The premise.
When we first drew this distinction, the test was simple. If the machine replaced the person on a task, that was automation. If the person stayed in the loop and the machine made them faster and better, that was augmentation.
The test has stopped working. In 2026 an AI agent can take a whole task, a research brief, a first model, a reconciliation, and return it finished. By the old test, that is automation. Yet the senior person who delegated the task has not been replaced. They are doing more of the work that only they can do. They have been augmented.
The distinction still decides how an AI program should be scoped, funded, and measured. What has to change is where the line is drawn.
What the evidence shows.
Delegation is rising. In the largest published series on how one widely used AI system is put to work, the share of use in which the system produces the work with little input from the person rose from 41 percent at the start of 2025 to 45 percent by the end of it, and for a time passed collaborative use. Where businesses build AI into their own systems, about three quarters of use already follows that pattern.
Delegation is easy to start and hard to scale. In one long-running global survey, four in ten large organizations now report scaling AI agents, up from about a quarter a year earlier. Among smaller organizations the share is about one in five, and it has not moved. In the same survey, fewer than four in ten respondents attribute any effect on enterprise earnings to AI.
The form of use may matter for people. Early research finds that in occupations where AI use leans toward automation, employment of early-career workers has declined or grown more slowly. Where the use is collaborative, no such pattern appears. The evidence is young, and it points one way.
Read together, these say that work is moving from collaboration to delegation whether or not leaders plan for it. The open question is whether that delegation is designed as automation or as augmentation.
The new line: who owns the outcome.
The old test asked whether a person was in the loop. A better test asks three questions of any piece of work an AI system performs. Who decides what good looks like? Who checks the result before it counts? Who answers for it afterwards?
If a named person does all three, the work is augmentation, however much of the task the system performed. If no one does, it is automation, and it should be held to the conditions automation needs: high volume, clear rules, and few exceptions.
That gives three operations, not two.
Automation. The system does the task, and the process answers for it. Invoice matching and routine document classification belong here.
Assisted augmentation. The person does the task with the system's help. A memo that took half a day takes thirty minutes, and four and a half hours of senior time go back to work that only a senior person can do.
Delegated augmentation. The system does the task. The person sets the brief, reviews the result, and owns the outcome. This is the newest of the three, and it is where most of the opportunity now sits.
The short form: delegate the task, keep the judgment.
What delegated augmentation demands.
i. The brief becomes the work. A senior person who cannot say what good looks like cannot delegate, to a person or to a system. More programs stall here than on the technology.
ii. Review becomes the constraint. An agent can produce ten drafts in the time a principal can properly review one. Output that nobody has checked is not leverage. It is risk. A program should be sized to how much the accountable person can review, not to how much the system can produce.
iii. Oversight is set by risk, not by habit. Three levels are now in common use: the person approves each step, the person is notified and can intervene, or the system acts alone. The last suits only work that is low in risk, easily reversed, and high in volume, which are the conditions of automation. The level should be chosen decision by decision, by what an error would cost. Anything that reaches a client, moves money, or carries legal weight stays at approval.
iv. The apprenticeship has to be rebuilt. Automation removed the repetitions through which people learned to handle exceptions. Delegation does the same to the first drafts through which juniors learn judgment, and one early analysis finds that young workers are being hired into the most exposed occupations at a slower rate than before. The finding is tentative, and it is the right thing to watch. A firm that hands every first draft to a system has to decide, deliberately, how its next senior people will learn to judge one.
What stays the same.
Scope by the cycle, not the task. Begin where senior judgment is being spent on work that does not need it.
Fund on leverage, not labor saving. The return is more work of higher quality from the same senior people. It does not appear in the headcount line.
Measure the leading indicators first. Cycle time, the floor it sets for quality, and where senior time is redeployed, reviewed at 30, 60, and 90 days. The financial result follows and belongs in the board update a year on.
Delegation adds a fourth measure: review load. Track the hours the accountable person spends checking delegated work, and how often that review changes the result. If review load rises faster than useful output, delegation has outrun its design.
Three ways it fails.
Augmentation measured as automation. The headcount line does not move, the program is judged a failure, and the gain in throughput and quality goes uncounted. This remains the most common failure.
Delegation mistaken for automation. The work is handed to a system with no named owner and no review. Quality falls quietly, and the first person to notice is a client.
Volume mistaken for value. An agent makes output cheap. Fifty documents are not an answer. The measure is the decision the work made possible.
Closing.
The first question in any AI program is still whether the work is automation or augmentation. In 2026 it needs a second question, asked of every piece of work that is delegated: who owns the outcome?
Name that person, give them the time to review, and set the level of oversight by what an error would cost. The system can now do the task. It cannot answer for it.
Sources. Anthropic, Economic Index reports of September 2025 and January 2026, and "Labor market impacts of AI" (March 2026). McKinsey, "The state of AI in 2026". Stanford Digital Economy Lab, "AI Economic Indicators: June 2026 Update". IDC, "Autonomy on a Leash: A Framework for Agentic AI Governance" (August 2026).