An AI agent needs a stop plan before it gets access to your business systems. Name the person who can pause it, identify the switch that prevents new actions, and know how to cut off each connected account if the first switch fails. Keep a record of what it already did. Restart only after someone has checked the cause and the affected work.
This matters even when the first job looks harmless. An agent that drafts a weekly report may also search a shared drive, read messages or update a task list. A bad answer is usually easy to correct. An unapproved email, wrong record change or repeated action can travel further. The National Cyber Security Centre’s guidance on agentic AI calls for defined boundaries, monitoring and an emergency way to halt autonomous activity. It says stopping the agent process alone may not be enough if the wider system can still act.
What counts as an AI agent going wrong?
I would decide this before the first live run. Otherwise the person watching has to invent a threshold while the system is still working. Write down both the expected result and the events that require a pause.
For a small business workflow, sensible stop triggers might include an external message sent without approval, a customer record changed outside the agreed task, an attempt to read a folder the agent was not meant to use, a repeated action after an error, or a result that has no usable source. A missed formatting rule is usually a correction. An unexpected payment, disclosure or deletion is a stop event. The exact line depends on what the agent can reach and what a mistake would cost.
Separate three states in the runbook: pause, when the workflow stops taking new work; contain, when connections and credentials are disabled so it cannot continue through another route; and recover, when you review what happened and correct any effects. A dashboard button labelled “pause” may only stop a schedule. It may leave a current run, queued action or connected service active. Test what that button actually does.
Who should be able to stop it?
Give one named owner responsibility for the workflow and a deputy for absences. The owner should know the approved task, the connected systems, what needs human approval and where the run history lives. The person who spots a problem must know how to reach them immediately. If an agent runs outside office hours, either arrange cover or limit it to work that can safely wait until someone is available.
The stop procedure should be short enough to use under pressure. A single page can hold the workflow name, owner and deputy, stop triggers, the control that blocks new runs, the service accounts or integrations to disable, the log location and the person who decides whether to restart. Keep the page somewhere the team can reach if the AI service itself is unavailable.
The NCSC recommends named responsibility and human oversight where unintended activity could have significant consequences. It also recommends monitoring that lets people investigate and intervene. Those controls are more useful when the person on duty has actually tried them than when they exist only in a supplier’s feature list.
Where is the real stop switch?
Trace the full route from trigger to action. A customer form might start an automation, which calls an AI service, reads a knowledge base, writes to a CRM and sends an email. Stopping the AI call may leave the automation queue or mail step intact. Disabling the CRM connection may stop record updates but not messages. Draw that route before the pilot and test each control.
| Layer | What to check before launch | What to do in a stop |
|---|---|---|
| Trigger and queue | Where new jobs enter and whether queued jobs continue | Disable the trigger and hold or clear the queue according to the incident plan |
| Agent run | Whether a current run can be cancelled | Cancel it, then confirm it has stopped making tool calls |
| Connected tools | Which systems allow reading, writing, sending or deleting | Disable the affected connection or revoke the service account’s access |
| Outbound action | Whether emails, payments or record updates wait for approval | Hold pending actions and check what has already gone through |
| Monitoring | Where the activity and approvals are recorded | Preserve the relevant records before routine rotation or deletion |
For a first deployment, give the agent only the access it needs. Read-only access is enough for many useful drafts. If it needs to change records, use a narrow account rather than a person’s broad login. Keep sending, spending, deleting and permission changes behind a real approval step. Our guide to AI access to your CRM explains how to separate reading, drafting and writing rights.
The NCSC says credentials and network access need their own controls, and that an emergency shutdown may have to interrupt communications beyond the agent process. For a small firm using a hosted service, the practical equivalent is to know which administrator can disable the workflow and revoke each integration. Ask the supplier what happens to a run already in progress when access is removed. Verify the answer in a safe test account.
What should you do in the first few minutes?
Use a simple sequence, with the person in charge free to escalate faster if customers or sensitive data are affected:
- Stop new work. Pause the trigger or schedule. Record the time and the last known task.
- Contain the active run. Cancel it and disable the connected action that could cause more harm. Check that queued actions are held.
- Preserve the evidence. Save the run ID, inputs, tool calls, approvals, outputs and relevant system logs. Restrict access to these records because they may contain personal or commercial data.
- Check the effects. Look in the actual email, CRM, payment or document system. Do not assume a cancelled run means no earlier action completed.
- Assign the response. Name who will contact affected people, correct records and decide whether a privacy, security or consumer issue needs specialist handling.
This sequence is an operating example, not a universal incident clock or legal reporting rule. Your existing security and data protection procedures still apply. If personal information has gone to the wrong place, involve whoever handles data protection for your business. If money or a customer commitment is involved, involve the person with authority to put it right.
The ICO’s current call for evidence on agentic AI asks about data security, transparency and accountability. It closes on 20 November 2026 and is gathering evidence, not setting new final guidance. That is a timely reminder to decide who can explain what your agent did with personal information. The NCSC also advises protecting activity logs so they remain useful when an incident is reviewed.
What if the agent has already contacted a customer?
Read the actual message or decision, not just the AI summary. Find out who received it, what it said, whether it was sent once or repeatedly, and whether the customer acted on it. Correct the record and response through the same channel where practical. Keep the correction factual and make it easy for the customer to reach a person.
The Competition and Markets Authority’s guidance on using AI agents with consumers says businesses remain responsible for an agent’s conduct, including when a supplier provides it. It calls for monitoring and quick correction when the agent gives wrong information about products, prices or rights. This is a stronger reason to keep customer-facing actions under review than a vague promise that the model is accurate. For the approval boundary, see our article on AI customer emails.
How do you test the stop plan before launch?
Run a drill on a test workflow that cannot reach real customers or live financial records. Give the agent a harmless repeated task, then introduce a known failure such as an unavailable data source or an instruction that requests an out-of-scope action. The aim is to test your controls, not to see whether a prompt sounds reassuring.
Ask someone other than the builder to follow the one-page runbook. Can they find the current run? Can they stop new jobs and the active job? Can they identify all connected accounts? Can they tell which actions completed before the stop? Can they retrieve the evidence without opening the AI service? Record the gaps and repeat the drill after fixing them.
Put a limit on each run as well: how many records it may change, messages it may draft, steps it may take and how long it may continue without review. Technical limits help when the person watching misses the first warning. Start with a small batch and explicit approval, then widen only when the observed results justify it. Our 30-day AI pilot scorecard offers a way to record quality and exceptions alongside time saved.
The UK government’s agentic AI resilience factsheet, updated on 3 November 2026, describes possible future measures for regulated essential and digital services, including monitoring and evidence preservation. These are proposals subject to consultation and further process. They are not a new blanket requirement for every SME. The useful lesson for a smaller business is operational: if you cannot see what an agent is doing, stop it or reconstruct its actions, you are giving it more freedom than you can manage.
When is it safe to restart?
Do not restart because the alert has stopped. First establish the cause: a bad instruction, an unexpected document, overbroad permission, wrong source data, a supplier fault or a missing approval gate. Check the affected systems for completed actions and correct them. Then change the control that failed, test the same scenario in a safe environment, and get the named owner to approve a limited restart.
I would begin with read-only or approval-required work, watch the first few runs and keep the stop control close at hand. If the workflow cannot show what it did, treat that as an unresolved gap. A useful AI service should survive an ordinary question: “How would we stop this today, and what would we know afterwards?”
If you need to design or test that boundary for one workflow, talk to Ajairu about AI governance.