16 October 2026 EU AI Acthuman oversightrecurring work

Human oversight that is real

"A human reviews it" is not a gate. Name who checks, who can stop, and who they call.

Two colleagues review a draft at a desk; one marks the page with a red pen beside a computer.

"A human reviews it" is not a gate.

In a lot of companies, that sentence is the whole oversight plan. It sits in a policy. The work continues. The sentence does not name the person, the moment, what they may change, or how they stop a bad draft from leaving.

Pick one use that already happens. An assistant drafts a customer reply, and someone might send it. That is enough for this exercise. Do not start by classifying the tool.

Check, approve, escalate, stop

Check means a named person reads the draft before it goes out. They look at the facts the assistant used, and at the wording. "Someone on the team" is not a name.

Approve means that person can send the draft, or send it after a change they made. A nod in a meeting is not an approval if the draft already left.

Escalate means there is a second named person, and a reason to call them. A price, a complaint, a personal detail, or anything the first person is not allowed to decide. "Ask around" has no destination.

Stop means the first person can refuse to send the draft, and can turn that use off until a named owner has looked at what went wrong. A review that cannot stop the output is a comment.

Write those four lines for one use. One paragraph is enough.

The paragraph to write

Replace the brackets. Keep it to one use.

Before a customer reply drafted by [tool] is sent, [named person] checks the facts and the wording. They may send it, change it, or refuse it. If it includes [a price, a complaint, personal data, or another limit they cannot decide], they pass it to [named second person] and do not send it. If the draft is wrong in a way they cannot fix, they stop that use until [named owner] has looked at it.

That is the gate. The policy sentence is not.

What the paragraph does not do

The paragraph is an operating habit for ordinary recurring work. It does not fulfil Article 14. It does not fulfil Article 26(2). It does not classify a system.

Article 14 of Regulation (EU) 2024/1689 applies to high-risk AI systems under either Article 6 route. It does not apply to every workplace tool. The system has to be designed so a person can oversee it while it is in use, including the ability to disregard the output or interrupt it.

High-risk is a classification with two routes, not a label for Annex III alone.

Article 6(1) covers an AI system that is a product, or a safety component of a product, under the Union legislation in Annex I, when that product must undergo a third-party conformity assessment. Paragraphs 1a to 1c, added by Regulation (EU) 2026/1744, narrow that route. A system used only for non-safety user assistance, performance optimisation, service efficiency, automation, convenience, or quality control is not a safety component. A system whose failure or malfunction would endanger health and safety still is. A conformity assessment that exists only for risks other than health and safety, such as radio spectrum or electromagnetic interference, does not meet Article 6(1)(b).

Article 6(2) is separate. Systems referred to in Annex III are high-risk, unless Article 6(3) applies. That derogation covers a narrow procedural task, improving a human activity that is already finished, detecting decision patterns without replacing human review, or a preparatory task, and only where the system does not pose a significant risk of harm. The provider documents that assessment. An operator does not create it by adding a sentence under a customer reply. An Annex III system that profiles natural persons remains high-risk.

An ordinary workplace assistant that drafts a customer reply sits in neither route. It is not an Annex I safety component, and drafting a reply is not an Annex III use. A use in hiring, credit, education access, biometrics, or as a safety function in a regulated product may be high-risk. This article does not decide which one you have. If the use sits near those areas, stop the exercise and involve a qualified specialist. Do not decide from the vendor name.

The dates are not the general line "it applies from 2 August 2026." In the consolidated text of 27 July 2026, Chapter III, Sections 1, 2 and 3 apply from 2 December 2027 for high-risk systems under Article 6(2) and Annex III, and from 2 August 2028 for high-risk systems under Article 6(1) and Annex I. Article 14 is in Section 2. Article 26 is in Section 3. Article 26(2) says a deployer of a high-risk system assigns oversight to a natural person with the necessary competence, training, authority and support. A named reviewer, without that competence and authority, does not meet it. A paragraph written on 15 October 2026 does not meet it either. Those sections are not yet applicable.

"Human in the loop" does not close the gap. A person who cannot override or stop the output is not overseeing it.

This is practical orientation, not legal advice. It does not classify a system, meet Article 14, or meet Article 26(2). For a use that may be high-risk, involve qualified legal and technical specialists and read the current official text.

Earlier pieces in this series:

Is recurring work behind the question?

Tell us what keeps happening in the free fit check. We will review whether Groundlens is the right next step.

Check if it fits