Twenty percent of employed US adults say AI now handles at least one task they or their team previously handed to a coworker or an outside contractor (Epoch AI, 2026). Not assisted with. Handled.
That is a sourcing decision โ the kind your organization normally wraps in a requisition, a vendor review, and a budget line. It was made by individual contributors, one task at a time, and it appears in no system you own. AI task delegation of this kind does not show up in headcount, in contractor spend forecasts, or in any adoption dashboard, because the dashboards were built to count tool usage, not work reallocation.
The uncomfortable part is not that it happened. It is that your capacity model is now wrong, and wrong in a direction you cannot see.
What the Epoch AI/Ipsos Survey Actually Measured
The survey was conducted by Ipsos on behalf of Epoch AI using the probability-based KnowledgePanel, fielded 10โ19 July 2026 among 1,106 employed US adults, with estimates weighted to the employed US population (Ipsos, 2026). Probability sampling matters here: this is not an opt-in panel of AI enthusiasts telling you how transformative their tools are.
The headline is the 20%. The texture is in the task-level breakdown. Workers reported AI taking over work previously delegated to humans most often in analyzing data (7.1%), reading work documents (5.7%), and maintaining records (5.3%) (Epoch AI, 2026).
Read that list again as an operator. Those are not creative tasks or edge experiments. They are the exact categories a 200-person company routinely pushes to a junior analyst, a temp, an offshore team, or a freelance researcher. They are also, almost without exception, tasks whose output someone else depends on downstream.
One more figure from the same survey, and it is the one that should hold your attention: across AI-assisted tasks, 66% of outputs were used unchanged or with only minor edits, against 5% that were majorly reworked or redone.
A Make-or-Buy Decision With No Requisition
Every organization has a process for deciding whether work gets done inside or bought outside. It is slow for a reason. Someone scopes the work, someone prices it, someone checks whether the vendor is any good, and someone owns the result.
What the Epoch data describes is that same decision, made at the individual level, in the time it takes to open a chat window.
The employee is not doing anything wrong. From where they sit, this is initiative: the report that used to take four days of back-and-forth with a contractor now takes an afternoon. They are being rewarded for exactly this. No policy told them a task reassignment from a human to a model is a different category of decision than choosing a faster way to do their own work.
But it is a different category, for one structural reason. When you delegate to a person, you get a second mind on the problem whether you asked for one or not. Contractors ask clarifying questions. Junior analysts flag data that looks wrong. Colleagues push back when a request does not make sense. That friction is not a defect of human delegation โ it is an unpriced quality control mechanism that came bundled with it.
Delegate the same task to a model and the friction disappears. The model does not ask why the request is strange. It answers.
The Review Gate Vanished at the Same Moment the Work Moved
Put the 20% and the 66% side by side and the risk becomes concrete. Work is moving from humans to models at the same time that two-thirds of model output ships with no meaningful revision. The task changed hands and the inspection step left with the previous owner.
This is not hypothetical damage. Adaptavist's 2026 study of 2,500 knowledge workers found 52% regularly correct AI-generated work produced by colleagues (Adaptavist, 2026). The correction is happening โ it has simply relocated from a defined handoff point to whoever is unlucky enough to be downstream.
Research from BetterUp Labs and Stanford Social Media Lab put a number on that relocation: roughly 40% of US desk workers reported receiving AI-generated "workslop" in the previous month, taking about two hours to resolve per incident, which the researchers costed at approximately $186 per worker per month (BetterUp Labs, 2026).
Two hours per incident, absorbed by a colleague, logged nowhere.
There is a distributional wrinkle worth naming. The three task categories where this shows up most โ analyzing data, reading documents, maintaining records โ are concentrated in exactly the functions that a mid-market operation tends to run thinnest: finance support, ops coordination, compliance administration. These are the roles where one person often is the process, and where a contractor was historically the pressure valve. When the pressure valve is replaced by a model and nobody re-specifies who reviews the output, the failure does not announce itself gradually. It announces itself as a single bad number in a board pack.
That is the shape of the problem. The savings are captured by the person who delegated to the model and are visible to them. The cost is paid by the person receiving the output and is invisible to everyone. Net productivity can be flat or negative while every individual involved honestly reports being faster.
Why Your Capacity Model Is Now Wrong in an Unrecorded Direction
Here is where this stops being a governance concern and becomes an operations problem.
If a meaningful share of your team has quietly rerouted tasks away from contractors and coworkers, several of your planning inputs are now measuring something other than what you think.
Falling contractor spend reads as discipline. It may be discipline. It may also be unreviewed work substitution. Those two have identical signatures in the P&L and opposite implications for risk.
Internal handoff volume drops. Fewer tickets, fewer requests between teams, faster cycle times on paper. Some of that is genuine. Some of it is work that stopped being handed off because it stopped being checked.
Headcount models drift. If you plan next year's hiring on this year's observed workload, and this year's observed workload has been silently absorbed by models with a 66% ship-unchanged rate, you are extrapolating from a baseline that already contains unmeasured quality debt.
At 50โ500 FTE this bites harder than it does at enterprise scale. A large organization has procurement gates, a vendor management function, and enough process instrumentation that a shift in how work gets sourced eventually surfaces in someone's report. A mid-market operation has a Head of Operations, a finance lead, and a set of spreadsheets built on the assumption that work is done by the people on the org chart.
Nobody is doing anything visibly wrong. That is exactly why it persists.
What This Evidence Does Not Say
Three limits, stated plainly, because a plan built on an overreading of this data will fail.
Twenty percent is not twenty percent of your workload. The figure is the share of workers reporting at least one such task, not the share of tasks or hours displaced. The per-task numbers โ 7.1%, 5.7%, 5.3% โ are the better guide to volume, and they are single digits.
Ship-unchanged is not the same as wrong. The 66% figure tells you review intensity dropped, not that the output is defective. For a routine record update, unchanged may be entirely correct. The concern is that the review rate is now uniform across task types where the error cost is very much not uniform.
Self-report has known limits. Workers may not accurately recall whether a task would previously have gone to a contractor. Probability sampling controls for who was asked, not for how well people remember counterfactuals.
None of that changes the operational conclusion. Even at the conservative end, work is being resourced by a mechanism your capacity planning does not observe, and the quality gate that used to be free is gone.
Instrumenting AI Task Delegation This Quarter
The fix is measurement discipline, not a policy memo. Four moves, none requiring new spend.
Ask the question directly, once. In your next team review, ask each function one question: which tasks did we stop sending to a contractor, a shared service, or another team in the last six months, and where did they go? You will get an answer in an afternoon. The Epoch data says roughly one in five of your people has one.
Instrument reallocation, not logins. Adoption metrics that count seats and sessions are structurally incapable of detecting this. Track the things that move when work is reassigned: contractor invoice volume by category, inter-team request counts, and rework incidents. Stanford's Digital Economy Lab has been making this same argument about measuring AI at the task level rather than the tool level (Stanford Digital Economy Lab, 2026).
Re-attach the review gate where peer delegation used to supply one. This is the highest-value item and the cheapest. Classify tasks by the cost of an undetected error, not by whether AI is involved. For the high-cost set, name a reviewer explicitly. Human delegation supplied that reviewer for free; model delegation does not, and it has to be reinstated deliberately.
Make task reassignment a declarable event. Not an approval process โ a disclosure. When someone moves a recurring task from a person to a model, it gets noted in the same place a headcount or vendor change would. One line, no ceremony. Visibility is the entire objective.
The decision in front of you this quarter is narrow and binary. Either you can name the AI task delegation that has already happened in your operation โ which work moved from a human to a model, and who checks it now โ or you cannot. If you cannot, you are not running a lean operation. You are running one whose true cost structure and quality exposure are being set, task by task, by people who were never told they were making that call.
