The most common concern we hear in first conversations about AI is not the price and not the technology. It is loss of control: what if the thing sends something wrong to a customer? The concern is justified, because language models make mistakes, and they phrase them with the same confidence as their correct answers. The response, though, is not to skip AI. It is an architecture question: you build the automation so that the AI proposes and people decide, precisely at the points where it matters.

Why “human in the loop” is more than an OK button

The technical term for processes with human control is human in the loop, and it is often misread as: put a confirm button in front of every action. That is the worst implementation, because a button pressed a hundred times a day gets pressed blindly after two weeks. The control then exists only on paper.

A workable implementation asks a different question: which decisions actually need human judgment, and which do not? Sorting an email into a category does not; it can be corrected afterwards. Sending out a quote does. The craft lies in drawing that line deliberately instead of approving either everything or nothing.

How rarely that line gets drawn deliberately shows in McKinsey’s global State of AI survey (2025): only 27 percent of organizations using generative AI have people review all AI output before it is used. A roughly equal share reviews at most a fifth of it. The rest sit somewhere in between, often without anyone ever having decided.

The tool behind it: confidence thresholds

Language models can be set up to deliver an assessment of their own certainty alongside their results, or downstream checks can evaluate that certainty: do the extracted values fit together? Does the result contradict the master data? This creates the most important switch in the whole workflow: confident cases pass through, uncertain ones go to a person.

In fairness, one thing has to be added: a model’s self-assessment is not a measuring instrument. Models are sometimes convinced of wrong answers. That is why we always combine it with hard, programmed checks that are non-negotiable: totals must add up, references must exist, amounts above a limit always go to approval, no matter how confident the model feels.

Three tiers of control, matched to risk

In practice, a simple three-way split has proven itself, defined per action type. It looks like this:

Tier	What it covers	Examples	Human role
Automatic with logging	Low-risk, correctable steps	Categorizing, filing, enriching data	Spot checks and correction after the fact
Draft with approval	Anything that leaves the building or moves money	Quotes, customer replies, bookings	A named person reviews and approves
Always manual	Discretion or legal weight	Goodwill questions, personnel matters, contract changes	The human decides, the AI contributes groundwork at most

The underrated problem: approval fatigue

Whoever builds in approvals must also budget for their wear. If nine out of ten AI proposals are good, the person gets used to waving them through, and the tenth, flawed one slips through with them. That is not a character flaw of the employee, it is predictable behavior, and a good workflow plans against it.

Concretely: the review screen shows not only the result but its basis next to it, for example the original document beside the extracted values, so that checking is faster than trusting. Unusual cases get flagged rather than filed. And the volume has to be right: if a person is expected to clear four hundred approvals a day, the person is not the problem, the threshold is set wrong.

No log, no control

The second pillar next to approvals is traceability. Every step of the workflow belongs on record: what came in, what the model recognized, which checks ran, who approved when, what was executed. It sounds like bureaucracy, but it is the foundation for everything else. Without a log, an error cannot be traced back, a complaint cannot be resolved, and nothing can be demonstrated to an auditor or your tax advisor.

The log has a second use that often gets overlooked: it is the measuring instrument for the quality of the workflow itself. The corrections your staff make show where the model is systematically off, so improvements can be targeted instead of anecdotal.

Control is not a transitional phase

One might assume the approvals are training wheels, to be removed once the AI is good enough. For some steps that is true: if the logs show over months that a specific class of cases passes correctly practically every time, you can lower the threshold for exactly that class. That is a decision based on data, and it stays reversible.

For other steps it is never true. Whatever binds legally, moves money or touches customer relationships keeps its approval, regardless of how good the model gets. Not because the technology could not handle it, but because responsibility cannot be delegated. An AI cannot answer for a mistake; only the organization that runs it can. Automation you can control means exactly that in the end: you know at any time what it is doing, you can intervene, and the decisions that carry weight still carry a name.