Brandon Sammut, Chief People and AI Transformation Officer, Zapier, says human oversight in AI systems is often positioned too late in the decision-making process, limiting accountability and reducing the effectiveness of review.
There is a very real concern about AI in the workplace focused on replacement. Which roles will disappear? Which tasks will be automated away? Yet a more immediate problem is already affecting organizations deploying AI today.
Many companies have added checkpoints and approval steps to AI-driven workflows. On paper, humans remain in the loop. In practice, however, those humans are often positioned at the end of a process, after the reasoning has run, the data has been shaped and the decision has effectively been made. At scale, this creates routine approvals rather than meaningful oversight.
Getting human-in-the-loop (HITL) AI right means fixing where people intervene in a workflow and what they can actually see when they get there. That shift can dramatically improve reliability and confidence when scaling AI systems.
The most common structural error in HITL design is late placement. Human review happens after the AI agent has already committed to a reasoning path, leaving teams with a choice between approving and disrupting. In reality, it is rarely a genuine choice. By the time a recommendation reaches review, it has already shaped downstream expectations and challenging it means absorbing the cost of delay or rework. Most teams approve and move on.
For consequential decisions, intervention must happen earlier, before the AI agent’s direction is set. In a workflow routing high-value sales opportunities, for example, the checkpoint belongs where routing criteria are defined, not when a ranked list appears in someone’s queue. In a contract review workflow, an HITL checkpoint should occur when the agent is configured to identify clause types, rather than after it produces a marked-up draft. This shifts reviewers from signing off on outputs to shaping how the agent approaches work.
Visibility is equally important. Even when HITL placement is correct, reviewers need access to the reasoning behind conclusions. A reviewer positioned at the right point in a workflow but handed only an output is still being asked to support a decision they cannot properly evaluate.
Bridging that gap starts at the platform level. Effective AI platforms provide governed access layers that record what the agent accessed and the permissions it used, surfacing that context alongside every output. Reviewers can then intervene with a complete picture rather than relying on assumptions.
Organizations must also assign review rights to people who own outcomes. Too often, approvals are based on workflow position rather than accountability. The individual next in line may lack the authority, context or experience to challenge a recommendation. Fewer reviewers making informed decisions generally produce better outcomes than broad oversight spread across every output.
The strongest HITL implementations begin by identifying where human judgment adds irreplaceable value. That principle must be embedded in AI infrastructure from the start through governance, auditability and clear controls. As AI agents connect across more tools and systems, oversight must travel with them. Bolt-on accountability does not scale. Infrastructure-level accountability does.
