What does Article 14 of the EU AI Act require?
Article 14 says high-risk AI systems "shall be designed and developed in such a way, including with appropriate human-machine interface tools, that they can be effectively overseen by natural persons during the period in which they are in use." Oversight measures must be proportionate to the risks, the level of autonomy and the context of use.
The practical core is paragraph 4. The system must be provided to the deployer so that the people assigned to oversight are enabled, as appropriate and proportionate, to do five things. The table restates each point in plain words and shows what it means when the system is an AI agent. Read the official text (Regulation (EU) 2024/1689) before relying on any summary, including this one.
This guide is general information for teams designing agent workflows. It is not legal advice, and it does not say whether your system is high-risk. Ask counsel for that.
| Point | The overseer must be able to | For an AI agent, that looks like |
|---|---|---|
| 14(4)(a) | Understand the system's capacities and limitations and monitor it, detecting anomalies | A readable process: which steps the agent runs, what it may change, a live view of where each run is |
| 14(4)(b) | Stay aware of automation bias | Approvals that show evidence (tests, sources, diffs), not a bare "approve?" prompt; few enough stops that people still read them |
| 14(4)(c) | Correctly interpret the output | Outputs tied to the step and inputs that produced them, with the reason recorded |
| 14(4)(d) | Decide not to use, disregard, override or reverse the output | A gate before consequential actions where the person can decline, and a way to send work back |
| 14(4)(e) | Intervene or interrupt the system, through a stop button or similar procedure, to a safe state | Stop or cancel a run, or have it wait at a known step, without leaving half-applied changes |
sourcesRegulation (EU) 2024/1689, official text (EUR-Lex)EU AI Act, Article 14 (human oversight)
Does the EU AI Act apply to my AI agents?
Only some of it, and it depends on use, not on the word "agent". The Act is risk-based. Article 14 attaches to high-risk AI systems: those listed in Annex III (areas such as employment, access to essential services like credit, education, law enforcement and border control) and AI that is a safety component of products covered by the Union legislation in Annex I. An agent that drafts blog posts is very unlikely to be high-risk; an agent that screens job applicants may well be.
Roles matter too. Providers develop the system or place it on the market; deployers use it under their authority. Article 14 is mainly a design duty for providers; Article 26 gives deployers duties to use the system according to its instructions, to assign oversight to competent people and to keep logs. A company that builds an internal agent for its own high-risk use can carry both roles.
Even when your agents are not high-risk, Article 14 is a good engineering checklist. The same five abilities are what any team needs before it lets agents act on customer data, money or production systems.
sourcesEU AI Act, Annex III (high-risk areas)EU AI Act, Article 6 (classification of high-risk systems)
When do the human oversight rules apply?
Later than first planned. The Act entered into force on 1 August 2024 and applies in stages. On 29 June 2026 the Council gave final approval to the Digital Omnibus on AI, which moved the high-risk dates: 2 December 2027 for stand-alone high-risk systems (Annex III) and 2 August 2028 for high-risk AI embedded in products (Annex I), according to the Council press release of that day. Check the Official Journal and the Commission's pages for the consolidated text before you plan against these dates.
| Date | What applies | Source |
|---|---|---|
| 1 August 2024 | The AI Act enters into force | Regulation (EU) 2024/1689, Article 113 |
| 2 February 2025 | Prohibited practices and general provisions apply | Article 113 |
| 2 August 2025 | Obligations for general-purpose AI models apply | Article 113 |
| 2 December 2026 | Transparency marking for AI-generated content (grace period shortened by the omnibus) | Council press release, 29 June 2026 |
| 2 December 2027 | High-risk obligations for stand-alone systems in Annex III, including Article 14 | Council press release, 29 June 2026 |
| 2 August 2028 | High-risk obligations for AI embedded in Annex I products | Council press release, 29 June 2026 |
sourcesCouncil of the EU press release, 29 June 2026EU AI Act, Article 113 (entry into force and application)
What must deployers of high-risk AI do for oversight?
Article 26 turns oversight into named duties for the organisation using the system. Four of them shape how you run agents day to day.
| Paragraph | Duty | What to prepare |
|---|---|---|
| 26(1) | Use the system in accordance with its instructions for use, with technical and organisational measures | The process the agent follows, written down and versioned, not living in a prompt |
| 26(2) | Assign oversight to people with the competence, training, authority and support they need | A named owner for each gate, and the authority to say no |
| 26(5) | Monitor operation and inform the provider or authorities about risks and serious incidents | A view of running and stuck work, and an incident path |
| 26(6) | Keep automatically generated logs under your control for an appropriate period, at least six months unless other law provides otherwise | A retention decision and a place the logs actually live |
How do you design human oversight into AI agent workflows?
Turn each ability in Article 14(4) into structure the system enforces, then collect the evidence as the work runs. Prompted promises ("ask before you send") do not count: a model can skip them, and nothing records the miss.
- 1
Write the process as steps
Each step has its own instructions and a clear output. Overseers can read what the agent does before it does it (14(4)(a)).
- 2
Put gates before consequential actions
A stop the system evaluates, not the model, before anything is sent, published, paid or deleted. The overseer can decline there (14(4)(d)).
- 3
Show evidence at the gate
The approval request carries the diff, the test results, the sources. Fewer, better-informed approvals fight automation bias (14(4)(b) and (c)).
- 4
Make stopping cheap and safe
Runs must be stoppable at a known step without half-applied changes. Prefer processes where consequential writes happen only after the gate (14(4)(e)).
- 5
Name the overseers
Each gate has an owner with authority and training (Article 26(2)). Record who they are outside the tool, too.
- 6
Log with two identities
Record the automation and the verified person separately, with the step and the reason (Articles 12 and 26(6)). See the audit trail guide for the field list.
in the prompt read by the AI
Shipped. Nobody said yes.
as a gate checked by the engine
Waiting for you.
Where does ConvOps help, and where does it stop?
ConvOps is a workflow engine your AI client connects to over MCP. It supports several oversight practices directly. It does not make any system compliant, it does not classify your risk, and it does not replace your provider's or your own legal assessment.
| Practice | What ConvOps does | What you still own |
|---|---|---|
| Readable process (14(4)(a)) | Workflows stored as steps with instructions; the agent receives one step at a time; any session can see where a run is | Writing instructions a person can actually review |
| Stop before acting (14(4)(d)) | Approval gates the engine evaluates on every advance; the run does not move until approval is passed | Deciding where gates go; approval is a flag on the advance call, so restrict who can make it in your process |
| Interrupt (14(4)(e)) | Runs wait at gates; runs can be cancelled; ConvOps-dispatched executions can be stopped | Designing steps so a stop leaves a safe state |
| Named overseers (26(2)) | Three roles (owner, admin, member) per workspace | Training, authority, and naming overseers per gate |
| Logs (12, 26(6)) | Field-level audit on tasks and workflow instances, with the agent label and the server-verified person; step history with reasons | Retention, export and tamper evidence, which ConvOps does not provide today |
| Isolation | One workspace per system, isolated with Postgres row-level security | Data protection obligations under other law |
What should you do this quarter?
Even with the high-risk dates moved, the work is the same and it takes time. A short list:
- Inventory your agents and the decisions they touch; flag anything near an Annex III area for legal review
- For each flagged agent, write the process down as steps and mark the consequential actions
- Put a system-evaluated stop before each consequential action and name its owner
- Check where the human identity in your logs comes from, and fix any path where the caller supplies it
- Decide retention for agent logs and confirm the tools you use can meet it
- Rehearse a stop: interrupt a live run and check what state it leaves behind
Frequently asked questions
Is this legal advice?
No. It is general information about the regulation text and engineering practice, current as of October 2026. Whether your system is high-risk, and what you must do, needs advice from qualified counsel.
Are AI agents high-risk under the EU AI Act?
Not by default. Risk follows the use case. An agent used in an Annex III area, such as recruitment or credit decisions, may be high-risk; an agent that drafts internal documents usually is not.
When does Article 14 apply after the Digital Omnibus?
According to the Council press release of 29 June 2026, high-risk obligations apply from 2 December 2027 for stand-alone Annex III systems and from 2 August 2028 for AI embedded in Annex I products. Check the Official Journal for the consolidated text.
Does ConvOps make my AI system compliant?
No tool can do that on its own. ConvOps supports oversight practices: engine-evaluated approval gates, stored processes, cancellable runs, and an audit that keeps the agent label apart from the verified person. Classification, retention, export and your legal assessment remain yours.
Does ConvOps record who approved a gate?
It records the verified account behind the advance call that passed the gate, together with the agent's label. There is no separate approver field or signed approval, so design your process so that only overseers make that call.
What counts as a stop button for an AI agent?
Article 14(4)(e) asks for a stop button or similar procedure that brings the system to a halt in a safe state. For agents, that means a reliable way to cancel or pause a run and a process designed so consequential writes happen after the gate, not before.