When the human needs to step in, can they — quickly and reliably?
Baseline: 1Enhanced: 23 controls
Who it applies to
Enter for every AI system. Go as deep as your system letter, and implement in your mode.
Why it exists
Make sure the human can actually intervene — override a decision, pause the system, take control — and that the means to do it works under pressure. Oversight that can see a problem but cannot stop it is observation, not control. The intervention must be reliable, reachable in the moment, and governed by a procedure the operator knows, because an override mechanism discovered for the first time during a crisis is a mechanism that fails during a crisis.
What to actually do
Keep a way back to the manual process — from the first day. Before you rely on any AI, make sure you could stop using it today and carry on by hand: the manual procedure is still written down, someone still knows how to run it, you know how to suspend the tool’s access, and the decision to stop does not wait for anyone’s approval. This needs no technology and no structure, and it is what lets a small organisation recover from a failing AI before it has a kill switch.
Give the operator a reliable way to override and to halt. Build the means to reverse or block a specific AI decision and to stop the system as a whole, and place them where the operator can reach them in the moment they are needed. Connect the stop to the disengagement and fallback you built in the Governance security program and Technical Security — stopping the AI must not strand the process.
Write and rehearse the intervention procedure. Document when and how to intervene, and rehearse it, so that under stress the operator acts from practice rather than improvises from panic. An untested procedure is an assumption.
Controls
Baseline1 controls
Control 196
Be able to stop using any AI system and return the affected process to its manual form within the same day.
BaselineCorrectiveall rolesper system
What it is
The manual procedure is kept written and known; who may decide to stop is named and needs no further approval; how to suspend the tool’s access is known.
How to implement it · Minimum structure
Manual procedure: how the process was done before the AI, written down and kept current
Who can run it by hand: at least one named person who still knows how
Who may decide to stop: one named person, plus a deputy, needing no further approval
How to suspend access: disable accounts or keys, switch off the integration, pause the subscription
Who to tell when stopping: staff using the tool and, if service changes, customers
How to start
For each AI tool, write half a page: how the work was done before, who can do it, who may pull the plug, and how to cut access.
Evidence
The written manual procedure and the named person who may stop the AI.
Mode: light / standard / full
light
Half a page per tool in a shared folder; the owner-manager is the named person.
standard
Part of the process documentation, with a named decision-maker and deputy per system.
full
Built into business-continuity plans, with the access-suspension steps held by IT.
Enhanced2 controls
Control 197
Provide reliable override and halt mechanisms reachable in the moment.
EnhancedCorrectiveall rolesper system
What it is
Connected to the business-continuity fallback so stopping the AI does not stop the process.
How to implement it · Minimum rules
Override: reverse or block a single AI decision before or after it takes effect
Halt: stop the whole system or one function, e.g. a feature switch, disabling the integration, revoking its credentials
Reachable: available to the operator on duty from their normal interface, whenever the system runs
Fallback: on halt, work routes to the manual process or a safe default, not a dead end
Speed: halt takes effect within a defined time; every override and halt is logged
How to start
Identify the fastest way to stop this system today and test whether the operator on duty can trigger it without asking anyone.
Evidence
The override and halt mechanisms, who can reach them, and a test record.
Mode: light / standard / full
light
The provider's pause or disable setting, with the operator holding the access needed to use it.
standard
A switch in the application plus routing to the manual process, owned by the system owner.
full
Engineered halt controls with automatic fallback, monitored and tested by operations.
Control 198
Document and rehearse the intervention procedure.
EnhancedCorrectiveall rolesper system
What it is
Operators act from practice, not improvisation.
How to implement it · Minimum structure
Triggers: when to intervene, e.g. harmful output, drift alert, action out of bounds
Who acts, with a deputy, and the exact steps to override or halt
Switch to the fallback and who to notify
Conditions for restarting and who authorises it
Rehearsal: at least yearly and after significant change, with lessons recorded
How to start
Write the procedure on one page and run a 30-minute tabletop exercise with the operators.
Evidence
The documented procedure and rehearsal records with lessons learned.
Mode: light / standard / full
light
A one-page procedure and a yearly walk-through led by the owner-manager.
standard
A procedure per system with a recorded tabletop exercise by the system owner and operators.
full
Live drills scheduled by operations, combined with continuity exercises and tracked findings.
Evidence
The written manual procedure and the named person who may stop the AI; The override and halt mechanisms and their reachability; the documented procedure; rehearsal records.