Human in the loop, at the helm or in the lead? The AI oversight stack

As agentic AI moves from generating content to making decisions and taking action, boards and C-suite leaders are looking beyond human in the loop as their AI oversight model. They are adopting approaches that put people clearly in charge of AI, rather than simply part of its processes.

In 2024, Salesforce’s Paula Goldman championed human at the helm. In January 2026, Accenture Chair and CEO Julie Sweet raised human in the lead at Davos. Rather than competing ideas, the various approaches represent an evolution where each serves a distinct purpose, together forming an “oversight stack.” In the loop, people review AI’s processes and outputs. At the helm, people steer the system. In the lead, people own the outcomes and decide where the loop and the helm belong.

Human in the loop: Reviewing outputs

“The loop” suited early enterprise AI, which produced drafts for people to edit. As agents make decisions and take action, the loop is under strain in three ways:

• Human review cannot keep pace and cognitive overload can occur. Agentic AI often operates with concurrent systems that act in milliseconds and complete numerous actions before a reviewer can open a notification. When review timelines reflect AI speed rather than decision complexity, cognitive overload can follow
• Oversight drifts. Even where volumes are manageable, automation bias can lead people to accept a system’s recommendation despite contrary evidence. Approval becomes a click or silence. Process metrics can reinforce the pattern when they track completed reviews, turnaround times and sign-off rates rather than decision quality
• Efficiency can create long-term risks. An AI first philosophy asks teams to show why a task needs a person before adding one. Reducing human review can be a sound decision when the organization can see what its AI is doing, intervene when needed and stand behind the results. Longer-term risks arise when savings and speed are prioritized over accountability, or when organizations lose the expertise needed to recognize errors. Like talent debt, this creates “oversight debt,” a future cost not reflected on today’s balance sheet

Human at the helm: Steering the system

Human at the helm moves people from reviewing individual outputs to steering the system itself. People set controls, parameters and accountability to direct outcomes.

Much of this work happens in the AI harness, the software surrounding a model that manages tools, memory, permissions, safeguards and audit logs. Two organizations using the same model can have very different levels of risk depending on harness design.

The helm is also limited. Controls left unchecked can become less effective, and even a well-designed harness needs an accountable owner. Without one, organizations risk what researcher Madeleine Clare Elish called the moral crumple zone, where the person nearest an automated system bears the impact of failures, while those responsible for the broader design choices face less scrutiny. Board attention can follow a similar pattern. In WTW’s 2026 Directors’ and Officers’ Survey, 50% of respondents cited AI errors and misinformation as a primary concern, while only 32% cited weak governance or uncontrolled AI use.

Human in the lead: Owning the outcome

Human in the lead adds what the other two models assume but do not guarantee: ownership. Leaders define each AI system’s purpose, remain accountable for its results and decide where human judgment applies, whether in the loop, at the helm or both.

7 key practices

Organizations adopting the human in the lead oversight model tend to focus on seven practices that span the oversight stack:

In the loop

1. Protection against cognitive overload: Reviewers have manageable workloads and deadlines, the support and tools they need and dedicated time to focus. Workloads are monitored to prevent burnout
2. Review as supervision rather than gatekeeping: Audit trails, sampling and exception review replace preapproval of everything, making systemic issues easier to identify

At the helm

3. Boundaries built into the harness: Prohibited actions, protected data, spending limits and escalation triggers operate at machine speed, with ownership extending across the full technology stack
4. AI agents managed like a team of employees: Agents have clear objectives and boundaries, defined issues for escalation and a person accountable for their results

In the lead

5. Clear ownership for every agent: Each system has a named leader who is accountable for its mandate and outcomes.
6. Retained human expertise and the ability to intervene: Organizations remain capable of evaluating AI output, even as AI improves efficiency; they maintain and test ways to pause or stop every agent
7. Support for users as authors: Training and performance expectations help people stand behind what they produce with AI and question it when needed. WTW’s employee experience research suggests that confidence grows when people understand their role in AI-enabled work

Where the loop still belongs

The newer models don’t eliminate the loop. It still belongs in high-stakes or irreversible decisions, including those affecting people’s rights, livelihoods or safety. In such cases, people need enough time and information to decide. Article 14 of the EU AI Act formalizes this for high-risk systems. The difference is that leaders decide where the loop belongs, rather than applying it everywhere.

Board oversight of AI grows

Board attention to AI governance is increasing. Yet in WTW’s 2026 Directors’ and Officers’ Survey, only 55% of respondents said their boards have the skills and knowledge to oversee AI effectively. The report also noted that enforcement and securities class actions are already creating consequences for directors who do not sufficiently oversee AI.

Boards working to close that gap are asking practical questions: Who owns this agent’s outcomes? Could we pause it today? Can we recognize when it’s wrong? What capability would we need if this system failed?

This space previously covered the National Association of Corporate Directors’ (NACD) updated AI governance model, outlined by Dylan Sandlin, Dr. Helmuth Ludwig and Dr. Benjamin van Giffen. Its four elements align with this oversight stack: strategic oversight defines AI’s purpose; capital allocation weighs savings against oversight debt; AI risk is a factor in setting the boundaries of the loop and the helm; and technology competence helps leaders recognize when AI is wrong.

A line widely attributed to a 1979 IBM training presentation still holds: “A computer can never be held accountable, therefore a computer must never make a management decision.” Nearly 50 years later, the principle carries new weight. “Human in the lead” completes the oversight stack by keeping people accountable for AI at the level where they can exercise that responsibility effectively.

A version of this article originally appeared on Forbes on October 16, 2026.