Emergency stop for AI: anyone may halt, not anyone restart
Every machine has an emergency stop; AI almost never does. What machinery law has regulated for years and how to build halt and restart for AI.
In short
An emergency stop for AI is a single action that any person working with the system can trigger without asking and that brings it into a safe state: write permissions removed, queue paused, downstream scripts stopped. Restart is released by the responsible person only after a review, just as with a machine’s emergency stop, where disengaging alone restarts nothing.
Auf Deutsch lesen: deutsche Fassung
On every machine on the shop floor there is a red mushroom button on a yellow background. Everyone knows where it is. Everyone may press it, nobody has to ask first. In my AI seminars I then ask a single question: where is the emergency stop for your AI? At the machine everyone thinks of it. With AI, hardly anyone thinks of it at first.
Yet the question is no gimmick. An AI agent with write permissions in the ERP posts, orders and dispatches faster than anyone can reach IT by phone. In many companies today, switching off means: write a ticket, call the service provider, if in doubt cancel the subscription. That is not an emergency stop. That is a trip to the authorities.
What the emergency stop on a machine has to be able to do
Machinery law describes the emergency stop with surprising precision. The Machinery Directive 2006/42/EC currently in force regulates it in Annex I, section 1.2.4.3. The Machinery Regulation (EU) 2023/1230 adopts this almost word for word in Annex III, Part B. It applies from 20 January 2027; the text first published said 14 January, and a corrigendum in the Official Journal corrected the date. Six properties can be read from the text:
- It is clearly identifiable, clearly visible and quickly accessible.
- It stops the hazardous process as quickly as possible, without creating new hazards.
- It stays engaged until someone releases it by a deliberate action.
- Releasing it does not start the machine. It only makes restarting possible.
- It is available in every operating mode.
- It supplements the other safeguarding measures and replaces none of them.
One section further on, in 1.2.4.4, there is a sentence every plant engineer knows. Where several machines work together, the stop must also be able to stop the related equipment as soon as its continued operation may be dangerous. A linked installation therefore stops as a whole. These two sections are the best template for AI that I know.
“Disengaging the device must not restart the machinery but only permit restarting.” Machinery Directive 2006/42/EC, Annex I, 1.2.4.3
What an emergency stop for AI looks like
Translated to an AI system, it does not need a red mushroom. It needs a single action that every person working with the system knows and that takes effect without asking. What this action does technically depends on the system. For an agent that creates purchase orders in the ERP, it is typically three steps triggered together:
- The AI’s technical user loses its write permissions. It may continue to read, so that it remains traceable afterwards what it last saw.
- The queue of open jobs for the AI is paused. Nothing is deleted.
- The scripts that process the AI’s results further stop as well. That is the linked installation from 1.2.4.4, only without a conveyor belt.
The machine knows the safe state too. With AI that means: no half-finished transaction. A purchase order that is just going out to the supplier is either completed cleanly or withdrawn. Whatever is left pending from the moment of the halt is in the log with a timestamp. Anyone who defines that beforehand does not have to invent it in an emergency.
With purchased services, the button is often not in your hands at all. The provider can switch off its model; you cannot. What you do have in your hands is the access: revoke the API key, cut the connection to the ERP, lock the technical account. That is your emergency stop, even if the AI keeps running at the provider. So before buying, ask which of these actions you can trigger without the provider and how long each of them takes.
How quickly a chain of automation takes off is shown by a case I like to tell in the seminar. A script triggered the same production confirmation five times in a row, and the quantity was then in the system five times over. It was noticed by a colleague who knew that this quantity could not possibly exist in the building. No alert, no watchdog. A person with experience. Had an AI been attached behind the script, deriving purchase orders from stock quantities, the orders would already have been on their way before anyone picked up the phone.
Anyone may halt. Not anyone may restart.
This is where many companies get tangled up. They ask: who may switch off the AI? And because switching off sounds like responsibility, the answer lands with the management or with IT. That turns the button back into a trip to the authorities.
The machine solves this with a clear separation. Anyone may halt, without giving reasons, without asking. Restart is a separate, deliberate step: first the cause is clarified, then someone who knows the installation gives the release. Exactly this separation holds for AI too.
- The halt may be triggered by any person who works with the system. The colleague in internal sales who sees a strange email to a customer needs no permission.
- A halt costs time, nothing more. A false alarm is cheap. A halt that does not happen is expensive.
- Restart is released by the person to whom the system is assigned. Beforehand they review the log from the time of the halt and have corrected whatever went wrong.
- Every halt is recorded: who, when, why. Without any question of blame. Whoever pressed did their job.
The obvious objection: then everyone will be pressing it all the time. At the machine the low threshold has nevertheless proved itself, and there is a reason for that. The emergency stop is trained, and whoever presses it explains afterwards what they saw. That curbs playing around and lets the real emergency through. An emergency stop nobody knows about does not exist.
The button becomes most urgent when it is no longer you controlling the AI. The IT service provider of a mid-sized company had three security incidents within a few months, all via the same access route, the company’s Microsoft account. Someone sitting at an AI agent with stolen credentials does not wait for a ticket. At that moment the only thing that counts is how quickly someone in the company cuts the connection, and whether they dare to do so without reaching the management first.
Who releases the restart has to be settled beforehand. How to assign responsibility for an AI so that it holds in an emergency.
To assigning responsibilityWhat the EU AI Act, machinery law and ISO 42001 say about it
For high-risk AI the stop button is law. Article 14(4)(e) of the EU AI Act requires that the people to whom oversight is assigned can intervene in the operation or interrupt it through a “stop” button or a similar procedure, so that the system comes to a halt in a safe state. Article 26(5) obliges deployers to inform the provider and the market surveillance authority and to suspend use as soon as they have reason to consider that the system, when used as intended, presents a risk to health, safety or fundamental rights (risk within the meaning of Article 79(1)). The Digital Omnibus, Regulation (EU) 2026/1744, did not touch either provision. They apply to systems under Annex III from 2 December 2027.
For AI in machinery, the same omnibus rerouted the path. The Machinery Regulation is now in Section B of Annex I to the EU AI Act. The requirements for AI as a safety component of a machine are adopted by the Commission by delegated act under the Machinery Regulation, added to its Annex III, that is, to the annex that also contains the emergency stop. They are to apply from 2 August 2028. Anyone building machines with AI thus gets their stop button from the familiar set of rules.
For most systems in mid-sized companies, however, no law requires a stop button. A quotation assistant, an agent in purchasing, a chatbot in customer service are as a general rule not high-risk. The emergency stop is built anyway, for the same reason as on the machine: the damage arises faster than anyone can clarify who is responsible.
ISO/IEC 42001 does not turn this into a separate obligation, but it sets the direction. Control A.6.2.7 requires you to determine what technical documentation each category of interested parties needs for an AI system, and to provide it. The implementation guidance in B.6.2.7 names as possible content a plan for dealing with failures: rolling the system back to an earlier state, switching off individual functions, notifying users. Added to that are functions that allow operators to intervene in operation. Ongoing monitoring itself belongs in A.6.2.6. In the audit I ask at this point for the plan and, straight afterwards, for the proof that it works.
The check you can do on Monday
Four questions are enough to establish whether your emergency stop exists or is only written in a policy. Ask them for every AI system that acts by itself, that is, posts, orders, dispatches or changes data.
- Where is it? Name the one action and the place where it is triggered. “With IT” is not a place.
- Who can get to it? Every person who works with the system knows it and may trigger it.
- What happens then? Write permissions gone, queue paused, downstream scripts stopped, no transaction left half-finished.
- When was it last pressed? With date, time to standstill and the log of a test run.
The fourth question is the most uncomfortable. An emergency stop that has never been triggered is an assertion. On the machine it is tested regularly. For AI, a planned test run outside peak hours is enough: trigger the halt, look at the state, restart after release. Afterwards you know how long it really takes, and usually also which script nobody had on their list.
How robust an answer is can usually be heard in the first sentence:
- Weak: IT can switch that off at any time. Robust: every person in purchasing locks the technical user with one button; in the last test run everything was at a standstill after four minutes.
- Weak: we have an emergency plan. Robust: here is the plan, here is the log of the test run, and here are the two scripts we found and connected in the process.
- Weak: the provider can block the model. Robust: we revoke the API key ourselves, without the provider, in under a minute.
Then the good news, which is talked about too rarely. An emergency stop that works makes you bolder. Anyone who knows they can halt at any time is more willing to give an AI the permissions it needs to really take work off people. On the shop floor that goes without saying. Nobody would run a milling machine without an emergency stop, and nobody leaves it standing idle because of one. With AI it is the same calculation.
The emergency stop cannot replace anything, and machinery law says so too. Amount limits, clean permissions and approvals remain what they are. The button is the last line. It only has to be there when all the others have already been breached.
Primary sources
- Regulation (EU) 2024/1689 (EU AI Act), consolidated version of 27 July 2026, Articles 2(2), 14(4), 26(5) and 113, Annex I
- Regulation (EU) 2026/1744 (Digital Omnibus on AI), Article 3, amendment of Regulation (EU) 2023/1230
- Regulation (EU) 2023/1230 on machinery, consolidated version of 27 July 2026, Article 54 and Annex III Part B, 1.2.4.3 and 1.2.4.4
- Corrigendum to Regulation (EU) 2023/1230, OJ L 169 of 4 July 2023, p. 35 (date of application 20 January 2027)
- Directive 2006/42/EC (Machinery Directive), Annex I, 1.2.4.3 and 1.2.4.4
- ISO/IEC 42001:2023, AI management systems, Annex A.6.2.6 and A.6.2.7, Annex B.6.2.7
Setting up an emergency stop for an AI system
Define the one action
Determine what the halt triggers: remove the technical user’s write permissions, pause the queue, stop downstream scripts. All with a single action.
Describe the safe state
Determine what happens to transactions that are in progress at the time of the halt: complete them or withdraw them. There must be no half-finished postings.
Open access to everyone
Make the halt accessible to every person who works with the system, without a ticket and without permission.
Regulate the restart
Determine who gives the release after a halt and what is reviewed beforehand: cause, log from the halt onwards, corrections.
Train
Show everyone involved where the emergency stop is and when it should be pressed. Whoever presses it explains afterwards what they saw.
Run a test and record it
Trigger the halt once as planned, measure the time to standstill and document the date, the result and the gaps in the plan for failures.
Frequently asked questions
Does every AI need an emergency stop?+
By law, the EU AI Act requires a stop button only for high-risk systems (Article 14(4)(e)). In practice, every system that acts by itself needs one, that is, one that posts, orders, dispatches or changes data. An assistant that only suggests texts which a human reviews and sends already has its emergency stop in the human.
Is it enough to call the service provider or cancel the subscription in an emergency?+
No. An emergency stop takes effect in seconds and without asking. A call and a cancellation take hours to days, and in that time the system keeps working. The halt must be triggerable within your own company, for example by the AI’s technical user losing its write permissions.
Who may restart the AI after a halt?+
The person to whom the system is assigned, after they have clarified the cause and reviewed the log from the halt onwards. This corresponds to machinery law: disengaging the emergency stop does not restart the machine by itself, it only makes restarting possible.
From when does the stop button of the EU AI Act apply?+
Article 14 belongs to Chapter III, Section 2. Under Article 113 as amended by Regulation (EU) 2026/1744, it applies to high-risk systems under Annex III from 2 December 2027 and to products under Annex I, Section A from 2 August 2028. For AI in machinery, the requirements come via delegated acts under the Machinery Regulation, likewise from 2 August 2028.
Author & expert review: Lars Zimmermann · ISO/IEC 42001 Senior Lead Auditor & Senior Lead Implementer · ISO/IEC 27001 Lead Auditor & Lead Implementer (PECB)
Last updated: 10 November 2026. Researched and reviewed to the best of our knowledge; not a substitute for individual legal advice.
Questions about your own case?
In a free 15-minute intro call we assess where you stand on ISO 42001, ISO 27001 and the EU AI Act, honestly and without a sales pitch.