The Approve Button Is Not Oversight
"Human in the loop" has become the default answer to "is this agent safe?" Researchers and regulators are now asking a harder question: does a click count as a decision?
Ask almost any vendor how its new agent stays safe and you'll hear the same answer. A human approves before anything important happens. It's reassuring, it's easy to explain, and it's quickly becoming the standard design.
It's also starting to look like a fiction. When an agent sends approvals faster than a person can think about them, the person stops deciding and starts clicking. The approval step still exists on the process map. The oversight doesn't.
That's a workflow design problem, and workflow practitioners are the ones who will be asked to build these approval steps. The question isn't whether to put a human in the loop. It's whether the human you put there could realistically say no.
This month's launches lean on the approve button
Look at how the latest agent launches describe safety.
Date | Launch | How it describes oversight |
Oct 27 | AI recommendations stay subject to human review and approval, with oversight at "critical decision points." | |
Oct 27 | SAP Joule agents (TechEd) | Central governance of agents and fixed, deterministic workflows for compliance-sensitive steps. |
Oct 21 | New procedures and learned guidance go live only after a person approves them. Each agent has defined limits on what it can do without approval. |
None of these is a bad design on its face. But notice what the announcements leave out: which decisions count as critical, who the reviewers are, how many approvals they'll see in a day, and what happens when they're too busy to look. Those details decide whether the approval step is oversight or decoration.
SAP's approach is the interesting outlier. Instead of leaning on a reviewer, it leans on structure: some steps simply run in a fixed order every time. Hold onto that. It's half of the answer.
Approving becomes clicking
In Forbes this month, Guney Yildiz put the problem in three words. Once approvals run into the hundreds per hour, "approving becomes clicking." The person becomes a signature, not a decision-maker. And when the agent gets it wrong, the liability lands on that signature.
Regulators are noticing. Colorado's draft AI rules, which had a public hearing on October 26, would give consumers a right to meaningful human review of adverse decisions. More importantly, they set a test for when a company can say a person made the decision: that person needs real authority to decide, based on their own independent review and analysis.
Read that test as a workflow designer. A reviewer who sees only the agent's conclusion, has seconds to respond and has never overruled the system in practice would struggle to pass it. The approval step is on the diagram. The authority and the independent review are not.
Oversight degrades the overseer
The research says it's worse than a volume problem. In AI Agents Push Humans Out of the Loop, Margaret Mitchell and Avijit Ghosh of Hugging Face and Samir Passi of Data & Society argue that the act of overseeing an agent wears down the skills oversight depends on: awareness of what's happening, vigilance, judgment and domain expertise.
They trace this back to Lisanne Bainbridge's 1983 "ironies of automation": the more capable the system, the less prepared the human is to step in at the moment it matters. They describe several ways this plays out with agents:
Approval fatigue. People asked to get work done, grant permissions and judge every step stop scrutinizing what they approve.
Shortcuts. Reviewers treat an agent's stated plan as proof of what it did, or a passing test as proof the work is right.
Automation bias. Under load, people accept suggestions, especially fluent and confident ones.
A bad feedback loop. If quick approvals feed back into how the system is tuned, tired reviewers end up rewarding the behavior that weakens their own oversight.
It's a position paper, and the authors are open that some evidence is suggestive. But the direction is hard to argue with. An approval step you add today becomes a weaker control every month it runs unchanged.
When a review point is real
The fix isn't more approval steps. It's fewer, better ones. A review point is real only if the reviewer could realistically decide otherwise. Before you put one in a workflow, test it against six questions:
Time. At expected volume, how many seconds does the reviewer get per item? If the answer is under a minute for a consequential decision, it's a click.
Information. Does the reviewer see the inputs and the evidence, or only the agent's conclusion? Colorado's test asks for independent review, not agreement.
Authority. Can the reviewer reject, change or escalate without asking permission? Is anyone measured on how fast they approve?
Independence. Is the reviewer someone who benefits from approving, like the person whose queue the agent is clearing? The paper recommends separating the two.
Evidence of judgment. What's the override rate? If no one has rejected anything in three months, you don't know whether the reviewer is deciding or nodding. Seed known-bad items to find out.
Skill upkeep. Does the reviewer still do some of the work unassisted? If not, their ability to spot errors is quietly eroding.
A step that fails most of these isn't oversight. Take it out, or redesign it until it passes. Keeping it in gives everyone false comfort and puts the liability on the person clicking.
Limits upstream for everything else
Most agent actions will never get a real reviewer, and that's fine. They need a different kind of control: limits set in advance, built into the workflow, that don't depend on anyone paying attention in the moment.
The Forbes piece points to energy trading desks as the model. One human decision can redirect a cargo worth tens of millions. So experienced desks keep the final step human and put hard limits beneath it. The agent can act fast, but only inside a box someone drew deliberately.
Kind of decision | Control that works |
High-stakes, can wait | A real review point that passes the six tests above |
Fast or high-volume | Caps set in advance: spend limits, record limits, rate limits |
Must follow rules exactly | A fixed sequence the agent can't reorder, as in SAP's approach |
Anything that can run away | A stop switch someone owns and has tested |
This is explicit delegation in practice. Deciding what an agent may do alone, up to what limit, and where it must stop is a design decision made once, carefully. It's far more reliable than hoping someone catches the problem on approval number 340 of the day.
The real signal
"Human in the loop" made sense when the loop moved at human speed. Agents changed the speed, and the phrase didn't change with them. Researchers, journalists and now regulators are pointing out the gap.
The practitioners who get this right will design fewer approval steps, make each one count and draw firm limits around everything else. Everyone else will keep adding approve buttons, and keep calling it oversight.
Sources
The Human In The Loop Is Where AI Agents Park The Liability, Forbes, Oct 20, 2026
AI Agents Push Humans Out of the Loop, Mitchell, Ghosh & Passi, arXiv, Aug 2026
Dayforce Introduces Dayforce Deploy to Bring Agentic AI to HCM Implementations, GlobeNewswire, Oct 27, 2026
SAP advances its autonomous enterprise push with smarter, governed Joule agents, SiliconANGLE, Oct 27, 2026
Kognitos Launches AI Employees and expands Context Graphs for Finance and Banking, GlobeNewswire, Oct 21, 2026


Comments