According to a representative survey run by Epoch AI, 20% of employed Americans hand at least one task previously performed by a person over to AI.

Why the framing of the question matters

What the survey measures is not "do you use AI". That question has returned high numbers for a long time and says very little — using a search engine can count as using AI.

The question here is narrower and sharper: are you giving AI a task a person used to do? That reframes the tool from an assistant into a substitution. The 20% is a figure measured against that definition.

The real finding is in the second sentence

The more striking part of the survey is not the delegation rate but how the output is handled: respondents generally accept what the AI produces with little or no editing.

That indicates one of two things. Either the output is genuinely good enough and needs no correction; or the effort required to correct it comes close to doing the work from scratch, so nobody does. Those are very different worlds, and the survey does not say which applies.

Where the review step went

When you hand a task to a colleague, a review layer arrives with it automatically: that person asks questions, objects, comes back when something does not make sense. When you hand it to a model, none of that happens. The model does not ask, does not object, and answers with the same assurance when it does not know.

For organisations that is a hard risk to measure. Productivity gains are countable — how many tasks, in how much time. The review that disappeared only becomes visible when an error escapes.

What organisations can do

The practical response is not to block the delegation but to set review according to the kind of task delegated:

  • Reversible work (draft text, summaries, idea lists) — review may be unnecessary; errors are cheap.
  • Outbound work (customer-facing writing, reports, proposals) — a second pair of eyes is essential.
  • Work that becomes a decision (analysis, recommendations, figures) — verifiable elements need checking one by one.

That is the real gap the survey exposes: the 20% substitution has already happened, but the organisational habit of making that distinction has not yet settled in.

The limits of the figure

The survey is self-reported, which introduces error in both directions. Some workers under-report using AI, particularly where they are unsure of their employer's stance. Others read the question differently because they do not count using a tool as "delegating a task".

The 20% also does not mean that 20% of work is being delegated; it is the share of workers delegating at least one task. What that amounts to as a proportion of total work does not follow from this survey.

Where it is heading

The direction, though, is not in doubt. Asked a year ago the figure would plainly have been lower, and asked in two years it can be expected to be higher. The point is less the number than whether organisations are living through this transition by plan or without noticing.

So far the picture looks closer to the latter: the delegation is happening on employees' own initiative, ahead of any policy. Policy that arrives afterwards usually arrives after an error has escaped, and usually in the form of prohibitions — when what is actually needed is not a ban but a framework stating how much review each kind of work requires.