When AI Agents Go to Work, What Will Humans Do?

Agentic AI is a top priority for many organizations, with the potential to transform how work gets done. As companies explore and implement AI agents that can perform tasks, make decisions, and take actions with varying degrees of autonomy, an important question is emerging: What will humans actually do? Human-in-the-loop oversight has become a central element of AI governance, particularly as organizations move toward more autonomous systems. In my research, I’ve seen human oversight emerge repeatedly as a governance priority. But simply saying that a human will remain in the loop isn’t enough. What will that person actually be responsible for? Will they be making decisions, directing agents, or simply checking their work? And will they have the knowledge and authority to intervene when something goes wrong?

I wanted to explore how data and AI professionals themselves envision these responsibilities. As part of TDWI’s 2026 Teams and Salary Survey, I asked, “If you were a human-in-the-loop for an agentic application, what would you be doing?” Of the 205 survey respondents, 136 answered the question. Interestingly, approximately one-third of those who responded did not articulate a clear responsibility, including some who said they weren’t sure or didn’t know. Others described a range of activities, from reviewing and validating agent outputs to providing business context, making decisions, directing agents, and handling exceptions. Their responses offer an early look at how professional roles may evolve as organizations adopt agentic AI.

Four Emerging Human Roles in an Agentic Workplace

Four themes emerged from the responses. These are not mutually exclusive roles. In fact, many respondents anticipated taking on multiple responsibilities.

The reviewer and validator. One of the most prominent themes was checking the work performed by AI agents. Respondents described reviewing outputs for accuracy, identifying errors, validating recommendations, and ensuring that results met expectations. One respondent envisioned “Double-checking samples of and outliers from agent results. Looking for ways to improve the process.” But as agents take on increasingly complex tasks, will human reviewers have enough visibility into how results were produced to evaluate them effectively?

The context provider and decision-maker. Other respondents anticipated applying business knowledge, interpreting results, evaluating risks, and making decisions. One described the responsibility as “Reviewing and making calls on high value, high risk decisioning.” This role goes beyond checking whether an output is technically correct. It involves understanding business objectives, organizational constraints, and the consequences of decisions. Even when agents can recommend or execute actions, humans may need to determine whether those actions are appropriate.

The agent orchestrator. Some respondents envisioned designing, directing, or managing agentic workflows rather than performing the underlying tasks themselves. One described the shift as “Most of what I do today, but orchestrating agents instead of human developers.” These responsibilities could include defining tasks, establishing operating boundaries, coordinating agents, and determining when human intervention is necessary. This suggests a potential shift in the skills required of data and AI professionals.

The exception handler and problem solver. Respondents also anticipated stepping in when agents encounter problems or situations requiring human judgment. One described the role as handling “complex exceptions, verify high-risk decisions, and provide course corrections when the autonomous agent hits a roadblock.” In this scenario, agents perform much of the routine work, while humans address unusual, ambiguous, or high-risk situations. However, if professionals spend less time performing routine tasks themselves, how will they maintain the expertise needed to resolve difficult problems?

Will Human Roles Differ by Job Function?

There were also some interesting differences based on respondents’ current jobs. Some architects envisioned designing and orchestrating agentic systems, while analysts and BI practitioners often focused on validating outputs and ensuring data accuracy. Some executives and senior leaders emphasized approvals, governance, and high-risk decisions. Although the sample sizes are too small to draw firm conclusions about differences across job functions, the responses suggest that human-in-the-loop responsibilities may evolve differently depending on a professional’s existing expertise and accountability.

Human Oversight Needs to Be Designed, Too

Perhaps the most telling finding is that a substantial share of respondents could not clearly articulate what they would do as humans in the loop. This may reflect the early stage of agentic AI adoption, but it also raises questions about whether organizations are adequately considering how human responsibilities will change.

There is also the question of meaningful work. One respondent envisioned “Looking for another job where I could use more of my brain.” That response stood out to me. If humans are relegated to simply approving outputs or handling problems, what happens to their opportunities to develop expertise, exercise judgment, and contribute value?

As organizations design agentic systems, they need to design human roles with equal care. That means determining where human judgment is essential, when intervention should occur, what information people need to make informed decisions, and who is accountable for the outcomes.

Human-in-the-loop should not simply be a governance checkbox. The future of agentic AI is not just about what work agents can take over. It is also about what work we want humans to continue doing, how those responsibilities will evolve, and how we ensure people can perform them effectively.