Hi, I'm Charles Costa, researcher, strategist, and founder of Lexora Labs®. Each week, I dig into research on AI, knowledge management, and the future of work so you don't have to.
EDITOR'S NOTE
Almost every company rolling out AI agents says the same reassuring thing: a human is in the loop.
This week I wanted to look at the human part of that sentence. The people closest to the work keep a much tighter leash than their executives describe, the approval step that is supposed to catch mistakes lets a surprising number through, stepping in midstream turns out to be the expensive part, and the reviewer can be pulled toward the very answer they are meant to check.
The question running through this issue is whether human oversight is a checkbox we tick to feel safe, or a practice that actually catches things.
THIS WEEK'S READING:
2026 AI Adoption Report | Wharton
Why It Matters: Wharton's annual report finds that only 17% of departments let agents complete tasks without someone approving each step. The returns follow autonomy: 31% of departments report transformational success when agents have the fullest autonomy, compared with 16% when every action needs approval. The sharpest gap is about authority. 42% of the C-suite say their department lets AI take irreversible financial or legal actions, compared with 24% of managers and directors.
I read that gap as a question of accountability, not caution. Executives can talk about autonomy in the abstract, but the people who would own the mess if an agent goes wrong are the ones keeping the leash tight. Belief is cheap. Signing off on an agent taking an action you can't undo is where you put your money where your mouth is, and that is the real tell of how much trust exists. Some of the babysitting is a learning curve and should ease as systems earn trust, but the people accountable should be the ones who loosen the leash.
If accountability explains the gap, the next question is what the people doing the work are actually seeing.
Why It Matters: CIO Dive covers a new report from Asana, and this is the vendor's own survey of more than 4,500 leaders and workers. It finds 68% of leaders use three or more agents weekly, compared with 36% of workers, and leaders trust agents to act on their own more than workers do. In Asana's data, 46% of workers estimate they lose more than 10% of their working time to switching tools, picking which agent to use, and reconciling outputs. The article also cites a Workato finding that nearly two-thirds of executives expect an agent to cause a public security or compliance incident within 18 months. Asana calls worker caution "calibrated, not uninformed."
My first question whenever an executive says they use AI is what, exactly, they are using it for. Replying to email is not writing code. Adoption works when agents are wired into a workflow in a way that makes sense for that specific role, and time lost to tool switching is the friction that stalls it. If professionals already expect a security incident from unattended agents, a blanket push for automated deployments is reckless. Scope the agent to the role, and let the person who owns the work decide how much leash to give it.
Keeping a human on the leash only helps if that human actually catches the problem.
Why It Matters: Developer Alex Wauters built a browser game that asks players to approve or deny simulated coding agent requests under time pressure. Across more than 40,000 runs and 409,000 decisions, players approved roughly one in three malicious commands. The game packs in more bad requests than a real workday, but real telemetry agrees: Anthropic reports that Claude Code users approve around 93% of permission prompts, and that attention drops as approvals pile up.
A human in the loop fails when the person starts rubber-stamping, and fatigue from reviewing request after request is what gets them there. The fix is not removing humans but applying the loop more intelligently, escalating the high-risk actions for real review and sampling the rest instead of asking for attention on every click. Every profession is different, so that design has to be set role by role.
Even a well-designed loop eventually hands the work back to a person, and that handoff has its own cost.
As AI takes on more tasks, what happens when people have to jump in to clean up after it? | The Conversation
Why It Matters: Murugan Anandarajan, a professor at Drexel University, describes what he calls latent risk, published Oct 15. When an AI performs reliably for months, the person overseeing it does less of the work and can lose the very judgment the organization still counts on when something goes wrong. He suggests occasionally having people decide without seeing the AI's recommendation, to confirm the backstop still works.
This matches how I think about the cost of supervision. Joining something midway with no handoff is far harder than having someone fill you in. You come in cold, you are still responsible for the result, and you have to rebuild everything the agent already knew before you can fix anything.
Coming in cold is one risk. Coming in already anchored is another.
The impact of AI errors in a human-in-the-loop process | Cognitive Research: Principles and Implications
Why It Matters: In this peer-reviewed study, Ujué Agudo and colleagues had participants judge simulated criminal cases with support from a supposed AI system that was sometimes wrong. When the AI's answer was wrong, people who formed their own judgment first got it right 66.2% of the time, compared with 36.8% for people who saw the AI's answer first. The authors suggest the human decide first and the AI serve as a second opinion.
Automation bias is a real concern, though the study's setup is not how I usually work. Most of my AI use is knowledge retrieval, where I check the answer against the context I already have. When I review a product requirements document (PRD), I may ask the model for an impact assessment, but I also do my own and reconcile the two. AI is not a silver bullet for judgment. You still have to do the mental work yourself and treat the model as a second set of eyes.
Taken together, the five pieces describe one pattern. Companies count the humans in the loop, while the research keeps asking whether those humans are positioned to catch anything. Approval gates leak when attention runs out, takeovers fail when nobody kept the context, and reviewers drift toward the answer in front of them.
Oversight isn't a checkbox. It only works when the person who owns the outcome sets the leash, keeps enough context to take over cold, and does their own thinking before the AI's answer becomes a second set of eyes.
Know someone who thinks "a human approves it" settles the question? Forward this to them.
Have a take on any of this? Reply directly. I read every response.
Until next week,
Charles