Stopping an AI agent safely means suspending everything it does outside the organisation, the writes to other systems, the emails and the robots, without losing the cases that were half done. A good stop mechanism records who stopped it and why, lets people keep making decisions and, on restart, moves the cases forward in the order in which they stopped.
AI agents already handle orders, supplier invoices, complaints and bookings in many organisations. The more work they do, the more it matters to answer a simple question. If something goes wrong, how do you halt everything without creating a bigger problem? This article explains what stopping an agent means, when it makes sense, how it is done and what the organisation should decide before it needs to.
What does it mean to stop an AI agent?
An AI agent is a program that reads what comes in, such as an email or a document, proposes what to do and carries out actions in other systems, such as creating an order in the ERP or replying to a customer. Stopping an agent is not switching off a server. It is preventing the agent from having any effect outside the platform, while keeping the state of each case so the work can continue later.
There are three levels of stopping, from the narrowest to the widest.
- Holding a case. The case waits for a person, for instance because it went over a value limit that requires approval.
- Pausing a process. Every case in that process stops, and the other processes carry on.
- Stopping everything. All external actions and all inputs stop at once.
The distinction matters because each situation calls for a different level. A supplier sending odd invoices justifies pausing one process. A failure in a connected system justifies stopping everything.
When does it make sense to stop the agents?
Most of the time, oversight happens case by case, with people validating whatever is unclear. Stopping is for situations where the problem lies not in a single case but in the whole. Some examples follow.
- A connected system, such as the ERP or a bank portal, is under maintenance or returning errors.
- A business rule has changed and the processes have not yet been updated.
- The team spots an unusual pattern, such as many orders from the same customer in a short period.
- There is a suspicion that a mailbox has received fraudulent messages.
- The organisation is about to run a migration or a year-end close and wants to freeze all writes.
Regulation points the same way. Article 14 of the European Artificial Intelligence Act provides that, for high-risk systems, people must be able to intervene and interrupt the system through a stop button or a similar procedure that brings it to a halt in a safe state. Even outside those cases, it is a sound reference for any organisation that puts agents to work.
How do you stop everything with a single command?
A general stop has to be fast and complete. If you have to switch off agents one by one, or remember which robots are active, someone will forget one. Good practice is therefore a single command that suspends, at the same time, the writes to connected systems, outgoing communications, the robots and the intake of new cases.
In EngiAIFlow, whoever administers the installation can interrupt, with a single command, all of the platform's external actions, the writes, the communications and the robots, as well as all inputs. A reason is mandatory and is recorded in the audit log, with the author and the date. When activity resumes, it is clear why it stopped and who decided.
Two precautions make the stop safe.
- Stop before the external effect. Cases should wait before the step that writes to a system or sends a message, not halfway through it.
- Do not lose work in progress. Each case keeps the point it had reached, with the data already read and the decisions already taken.
How do you pause just one process?
Not every problem calls for a general stop. Imagine an organisation that has automated orders, supplier invoices and delivery bookings. If its main supplier changes the invoice layout, it makes sense to pause only the invoice process while the rule is adjusted, and let orders and bookings carry on.
In EngiAIFlow, pausing a process also suspends the cases in progress in that process, and the rest continue. It is the way to contain a local problem without stopping the whole operation.
There is also a finer level that works without anyone pressing a button. Limits such as "above this amount, ask for approval" hold the case before the step until a person approves it. This kind of limit avoids many stops, because unusual cases reach the team before they have any effect.
How do you restart without duplicating work?
Restarting is the part most often forgotten and the one that can cause the most errors. The two main risks are losing cases and repeating actions, such as creating the same order twice.
A well designed mechanism deals with both.
- Nothing is lost. In EngiAIFlow, cases wait before the step with an external effect, people keep consulting and deciding during the stop and, when activity resumes, the cases proceed in the order in which they stopped.
- Nothing repeats by itself. A write to a connected system is never repeated automatically. If the process is interrupted, a person decides whether the write has already been made.
Before restarting, it is worth confirming that the cause of the stop has been resolved, reviewing the cases that have waited longest and letting the teams that depend on those processes know.
What should the organisation decide before it needs to?
A stop only works if it was prepared before the emergency. These are the decisions worth taking calmly.
- Who can stop. Name the people allowed to stop everything and to pause each process.
- In which situations. Write down the criteria that justify each level of stop, so the decision does not depend on who happens to be on duty.
- Which limits to apply. Set the amounts and conditions that hold a case and ask for approval.
- How to communicate. Agree who informs the teams, customers or suppliers affected.
- How to restart. Keep a short checklist for returning to normal activity.
A useful exercise is to rehearse a stop on a quiet day. Activity is interrupted, everyone checks that cases are waiting, and activity is resumed, so that everyone knows what to do on the day it is real. To learn more about the balance between what runs alone and what people validate, read how to decide what automation does alone and what people validate.
Frequently asked questions
What does stopping an AI agent mean?
It means suspending the actions the agent takes outside the platform, such as writes to systems, emails and robots, while keeping the state of each case so the work can continue later.
Does stopping the agents lose the cases that were in progress?
It should not. In EngiAIFlow, cases wait before the step with an external effect and, when activity resumes, they proceed in the order in which they stopped.
Who can stop the agents in EngiAIFlow?
Whoever administers the installation can interrupt all external actions and all inputs with a single command. A reason is mandatory and is recorded in the audit log, with the author and the date.
Is it possible to stop just one process?
Yes. Pausing a process also suspends the cases in progress in that process, and the other processes carry on.
On restart, can the same order be created twice?
In EngiAIFlow, a write to a connected system is never repeated automatically. If the process is interrupted, a person decides whether the write has already been made.