EU AI Act Article 14: human oversight, applied to AI agents
What Article 14 of the EU AI Act asks of high-risk AI systems, when it applies after the Digital Omnibus, and where an approval step fits and falls short.
· 9 min read
This explains a regulation and how one tool relates to it. It is not legal advice, and using VariantGate does not make a system compliant. Whether the AI Act applies to you, and what it asks of you, is a question for your own counsel.
Article 14 of the EU AI Act says that high-risk AI systems must be built so that people can oversee them while they are in use, and stop them. For AI agents, which act rather than advise, the most direct form of that oversight is a person approving an action before it happens. Here is what the article requires, when it applies, and where an approval step fits and falls short.
First: is the agent high-risk?
Article 14 applies to high-risk AI systems, not to every AI system. Article 6 defines two routes:
- Article 6(1): a safety component of a product, or a product itself, covered by the EU product laws in Annex I and needing a third-party conformity assessment.
- Article 6(2) and Annex III: systems used in eight areas: biometrics; critical infrastructure; education and vocational training; employment and workers’ management; access to essential private and public services (such as creditworthiness); law enforcement; migration, asylum and border control; and the administration of justice and democratic processes.
Article 6(3) takes some Annex III systems back out, where they do not pose a significant risk of harm, for example when they perform a narrow procedural task. A coding agent or a customer-support bot is usually not high-risk. An agent that screens job applicants or decides who gets credit may well be.
What Article 14 asks
In the regulation’s own structure, paraphrased:
- 14(1)
- The system is designed, including with suitable human-machine interface tools, so natural persons can oversee it effectively while it is in use.
- 14(2)
- Oversight aims to prevent or minimise risks to health, safety or fundamental rights, including under reasonably foreseeable misuse.
- 14(3)
- Measures match the risks, the level of autonomy and the context. The provider builds them in, or identifies them for the deployer to put in place.
- 14(4)
- The people assigned oversight are enabled, as appropriate and proportionate, to do five things, listed below.
- 14(5)
- For remote biometric identification under Annex III point 1(a), no action is taken on an identification unless two competent people have separately verified it.
The five abilities in Article 14(4):
- (a)
- Understand the system's capacities and limitations, and monitor its operation, including to detect anomalies and unexpected performance.
- (b)
- Stay aware of automation bias: the tendency to rely on the system's output too much.
- (c)
- Correctly interpret the system's output.
- (d)
- Decide, in any particular situation, not to use the system, or to disregard, override or reverse its output.
- (e)
- Intervene, or interrupt the system through a 'stop' button or similar procedure that brings it to a halt in a safe state.
Two deployer duties sit next to it. Article 26(1) says deployers use the system according to its instructions for use. Article 26(2) says they assign human oversight to people with the necessary competence, training and authority, and the necessary support.
When it applies: decided and proposed
Decided. The AI Act originally applied its high-risk rules from 2 August 2026, and those for Article 6(1) systems from 2 August 2027 (Article 113). The Digital Omnibus on AI, Regulation (EU) 2026/1744 of 8 July 2026, moved them. It was published in the Official Journal on 24 July 2026 and entered into force on 27 July 2026. Chapter III, Sections 1 to 3, which hold Article 14 and Article 26, now apply:
- 2 December 2027
- for Annex III high-risk systems (Article 6(2))
- 2 August 2028
- for high-risk systems under Annex I product laws (Article 6(1))
The omnibus did not change the text of Article 14 or Article 26, or the Annex III list. The prohibited practices and AI literacy duties have applied since 2 February 2025, with some new prohibitions added by the omnibus from 2 December 2026.
Proposed, and not what was adopted. The Commission’s proposal of 19 November 2025, COM(2025) 836, would have tied the high-risk dates to the availability of standards. The adopted regulation uses the fixed dates above. A separate Digital Omnibus, COM(2025) 837, which deals with the GDPR, the cookie rules of the ePrivacy Directive, the Data Act and a single point for reporting incidents, was still before Parliament and Council in November 2026 and does not move these dates. Check the sources below for anything newer.
Where an approval step fits
An approval step does not make a system compliant. It can be one of the oversight measures Article 14(3) talks about, for an agent whose actions are its output. Point by point:
- (a) monitor
- Each request names the agent, the key it used, the action, the amount and the details. The record keeps every decision. Understanding the model's limits is still the provider's instructions for use.
- (b) automation bias
- Standing rules keep the questions rare, so each gets attention. Amounts are marked as stated by the agent. Two-person approval adds a second reader. No tool can make a person read.
- (c) interpret
- The summary is one plain sentence with the amount, recipient or count, written for the person deciding.
- (d) override
- Deny, or approve a narrower version with new wording or a lower amount. Silence and errors mean no. An action already taken cannot be reversed from here.
- (e) stop
- The emergency stop pauses every agent, or one, and nothing is allowed until an owner resumes. It stops only agents that ask before acting.
- Art. 26(2)
- Rules can send an action to the people with the authority to decide it. Their competence and training are yours to provide.
What it leaves to you
- Deciding whether the system is high-risk, and whether you are its provider, its deployer or both.
- The rest of Chapter III: risk management, data governance, technical documentation, the system’s own logs, accuracy and robustness, and the conformity assessment.
- Making sure the agent cannot act without asking. A check in the code that acts, not an instruction in a prompt, is what makes oversight more than a request to the model.
Article 26(6) asks deployers to keep the system’s automatic logs for at least six months. The record of decisions is not the system’s own log, but it shows who allowed each action and when, and it can be exported and checked for tampering. The emergency stop page covers the stop button in detail. The field to act on is always allowed.
Sources
- Regulation (EU) 2024/1689 (the AI Act): Articles 6, 14, 26 and 113, and Annex III. EUR-Lex.
- Regulation (EU) 2026/1744 (Digital Omnibus on AI), amending Article 113. EUR-Lex.
- Commission proposal COM(2025) 836, 19 November 2025. EUR-Lex.
- European Commission: AI Act, regulatory framework and timeline.
- European Parliament legislative train: Digital Omnibus on AI.