ipsos research into the same worker population found that access type strongly predicts how central AI becomes to someone's job: workers on free tiers were far less likely to describe AI as core to their work than self-paying subscribers, and self-paying subscribers were in turn far less likely than workers with employer-provided tools to say they use AI as much for work as for personal tasks. Put plainly, the more an employer invests in giving a worker sanctioned AI access, the more that worker treats AI as a genuine part of the job rather than a personal side tool.

That gap creates a two-track adoption curve. One track is organization-led: companies that have issued licenses, set policy, and built AI into workflows, where usage is deeper and more integrated. The other is worker-led: employees at organizations that haven't formally adopted AI tools, who are nonetheless integrating them on their own initiative and their own dime. The second group is arguably the more important one to watch, because it represents demand that is running ahead of institutional readiness — exactly the kind of gap that tends to close fast once it becomes visible to management.

How Much Humans Still Check the Work

Delegating a task to AI is not the same as trusting it blindly, and the survey's editing data shows a workforce that remains cautious about output quality even as it hands off responsibility. Respondents were asked how much they typically revise AI-generated output before using it, and the results split roughly as follows:

1. Used with no changes at all: 6% of AI outputs
2. Used unchanged or with only minor edits (combined): 66% of AI outputs
3. Required more extensive revision: 27% of AI outputs
4. Majorly reworked or effectively redone: 5% of AI outputs

Read one way, that 66% figure looks like strong evidence of trust: two-thirds of the time, what AI produces is good enough to ship with light polish at most. Read another way, it means roughly a third of AI output still demands substantial human labor to fix, edit, or scrap — a meaningful tax on the productivity gains that AI vendors like to advertise. The survey also captured a related wrinkle on time savings: workers who let AI handle most or all of a task reported time savings 53% of the time, compared to 37% when AI only assisted with part of a task — but across all AI-assisted work, roughly one in six tasks actually took longer with AI involved than without it. Delegation, in other words, is not a guaranteed efficiency win; it is a bet that pays off more often than it doesn't, but still fails often enough to matter.

Which Job Categories Are Most Exposed

The task-level data doubles as a rough map of occupational exposure, even though the poll did not collect detailed breakdowns by job title. Reading the task categories against typical job functions suggests a clear hierarchy of who is feeling this shift first:

• Software engineers and systems designers sit at the leading edge, with the highest AI adoption and the highest rate of full task handoff — consistent with a wave of reporting on AI-assisted coding tools reshaping entry-level engineering work
• Data analysts, research staff, and business intelligence roles follow closely, given data analysis is both the single most-delegated task category and one where AI adoption already exceeds 45% among practitioners
• Administrative, legal support, and document-review roles show meaningful but slower-moving adoption, tracking the "reading work documents" and "maintaining records" categories
• Roles built around physical presence, hands-on service, or relationship management remain furthest from delegation, since none of the tasks AI has broken into at scale require a body in a room

That hierarchy roughly mirrors what other 2026 workforce studies have found using very different methodologies — AI reaching knowledge work first and fastest, with physical and interpersonal labor lagging. What this poll adds is a rare, direct measurement of the moment task ownership actually changes hands, rather than just a measurement of tool usage or sentiment.

The Bigger Picture

Twenty percent is not a majority, and it should not be read as evidence that AI has swallowed a fifth of the American labor market. It is evidence of something narrower and, in some ways, more significant: that direct task delegation — the specific behavior of routing work to AI instead of to a human colleague — has crossed from anecdote into a measurable, stable pattern across a nationally representative sample. A year ago, most surveys were still asking workers whether they had "tried" AI. This one asks what they've stopped doing themselves, and gets a clean answer from a fifth of the workforce.

The editing data suggests this shift is happening with eyes open rather than blind faith — most AI output still gets reviewed, and a meaningful share gets substantially reworked. But the payment data suggests the shift is outrunning institutional oversight, with half of workplace AI use happening on personal accounts employers haven't sanctioned, tracked, or secured. For any organization still treating AI adoption as a future planning exercise, this poll is a reminder that the planning window has already closed for a fifth of the workforce — the delegation has started, whether or not the policy has caught up.