As AI takes on more work, people need the authority and information to question its recommendations and correct mistakes.
Companies are investing billions to expand what AI can do. As those capabilities enter the workplace, leaders need to be equally deliberate about how employees will retain control over the decisions that affect them.
In September, AMD agreed to acquire World Labs, the AI company co-founded by Fei-Fei Li, for roughly $8.2 billion. Li is also a founding co-director of Stanford’s Human-Centered AI Institute, and her emphasis on human agency has shaped how I think about AI in the workplace.
AI can reduce administrative work and give people more time to use their strengths. But those benefits depend on how organizations redesign work. A system that saves a manager time while making it harder for an employee to challenge an inaccurate assessment can undermine the very people it’s meant to help.
Consider a hypothetical AI agent that flags an employee for missing project milestones. The employee was reassigned to another priority, but that change never reached the system the agent uses. If the flag feeds into a performance rating before anyone checks the context, the employee could be penalized for doing exactly what their manager asked.
Preventing that outcome requires clear ownership of the process, escalation to a person who can resolve the issue, and enough visibility for that person to make an informed decision. Each of those safeguards needs to be part of how the work happens.
Ownership requires the authority to change the process
An AI policy can name an accountable department, but that alone doesn’t give it control over how an agent behaves. Leaders need to decide who owns the business process, who manages the technology, and who can authorize changes when something goes wrong.
In the performance example, HR might own the review process while IT manages the agent’s configuration and access to data. They need shared goals and clear lanes of decision-making, including agreement on what the agent should accomplish and where its authority ends. It might be allowed to flag a potential concern, but it should not be able to change a rating or initiate an employment action on its own.
The person responsible for the review process also needs the authority to fix it. If the agent repeatedly misses changes in employees’ assignments, correcting one flag will leave the same problem waiting for the next employee. The process owner needs to work with IT and managers to change how updated assignments reach the system, and be able to suspend the agent’s use while the problem is resolved.
Changing a process can affect several departments, and the person leading that change needs explicit support to work across those boundaries. Without it, they can end up with the responsibility to solve the problem without the authority to make the necessary changes.
Escalation needs a clear trigger and a designated decision-maker
Automation design often centers on what an agent should do next. Governance also requires defining when it should stop. An escalation path identifies the conditions that require human judgment, the person responsible for providing it, and what happens while the decision is pending.
Many companies still have work to do here. In Deloitte’s 2026 State of AI in the Enterprise report, 74% of surveyed companies expected to use agentic AI at least moderately within two years, but only 21% reported having a mature governance model for autonomous agents. Companies need to define how people will supervise that work as they expand their use of agents.
For our performance-review example, a flag should trigger a review by the employee’s manager before it can influence a rating or employment decision. That review cannot depend entirely on the agent recognizing that something’s missing. In this case, the agent has no record of the reassignment and may have no reason to doubt its conclusion.
The manager needs to check the flagged milestones against the employee’s current responsibilities and give the employee an opportunity to supply missing context. If the manager is unavailable, the workflow should route the issue to an authorized backup or keep it on hold. It should never proceed to an employment decision simply because no one responded.
Adding AI can make a flawed process run faster without fixing its underlying problems. In our example, that means producing performance flags faster while still losing track of changes in employees’ assignments. We need to rethink the operating model around that work: the roles people play, the skills they need, and how decisions and accountability are shared. Technology should enable that model.
Oversight requires information people can act on
Giving someone the final say is only useful if they can evaluate the recommendation in front of them. A manager who sees a performance flag without the evidence behind it may end up approving a conclusion they have no way to assess.
Humans fail loudly, while AI fails silently. My concern is that the more we rely on AI to do our thinking, the less likely we are to question its output. An agent can produce a confident recommendation from incomplete information, and a person reviewing it can mistake that confidence for accuracy.
In the performance-review example, the reviewer needs to establish whether the employee delivered on their actual responsibilities, which may have changed since those milestones were set. They need access to the records the agent used, the criteria it applied, and an explanation of why it raised the flag. The employee’s explanation of their reassignment may supply the context those records lack.
The organization should then record how the issue was resolved, including any corrections and the reason for accepting or rejecting the recommendation. That record gives the process owner a way to spot recurring problems and determine whether the changes they make are working.
Employees also need to know how to challenge a finding if the first review misses something. Their ability to explain their work and correct an inaccurate record is a practical expression of human agency. Oversight should protect that ability as well as help managers make better decisions.
HR can start with one workflow
HR already shapes job design, how people’s capabilities are deployed, and how performance is evaluated. Those responsibilities give HR a practical starting point for designing how people and agents work together. IT builds the infrastructure, and business strategy sets the direction, but HR can help establish who does the work, who makes decisions, and how people are held accountable.
Choose one workflow and decide how it should work, including what people and AI will each be responsible for. Define the outcomes you want and bring together the people who can change the process and its technology. Then test the changes to roles, governance, and accountability alongside the agent itself in a small part of the business.
For a performance-review workflow, success might mean reducing the time managers spend gathering information while ensuring that employees are evaluated against their actual assignments. Test whether reviewers catch missing context, whether employees can get errors corrected, and whether the process owner can address recurring problems. Those results will tell you more than the number of flags the agent produces.
Use what you learn to improve the process before expanding it. As AI takes on more work, people must retain the power to question its conclusions and correct its mistakes.
About the author
Kelley Steven-Waiss is co-founder and CEO of Novaworks, an AI-native HR operating system built on ServiceNow to help organizations manage employees, contractors, and AI agents.
