20% of employed Americans report delegating at least one task to AI, but “delegate” needs unpacking
A July 10-19, 2026 survey conducted for Epoch AI by Ipsos found that 20% of employed U.S. adults say they have delegated at least one work task to AI that previously went to a coworker or outside contractor. That’s the headline. Picture a marketing manager pasting a weekly analytics brief into a generative model instead of emailing a contractor. Small moments like that add up fast.
Two important caveats up front. First, the data are self-reported: respondents described their own behavior rather than the survey logging actual usage or time saved. Second, the survey wording that produced the 20% figure matters: “delegated” can mean anything from asking AI for a draft once to routinely replacing an external supplier. The methodology appendix is the right place to verify how “delegate” was defined and which denominators were used.
The three states the survey captures, and the numbers
To interpret the findings, separate three operational states the report distinguishes (or implies):
- Assistance: AI helps with part of a task (human still responsible for most of the work).
- Near-autonomy: AI performs most or all of a task, with limited human intervention.
- Task substitution: A task that previously went to a coworker or contractor is now handled by AI instead.
Epoch AI’s survey reports adoption concentrated in structured, information-heavy tasks. Among respondents who perform the listed tasks, the share reporting AI use was:
- Computer systems and software development: 57%
- Data analysis: 46%
- Reading work documents: 39%
- Record-keeping: 25%
Most of that use is assistance rather than independence. The report frames AI as a “versatile but usually not self-sufficient workplace tool.” Full or near-full task completion by AI was notable in software development (about 10%) and below 7% for each of the other tasks studied. Measured task-substitution rates (respondents saying AI has taken over tasks people used to handle) were: data analysis 7.1%, reading work documents 5.7%, and record-keeping 5.3%.
Time savings and how people edit AI output
Productivity is the obvious ROI question. Respondents report mixed results:
- When AI handles part of a task, workers reported saving time on 37% of those tasks.
- When AI does most or all of a task, workers reported saving time on 53% of those tasks.
- About one in six AI-assisted tasks now takes longer than before.
How teams treat AI drafts matters for speed and risk. Epoch AI reports that 66% of AI output is used unchanged or with only minor tweaks; 6% of AI output is used verbatim with no edits at all; and 5% of AI output is heavily reworked or mostly rewritten. Those figures suggest many teams accept AI-generated work as a near-finished input, which lowers friction and raises the stakes for verification, IP control, and accuracy checks.
What leaders should take from this, concise, practical implications
The headline isn’t “AI replaced X jobs today.” It’s “AI is shifting where work happens.” Three implications matter for any executive sketching strategy now.
- Think task-first, not job-first. These numbers align with the task-based view of automation. Technology reassigns specific tasks faster than it eliminates whole occupations. Expect job descriptions to fragment into a mix of machine-executed tasks and human responsibilities like synthesis, judgment, and stakeholder management.
- Measure before you mandate. Self-reports are signal, not proof. Run pilots that pair usage logs with objective KPIs (time per task, rework hours, error/incident rates, and customer impact). Track who reviews AI output and how much follow-up work it creates.
- Governance is an operational priority, not an afterthought. Low editing rates speed workflows but amplify compliance and liability risks. Start with lightweight controls: a model inventory, approved-prompt templates for sensitive tasks, mandatory verification for external-facing outputs, and clear escalation paths when AI outputs appear suspicious.
A short playbook: audit, pilot, measure, and scale
Here’s a compact plan teams can act on this week.
- Two-week audit: Map the top 10 knowledge tasks across functions. For each, record who does the work, how long it takes (baseline), and whether AI is already used and how outputs are reviewed.
- 30-90 day pilot: Pick one or two high-structure tasks (data analysis, code scaffolding, document summarization). Randomize workstreams or use control groups where possible. Primary metrics: time per task, edit rate on AI outputs, rework hours, and downstream customer or compliance impact. Review weekly.
- KPIs to track continually: percentage of task-hours AI-assisted, percent of AI outputs accepted unchanged or with minor edits, error/incident rate tied to AI outputs, and hours spent on oversight/validation.
- Skill investment: Train reviewers in prompt best practices, evidence-checking, and contextual verification. Those verification skills will be the durable complement to AI assistance.
How this fits with other surveys and why metrics differ
Different surveys measure different things. Gallup reported in August 2025 that about 45% of U.S. employees used AI at least a few times a year, with roughly 10% using it daily, a frequency measure. By contrast, Epoch AI’s survey headline captures self-reported task substitution, which is a different slice of behavior. Vendor surveys add another perspective but are often non-representative: Anthropic’s survey of roughly 9, 700 Claude users found about half of respondents believed AI could already handle 50% or more of their work, reflecting a product-specific, power-user sample rather than the broader workforce.
Limits, open questions, and what to validate next
The Epoch AI/Ipsos numbers are valuable but incomplete without the full methodology. Before you read the 20% headline as evidence of broad structural change, check these items in the report or from Ipsos:
- Exact question wording for “delegated” and any definitions of “AI” provided to respondents.
- Whether task-level percentages are “of all respondents” or “of respondents who perform that task.”
- Sampling frame, weighting variables, margins of error, and subgroup breakdowns by industry, role, and frequency of use.
Other open questions that matter for planning: how often do those 20% delegate (one-off vs. routine), which specific tools are in use (public models, enterprise copilots, internal systems), and whether task substitution translates into headcount change or role redesign over time. Longitudinal data, or usage logs paired with business outcomes, will be needed to answer those questions rigorously.
“AI is a versatile but usually not self-sufficient workplace tool.”, Epoch AI
Questions leaders are actually asking, short, honest answers
-
Is AI replacing coworkers or contractors right now?
Not wholesale. Twenty percent of employed Americans report delegating at least one task to AI that used to go to a human, but most AI use in the survey is partial assistance and full task automation is uncommon outside software development.
-
Does AI reliably save time?
Sometimes. Respondents reported time savings on 37% of tasks when AI handled part of the work and on 53% when AI did most or all of the task, but roughly one in six AI-assisted tasks took longer than before.
-
Can I trust AI output as-is?
Often teams accept it with little editing, Epoch AI reports 66% of AI outputs are used unchanged or with minor tweaks, but that acceptance increases the need for verification on regulated, external, or high-stakes outputs.
-
Should I worry about mass layoffs?
The survey points to task redistribution rather than immediate, large-scale job loss. Still, companies should prepare for role changes and reskilling as task composition shifts.
Practical next step for leaders: run the two-week audit, launch one focused 30-90 day pilot with clear KPIs, and create a lightweight governance checklist. The work isn’t whether AI arrives, it already has, but whether your organization measures its impact, manages the risks, and organizes people to capture the gains.