Every retail and CPG board deck I have seen this year has the same slide. Agents in the workflow, and a reassuring box that says “human in the loop.” Everyone nods. Nobody asks the only question that matters: who is the human, how many hours do they have, and what did those hours cost.
That is the tell. Human in the loop has become a principle people invoke and a budget line nobody funds. If you did not cost the reviewer capacity, the loop is decorative. And decorative oversight is not a risk control. It is theater with a compliance label on it.
The gate is cheap. The judgment is not.
Here is the part that should reframe your whole business case. A June 2026 paper titled ‘Oversight Has a Capacity’ argues that the approval gate itself is trivial to build; as its author Emre Turan puts it, the gate is the easy part. The hard part, the expensive part, is the human doing the reviewing. And that human is not the infinitely available oracle your architecture diagram assumes.
Put a number on it. There are three costs, and only the first one ever makes it into a spreadsheet: reviewer labor. A $35-per-hour operations reviewer who spends two minutes per approval costs $1.17 per reviewed action.
That looks trivial. It is not. Take a returns or refund agent handling ten thousand flagged actions a day. At $1.17 each you have just added over $4 million a year in reviewer labor to a workflow you sold to the CFO as pure savings. If that number is not in your model, your model is fiction. You costed the agent and pretended the human was free.
Gate everything and you can buy worse outcomes
The instinct when a board gets nervous is to gate more. Route more to humans. Feel safer. This is exactly backwards, and the research now says so out loud.
Reviewers fatigue. Attention decays across a queue. Approval number 190 of the day does not get the scrutiny approval number 9 got. The consequence is not linear. Because reviewers fatigue as the queue grows, safety is an inverted-U in the escalation rate. Past a certain point, adding more human oversight makes a system less safe, not more.
Read that again from the seat you sit in. Gating everything does not just cost more. It can buy worse outcomes. When that happens, you enter the worst quadrant of the whole matrix: paying full price for oversight while receiving none.
So the answer is not “more humans” and it is not “no humans.” It is the right humans on the right actions, sized to a queue depth they can actually clear with attention intact. Do not gate by vibes. Gate by consequence. An action deserves a human checkpoint when it is irreversible, costly, regulated, or high-blast-radius, and especially when it is more than one of those at once. Everything else runs unattended, or a human watches the stream and steps in on exceptions.
How to size and fund it
Three moves. All of them are yours to own from the Head of AI seat.
First, make the workflow the unit of account, not the agent. The agentic version has one extra requirement: the unit of account has to be the workflow, agent plus oversight together, or the number lies. Every business case gets a second line under the agent labeled reviewer hours, fully loaded. No line, no approval.
Second, size the queue against real capacity. Take your escalation rate, multiply by volume, multiply by minutes per review, divide by the productive attention hours a reviewer actually has before fatigue sets in. That gives you required headcount and a hard ceiling on how much you can safely gate. If the math says you need forty reviewers and you funded four, you do not have a control. You have a backlog masquerading as one.
Third, name the owner. Someone signs for the escalation threshold and the queue depth. That person is accountable for the tradeoff between throughput and scrutiny. If nobody owns it, it defaults to nobody watching.
Now the Cost of Doing Nothing. Skipping this is not prudence, it is unpriced risk parked on the balance sheet. As of August 2026, any retail AI system qualifying as high risk, for example those controlling consumer credit, BNPL, pricing, or segmentation, must undergo conformity assessments, risk management, and human oversight, with significant penalties for non-compliance. An unfunded loop is not a defense when the regulator, or the incident, arrives. It is the exhibit.
Here is the punch. Every board that says human in the loop without funding the hours has already made a decision. They decided the risk is acceptable and just did not write it down. Write it down. Fund the loop or shrink the scope. The retailers who cost the human honestly will scale agents faster than the ones still treating oversight as a free principle, because their controls will survive contact with reality. The rest will discover their loop was decorative at precisely the moment it needed to hold.