انسان در حلقه نیست؛ دروازه تصمیم است

When some teams talk about AI in operations, they quickly reach a reassuring sentence: “Don’t worry, someone will approve it at the end.” It sounds responsible, but it often hides a large ambiguity. Who approves? What exactly do they see? What are they allowed to change? If they reject it, where does the workflow resume? And what happens if nobody responds in time?

If those answers are unclear, human approval is not a control; it is a decorative button. A button that people either click just to get through quickly, or one that leaves work stuck in an ownerless queue. In both cases, the intelligent system has not become more reliable. It has simply gained an ambiguous point where, when something goes wrong, everyone can shift responsibility elsewhere.

I prefer to speak of a “decision gate” rather than the common phrase “human in the loop.” A person is not supposed to sit beside the model all day and watch its outputs. The human role should be defined at the moment when a decision carries real cost, risk, or consequence. This small shift in wording creates a major shift in system design.

A good intelligent workflow does not send everything to a human for approval from the outset. That only destroys the speed of automation and turns people into operators of repetitive approvals. On the other hand, letting a system execute every action freely is not a sign of maturity either. It usually means the cost of error has not been taken seriously.

The central question is this: for every action, we should know in advance which of four states it belongs to. Can the system perform it automatically? Should it only make a recommendation? Must it obtain approval before execution? Or should it not be allowed to perform the action at all? This classification must be part of the architecture, not an improvised decision made after seeing a model output.

For example, summarizing an internal report, initially classifying requests, or finding inconsistencies across several files can be automated in many situations. But changing an operational record, issuing a purchase order, revising a production schedule, recording a quality-control result, or issuing an instruction related to equipment belongs to a different category. Here, the issue is not merely whether the generated text is correct; it is the real-world effect of the decision.

In industrial environments, this boundary becomes even more serious. A system may use sensor data, maintenance reports, quality records, or inventory status to conclude that an action is needed. Its real value is not in merely generating an alert. Its value becomes clear when it turns that recommendation into a controllable action: what action, for what reason, at what confidence level, based on which data, and with what possible consequence.

The approval gate should sit immediately before the point where a side effect is created. Not too early, when the decision-maker faces incomplete information; and not too late, when the system has already performed part of the action and genuine reversal is no longer possible. If an order is to be placed, approval must happen before placement. If an operating parameter is to change, control must happen before the change is applied, not after an undesirable result appears.

But pausing the workflow alone is not enough. The person who must decide should not face a vague message such as, “Do you approve?” They need an evidence package: the system’s recommendation, the reason behind it, the key supporting data, constraints and assumptions, the risk level, an alternative option, and the likely result of each choice. Good approval does not ask people to guess; it enables them to exercise judgment.

This evidence package should fit the decision-maker’s role. A procurement manager does not need to see every technical detail of the data, but they should know how the recommendation affects cost, delivery time, inventory, and commitments. A quality lead should be able to see samples, acceptance or rejection criteria, and nonconformities. An operator or maintenance lead should know which signals led to the proposed action and what safety constraints apply.

The decision owner must also be clear. A “team” does not make a decision; identifiable people in defined roles do. When an approval request is sent, the system must know who the primary owner is, who the backup is, how long the response window is, and where escalation goes if no response arrives. Sending a notification to a noisy group is not approval design; it is leaving real risk to chance.

Not every decision is binary either. Approval or rejection is too simple for many operational tasks. A good gate must also allow revision. The decision-maker may agree with the recommendation in principle but change the quantity, timing, supplier, priority, or scope of execution. If the system offers only two buttons, it forces the person either to approve an incomplete decision or to stop the entire workflow unnecessarily.

Rejection should not be the vague end of a process either. The system must know what rejection means: full cancellation, returning for more data collection, referring the case to another person, logging it as an exception, or requesting a new analysis under a new constraint. If the rejection reason is not recorded, it becomes a lost learning opportunity. If it is recorded but never reviewed, we have merely created another warehouse of data.

Another critical point is state. When a workflow pauses for approval, it must later resume from exactly that state, rather than having the model analyze the issue from scratch and possibly produce a different recommendation. Inputs, tool calls, intermediate outputs, active policies, and the decision identifier must be retained. This is not just a technical detail; it is a condition for accountability.

Imagine that a system has collected prices for an order, checked inventory, reviewed budget constraints, and produced a recommendation. If a manager approves it with one adjustment, the workflow should execute that adjusted recommendation. It should not regenerate everything after approval and create a new decision. Human approval must be attached to a specific state, not to a floating and untraceable conversation.

Every good approval also leaves a trail. Not to suffocate the organization with forms and bureaucracy, but so that later we can understand what happened. What recommendation was presented? What evidence was available? Who decided? What did they change? What was the result? Did the policies work as intended? Without that trail, evaluating the system depends on people’s memory and scattered messages across different tools.

Over time, this trail helps us improve policies. If one type of recommendation is almost always approved without changes, perhaps it is time to move it toward automated execution with appropriate controls. If a category of requests is repeatedly rejected, the problem may lie in the data, rule, or model. If decisions are consistently delayed at one point, perhaps the wrong owner was chosen or the evidence package is insufficient. Here, data is not only for the model; it is for improving the process.

To begin, you do not need to build a complex, theatrical architecture. Choose one specific process with operational value and limited, controllable side effects. Classify the actions. Write the stopping policy. Build an evidence package for the decision-maker. Define the paths for approval, rejection, revision, no response, and escalation. Retain state and the trail. Then review real cases to see whether the human can genuinely make a better decision or has merely become a rubber stamp.

AI becomes useful when it increases the speed and reach of analysis, not when it hides responsibility in fog. In an intelligent workflow, a human should not be a tired guard pressing an approval button every few minutes. They should own a decision gate: with clear authority, sufficient evidence, the right to revise and reject, and a system that continues safely, traceably, and evaluably after their decision.

If we do not design this gate properly, automation either makes mistakes faster or hides work faster behind approval queues. But if it is designed well, we can gradually transfer some trust from individuals to the process—not by removing people, but by placing them exactly where their judgment has real value.

A simple checklist for every intelligent action is this: What is the real-world effect of this action? Which policy permits it? Who owns the decision? What evidence is needed to decide? Are approval, rejection, and revision all possible? What happens if nobody responds? From which state does the system continue? And what trail will let us review this decision later? If you do not have answers to these questions, you have not designed approval yet; you have only added a button.

In practice, the difference between a showpiece assistant and a usable system often lies here. A model may be able to make a good recommendation. But an organization can rely on it only when it knows where to stop it, what to ask from it, and how to return to work smoothly and safely after a human decision.

This design may not be the most exciting part of an AI project. Attractive demos rarely talk about it. But in a business dealing with orders, inventory, quality, time, money, and safety, this is the part that determines whether your system is an intelligent toy or a dependable colleague.

Instead of asking, “Where should we add a human approval?” the more precise question is: “Which decision must pass through this gate, with what evidence, under whose authority, and with what path back?” The answer to that question defines the gap between noisy automation and an intelligent workflow.