Article 14 expects a person to be able to interrupt a high-risk AI system through a stop button or similar procedure that halts it safely. For an agent, that means no half-finished actions: each step in flight is completed or reversed by a known rule, queued work is held, and the record shows who stopped what and why.
What does Article 14 require, in plain words?
Article 14 of the EU AI Act sets human oversight duties for high-risk systems. One of the abilities oversight measures must give people is to intervene in the system's operation or interrupt it through a stop button or a similar procedure that allows the system to come to a halt in a safe state.
Two words carry the weight. "Interrupt" means a person can act while the system is running, not only review afterwards. "Safe" means stopping must not itself cause harm. A switch that kills the process and leaves the consequences to chance does not meet the intent. The related duty to resist over-reliance is covered in Article 14 automation bias and the people who approve agent drafts.
Who has to provide and use the stop?
The provider must build the ability to interrupt into a high-risk system. The deployer must assign oversight to people with the competence and authority to use it, under Article 26, and must be able to suspend use when it has reason to think the system presents a risk.
That authority question is often skipped. If the only people who can stop an agent are engineers on call, oversight sits in the wrong place. The business owner and the people supervising the agent's work should be able to stop it without asking permission.
What makes stopping an agent harder than stopping a model?
A scoring model can be stopped between requests with little consequence. An agent performs multi-step work across systems. Stop it at the wrong moment and you can leave a refund issued but not recorded, an account closed with no notice sent, or a payment instructed with the confirmation still pending.
So the safe state has to be designed in advance, step by step. For each type of step the agent takes, decide one of three rules: let it finish because it is short and safe to complete; reverse it because a clean reversal exists; or freeze it and hand it to a named person because neither is safe. Steps at depth 4, such as moving money, usually fall in the third group, because reversing them can be harder than finishing them.
Scope matters too. A useful stop has levels: one action, one case, one capability, one agent, or every agent touching a process. Stopping a whole agent because one capability misbehaves causes avoidable disruption.
How do you build a stop that meets Article 14?
| Requirement | Control | Evidence it leaves |
|---|---|---|
| Ability to interrupt | Stop available to named business overseers at action, case, capability and agent level | Access list and stop events in the record |
| Halt in a safe state | Per-step rule: finish, reverse or freeze and hand to a person | Step catalogue with its stop rule |
| No silent loss of work | Queued and in-flight work moved to held, each with a named person | Held queue with owners after each stop |
| Customers not left in limbo | Cases affected by a stop routed to people with deadlines | Case list with completion times |
| Controlled restart | Restart needs approval from the agent's owner; depth may restart lower | Restart approvals and depth on restart |
| Stop works when needed | Stop tested on a schedule in a live-like environment | Test results, including time to halt |
Consider a lender whose credit agent is stopped mid-batch. Some applications have a draft decision, some have none, and a few have a decline drafted but no notice sent. Each needs a person to decide, and declined applicants may still be owed reasons, a point covered in whether adverse action notices apply when an agent declines credit. A stop that leaves those cases unowned creates a second problem.
What should the record show when someone presses stop?
The record should let anyone reconstruct the stop without asking the person who pressed it:
- who stopped the agent, in what role, and at what scope;
- the time, and the reason given;
- every action in flight, its stop rule and its final state;
- every queued item and the person it was handed to;
- customer-facing effects, such as delayed decisions;
- who approved the restart, when, and at what depth.
A stop is evidence that oversight works. File it like any other decision.
What will a supervisor ask?
- Who can stop this agent, and how quickly?
- Show us the last time it was stopped. What state was everything in afterwards?
- Which steps can be reversed and which cannot?
- How do you make sure customers affected by a stop are dealt with?
- When did you last test the stop, and what did you find?
What does Article 14 not cover?
It does not prescribe a physical button or a particular interface; a documented procedure can meet it. It does not decide your stop rules per step, which are design decisions. It does not address the data protection side of a stop, although the ability to interrupt is a mitigation worth recording in a GDPR Article 35 DPIA for the agent. And as a binding duty it applies to high-risk systems, though any agent that acts should have one. See AI agent regulation mapped to controls for the wider picture.
This is an explanation of the rules, not legal advice; confirm how they apply to you with counsel.
Questions people ask
Is dropping an agent to read-only the same as stopping it?
It is a form of interruption, and often the better one. The agent keeps reading and can help people working the held cases, but it can no longer draft or act.
Should the agent be able to stop itself?
Agents should escalate when they meet something outside their bounds, which is a kind of self-interruption. But the Article 14 stop must sit with people and must work even if the agent does not cooperate.
How fast should a stop take effect?
Fast enough that no new actions start after it is pressed. Steps already in flight then follow their stop rule. Measure the time in testing and set a target the business owner agrees.