Human in the Loop: Why SME Leaders Should Only Use AI When the Human Stays in the Decision
"Human in the loop" sounds good at every AI conference. In many companies it means, in practice, that somewhere a person clicks "OK". My position is stricter. A human is only in the loop if you can name, for each use case, the exact step where that person approves, and if the approval can actually change something. The audit disaster connects documented review and clear approvals to traceable evidence. For SMEs, knowledge validation makes that principle concrete by assigning a reviewer who can correct or reject a record before others rely on it. A manager who confirms suggestions day after day and never rejects one is automation with a signature.
The obvious objection is that approvals slow AI down and eat up the benefit. That is true if you put them everywhere. So an approval belongs only where a mistake hits a person or is hard to undo. Everywhere else, let the AI run.
Where a human has to approve
I use two questions. Does the output concern a specific person, such as an employee, an applicant or a customer? And is it hard to reverse because it ends up in a personnel file, a contract, a payment or an email to a customer?
- Yes to both: a human decides, the AI only supplies context. Example: an agreement from an employee review that goes into the personnel file.
- Yes to one: approval before it goes out, spot checks afterwards. Example: a summary sent to the project team.
- No to both: the AI works on its own and someone checks now and then. Example: internal research or a first draft.
One simple number tells you whether the approval is real. Over a fixed period, count how often reviewers change or reject a suggestion. If the rate is zero, you have either a perfect system or a review nobody takes seriously. I would bet on the second.
What the EU AI Act says
For high-risk systems, such as AI used in candidate selection or performance assessment, the AI Act requires human oversight. Article 14 obliges the provider to build the system so that people can understand it, discard an output and stop it. It explicitly names the risk of over-relying on the output (automation bias). Article 26(2) requires the deployer to assign oversight to people with the necessary competence, training and authority. For these systems this applies from 2 December 2027. Our article on the EU AI Act in practice covers what else an SME has to do until then. This is not legal advice.
How RECO builds it into the product
RECO processes employee conversations, so nearly everything it handles concerns a person. That is why we decided from the start where the manager makes the call.
- Preparing an employee review: RECO pulls together topics, earlier agreements and open items. The manager decides what to raise.
- During the conversation: discreet prompts for topics not yet covered. What gets said stays with the human.
- Follow-up: the AI proposes a summary, agreements and follow-up dates. The manager reviews, corrects and approves.
- Knowledge base: content is versioned, and for every recommendation you can trace which conversation or document it came from.
What leadership should do now
For every planned use case, write down who approves at which step, and place it using the two questions. Require every tool to show its sources. If it can't, it does not belong in processes that affect people. Train managers to check and add to AI suggestions instead of nodding them through, and record who approved what and why, so the decision still makes sense years from now. Measure the change rate regularly. It tells you more about your oversight than any policy.
Sources
Frequently asked questions
What is the difference between human in the loop and human on the loop?+
With human in the loop, a person approves each individual output before it takes effect. With human on the loop, the system runs on its own while a person monitors it and steps in when something looks wrong. In this article's terms that roughly matches the middle tier: approval where an output affects a person or is hard to undo, spot checks for the rest. For content that ends up in a personnel file, monitoring alone is not enough.
Related articles
In a 30 minute software demo we show you how RECO secures and transfers knowledge in your company. Request a demo now and we'll get back to you with a proposed time.
