Geyto Research · Mariano D'Angelo · Essay · · 2 min read
AI agents. What changes in adoption when the tool acts
An AI agent does not only suggest; it carries out steps of the work. What it demands of people, and what research says about delegating to automation.
In brief: the first wave of generative AI suggested: it drafted, summarized, proposed. AI agents act: they query data, complete steps and trigger processes. That changes the adoption question. It is no longer only about using a tool, but about delegating to it and supervising it. Research on automation has decades of evidence on what that demands of people.
From suggesting to acting
Agents are already entering organizations. In McKinsey's 2026 survey, 40% of organizations with more than a billion dollars in revenue say they are scaling AI agents, compared with 22% of smaller ones. In Grant Thornton's (2026), nearly three in four organizations give them access to their data and processes.
But organizations still delegate cautiously. In that same survey, only 5% let agents make high-stakes decisions without human review, and 60% limit them to moderate-risk tasks. The caution makes sense: what an agent demands is not the same as what an assistant demands.
What we already know about delegating to automation
In 1997, Raja Parasuraman and Victor Riley described four ways of relating to automation: use, misuse (relying on it too much), disuse (relying on it too little) and abuse (automating without thinking about the person who has to supervise). All four appear today with agents.
John Lee and Katrina See showed that what matters is not how much trust there is, but whether it is calibrated: whether it matches what the system actually does well. And Lisanne Bainbridge warned in 1983 that the more is automated, the harder it becomes to step in on the exceptional case, because the person supervising loses practice in exactly that.
What changes in the demand
With an agent, what the change asks of people moves:
- From doing to assigning. Someone has to decide which tasks are delegated, with which limits and which information.
- From executing to reviewing. The work comes to include checking what the agent did, sometimes without having seen the intermediate steps.
- From routine to exception. People step in on the cases the agent does not resolve, which may be the hardest ones.
- From task to accountability. Someone has to answer for the result, and that has to be defined beforehand.
Each of those shifts calls for situated capacity: judgement, coordination and practice where the work happens.
Both sides can adjust
The answer is not only to prepare people. The initiative can also be adjusted: start with moderate-risk tasks, keep human review where the cost of an error is high, widen the agent's scope as the organization builds the capacity to supervise it. The caution the data shows can be read that way: as a reasonable adjustment of the demand to the capacity available.
What to do about it
- Define for each agent what it decides, what it proposes and what it executes only with approval.
- Keep people in touch with the task so they can step in on exceptions.
- Assign owners for the agent's results before widening its scope.
- Review regularly whether trust in the agent is still calibrated to what it does well.
More in Trusting the system, in AI makes the question sharper and in Both sides can change.
Geyto Research publishes essays, evidence reviews and market notes on AI and technology adoption in organizations.
Frequently asked questions
What is an AI agent?
An AI system that does not only answer or suggest, but carries out steps of a task within defined limits: it queries data, fills in forms, sends messages or triggers processes.
Do companies already let agents decide on their own?
Rarely. In Grant Thornton's 2026 survey, nearly three in four organizations give AI agents access to their data and processes, but only 5% let them make high-stakes decisions without human review, and 60% limit them to moderate-risk tasks.
What changes in adoption with agents?
The demand shifts from using a tool to delegating and supervising: deciding what the agent is given, reviewing what it did, stepping in on exceptions and defining who is accountable for the result.
Sources
- Grant Thornton (2026). 2026 AI Impact Survey Report. Survey of 950 business leaders, February–March 2026.
- McKinsey & Company (2026). The state of AI. Survey of 1,719 participants in 97 countries, May–June 2026.
- Parasuraman, R. and Riley, V. (1997). Humans and automation: Use, misuse, disuse, abuse. Human Factors, 39(2), 230–253.
- Lee, J. D. and See, K. A. (2004). Trust in automation: Designing for appropriate reliance. Human Factors, 46(1), 50–80.
- Bainbridge, L. (1983). Ironies of automation. Automatica, 19(6), 775–779.
Share