Stanford puts 70 million U.S. workers within reach of AI agents: they want shared control, not surrender

By El Ecosistema Startup · September 14, 2026.

The Stanford SALT Lab published on arXiv the study "Future of Work with AI Agents: Auditing Automation and Augmentation Potential across the U.S. Workforce," which estimates that 70 million workers in the United States could see part of their tasks transformed by AI agents (systems capable of carrying out tasks autonomously, not just answering questions). To reach that figure, the team interviewed 1,500 professionals across 104 different occupations and assessed 844 specific tasks, drawing on O*NET, the US Department of Labor's occupational database.

The methodology, dubbed WORKBank, cross-references two variables per task: whether the worker wants an agent to automate it, augment it or keep it under human control, and whether technical experts believe current AI can perform it reliably. That cross-referencing produces four zones: the "green light" zone (capability and desire aligned, immediate adoption), the "red light" zone (automatable but rejected by the worker), the R&D opportunity zone (desired but still out of technical reach) and the low-priority zone. According to the study, 46.1% of the 844 tasks received approval for automation, and the dominant motive is not fear of losing one's job: 69.4% of the favorable responses cite freeing up time, followed by repetitiveness and quality improvement.

The most interesting finding, however, is another one. The study introduces the Human Agency Scale (HAS), a five-level scale running from "the agent does everything" (H1) to "the human does everything" (H5). The middle level, H3 —balanced collaboration between person and agent— was the preferred option in 47 of the 104 occupations analyzed. And in 47.5% of the 844 tasks, workers asked for more human involvement than the technical experts themselves considered necessary. In other words: the technology could already go further than people are willing to let it, according to the data gathered by the researchers.

As sector context, the study arrives alongside a regulatory shift already under way in the United States. Stanford HAI published in August 2026 the report "Designing Loyalty: AI Agents and Conflicts of Interest," which proposes imposing a fiduciary duty on agent developers —requiring them to prioritize the user's interest over the commercial one—, as reported by Forkast News. The Federal Trade Commission (FTC) issued on July 1, 2026 a proposed policy against deceptive steering (when an agent nudges the user toward an option that suits the company more than the user), and the SEC included AI-linked conflicts of interest among its supervisory priorities for 2026.

Our reading: this study contributes something missing from the debate on AI and employment, which usually moves in headlines of the "so many million jobs at risk" variety. Here the unit of analysis is not the job but the task, and that changes the conversation at its root. It confirms what we have been arguing in this space: AI does not replace entire occupations in one stroke, it redefines which tasks within each occupation still deserve a human. The very finding that information-processing tasks —analysis, knowledge updating— are delegated sooner than coordination, human interaction or quality assessment fits the pattern we have seen sector by sector: the administrative and repetitive gives ground first; judgment, relationships and discernment hold out longer.

The majority preference for H3 —co-piloting, not handing over the wheel— is, in our view, the study's finding with the most staying power. It says that resistance to full automation is not just fear of layoffs: it is a legitimate preference for retaining agency over one's own work, even when the technology could already do it alone. That is good news in the short term, because it suggests the transition can be designed with the user inside the process rather than replaced by it, but it should not be read as the problem disappearing: the 47.5% gap between what people want and what technical experts already consider feasible is precisely the space where the next labor battle will be fought, and not every sector or every worker will have the same bargaining power to defend their H3.

Over the long term, the finding points in the direction we defend: if the abundance AI promises materializes, the division of tasks between person and agent will stop being a matter of economic survival and become a real choice about which part of the work each person wants to keep. The fiduciary duty proposed by Stanford HAI and the interest already declared by the FTC and the SEC in agents' conflicts of interest are signs that this transition will not be left to the market alone: legislation is beginning before the gap between technical capability and social trust turns into a crisis, which is exactly the kind of evidence-based governance this analysis has been calling for.