A distribution company set up its first AI agent to handle returns. It worked: it read the customer’s email, checked the order in the ERP, decided whether a credit note was due and issued it. Three happy weeks. Until a customer wrote \u00abyou sent me 400 units and I only ordered 40\u00bb and the agent, obediently, issued a \u20ac4,800 credit note for the customer’s own typo.<\/p>\n\n
Nobody reviewed it because nobody was supposed to. The agent was designed to execute, not to ask. And that is the question almost no company asks before automating: it is not whether the AI makes mistakes \u2014it does, like anyone\u2014 but what happens when it does and nobody is watching<\/strong>.<\/p>\n\n Human oversight of AI agents<\/strong> is exactly that: deciding where, when and on what criteria a person steps into the process. Designed well, it does not slow automation down. Designed badly \u2014or skipped entirely\u2014 it is the reason so many AI projects end up switched off.<\/p>\n\n\n Quick answer:<\/strong> Human-in-the-loop means placing approval gates only on an AI agent’s sensitive actions (payments, shipments, external communications). Everything else runs on its own. You supervise the 5% that matters, not 100%.<\/p>\n<\/div>\n\n\n Gartner predicts that over 40% of agentic AI projects will be cancelled before the end of 2027<\/strong> \u2014 and not because of technical limits: the causes are escalating costs, unclear business value and inadequate risk controls<\/em> (Gartner, 2025). In plain terms: the models are not what fail, process governance is.<\/p>\n\n What happens in practice is a pendulum. A company starts with an agent on a short leash, sees it get things right, and gradually lets it run end to end. Fine for months. Then the edge case arrives \u2014the one that was not in the tests\u2014 and the agent acts with exactly the same confidence it shows when it is right: it issues the credit note, sends the email, approves the invoice.<\/p>\n\n The damage is rarely the mistake itself; it is that the mistake is irreversible and silent<\/strong>. A miscoded ledger entry can be fixed. An email already sent to a customer cannot. Nor can a transfer. That is why the design question is not \u00abhow often is the agent right?\u00bb but \u00abcan this specific action be undone?\u00bb.<\/p>\n\n There is a further layer: since 2 August 2026<\/strong>, the EU AI Act’s obligations for high-risk systems apply, and its Article 14<\/strong> requires effective human oversight \u2014 whoever oversees the system must be able to understand its limitations, interpret its output, disregard or reverse it, and stop the system through a stop button or equivalent procedure. If your automation touches recruitment, credit decisions or worker evaluation, this is no longer best practice: it is a legal requirement. We cover it in our EU AI Act guide for SMBs<\/a>.<\/p>\n\n There is no binary \u00absupervise or don’t\u00bb. There are three degrees, and a single process usually needs all three at different points.<\/p>\n\n The classic mistake is applying one single level to the whole process. Put human approval on all 40 steps and you have built a bottleneck \u2014 people will end up bulk-approving without looking, which is worse than no oversight because it creates a false sense of control. Put it on none, and you get the \u20ac4,800 credit note.<\/p>\n\n This is the method we apply when deploying agents in companies that already have their systems in place:<\/p>\n\n It is worth separating this from AI data governance<\/a>, which answers \u00abwhat information can the agent see\u00bb. Here we are dealing with the other half: what we let it do<\/strong>. Two distinct controls, and you need both.<\/p>\n\n The goal is not \u00abless automation with more control\u00bb, but more automation precisely because control exists:<\/p>\n\n Practical criteria to decide without overthinking it:<\/p>\n\n Only if you supervise everything. In a well-designed process, somewhere between 3% and 10% of runs reach a human; the rest go through on their own. The person shifts from doing 200 tasks to deciding on 8 \u2014 and those 8 are the ones that matter.<\/p>\n\n In-the-loop<\/em> means the agent stops and waits for approval before acting. On-the-loop<\/em> means it acts on its own while a person monitors the activity and can intervene or halt it. The first protects irreversible actions; the second scales with volume.<\/p>\n\n No. The effective human oversight obligation in Article 14 applies to systems classified as high-risk \u2014 CV screening, employee evaluation, credit scoring and biometrics, among others. An agent that classifies invoices or answers internal questions does not fall under it, although supervising it is still advisable.<\/p>\n\n By reducing the number of approvals and improving their context. If someone receives 80 approval requests a day, they will stop reading them: raise the thresholds. Every request should show what the agent proposes, on what data, and what happens if it is rejected.<\/p>\n\n It is the cheap part of the project: you are inserting decision and notification points into a flow you already have, not rebuilding it. It usually takes days, and it prevents the kind of incident that costs thousands of euros and a customer relationship.<\/p>\n\n An AI agent without oversight is not a more automated company: it is a company with more surface area for error and fewer people looking at it. The difference between automations that last and those that get switched off after six months almost never lies in the model \u2014 it lies in whether someone sat down and decided which actions the agent was allowed to take alone. If you already have agents running, or you are about to, that is the exercise still pending.<\/p>\n\n\n\nWhy full autonomy is the most expensive mistake<\/h2>\n\n
The three levels of human oversight for AI agents<\/h2>\n\n
\n
How to design approval gates, step by step<\/h2>\n\n
\n
What you gain when oversight is designed properly<\/h2>\n\n
\n
When does each level make sense?<\/h2>\n\n
\n
Frequently asked questions<\/h2>\n\n
Doesn’t human oversight cancel out the benefit of automating?<\/h3>\n
What is the difference between human-in-the-loop and human-on-the-loop?<\/h3>\n
Does the EU AI Act require me to supervise every agent?<\/h3>\n
How do I stop approval becoming a mindless \u00abOK\u00bb click?<\/h3>\n
How much does it cost to add this layer to an existing automation?<\/h3>\n