What is an AI kill switch?
An AI kill switch is a control that halts AI activity immediately, so no further AI step takes effect until a named person restarts it. In a bank, it should stop all AI work at once or a narrower scope, be held by named staff, work without the vendor's help, and leave a record of who used it and why.
A kill switch is the strongest of a family of controls. Below it sit holds, where a single AI step waits for a named signer, and blocks, where a step that breaks the bank's rules never runs. Together they give a bank real authority over AI in real time, rather than finding out what happened afterward.
Why does a bank need a kill switch if AI only drafts?
Even when AI only prepares work, its output flows into things that matter: spreads feed ratios, ratios feed grades, drafts feed memos, and analyses feed collections strategies. A data feed that breaks, a model update that behaves differently or a misconfigured rule can produce bad work at machine speed across many loans. The faster the AI, the more it can get wrong before anyone notices.
AI leaders are making the same point. In an October 2026 essay, Microsoft CEO Satya Nadella argued, as reported by OfficeChai, that companies should treat AI models like insiders with powerful access and keep controls, including the ability to contain them, outside the model. A kill switch is the simplest form of that idea.
What are the levels of AI control: allow, hold, block and stop?
A kill switch works best as the top level of a ladder, not as the only control. Most of the time, rules checked before each AI step do the work, and the kill switch is there for when something looks wrong across the board.
| Control | What it does | When it is used |
|---|---|---|
| Allow | The AI step runs and is recorded | The step is within the bank's rules |
| Hold for signer | The step waits for a named person to approve or reject it | The step needs authority the AI does not have |
| Block | The step does not run, and the attempt is recorded | The step breaks a rule |
| Narrow stop | All AI steps in one scope pause | A problem is limited to one workflow, model, data source or portfolio |
| Stop all | Every AI step halts at once | Something looks wrong and the scope is unclear |
When should a bank stop all AI activity?
The threshold should be low. Stopping AI for an hour costs little, because people can carry on with the work. Letting a problem run costs much more. Write the triggers down in advance so nobody hesitates.
- AI output contains figures that do not match the source documents across several files.
- A sudden rise in blocked or held steps that nobody can explain.
- Signs that customer data went somewhere it should not.
- An unannounced change in the model or vendor behavior.
- A suspected security incident involving any system the AI can reach.
- A rule or authority change that has not yet been reflected in the AI controls.
How narrow can an AI stop be?
A stop-all switch is essential, but it is a blunt tool. Narrower stops let the bank pause the part that is wrong and keep the rest running. Design the scopes before you need them.
- One workflow: pause covenant testing while spreading and memo drafting continue.
- One model: pause work by a specific AI model after an update.
- One data source: stop AI from reading a feed that looks corrupted.
- One portfolio or segment: pause AI work on a book of loans under special review.
- One borrower: stop AI steps on a single relationship during a dispute or workout.
Who should be able to pull the AI kill switch?
Keep the list short and named. Typically the chief credit officer, the chief risk officer and an on-call operations or technology lead can stop all AI activity, while workflow owners can use narrower stops in their own area. Anyone who sees a problem should know how to reach one of them quickly.
Restarting deserves more care than stopping. Many banks require a second person to agree before AI activity resumes, along with a short note of what was found and what changed. The switch must also belong to the bank. If stopping AI requires a vendor support ticket, it is not a kill switch.
What happens to work in progress when AI is stopped?
A stop should be clean. AI steps already completed and recorded stay as they are. Steps that were proposed but not yet allowed should not run. Steps held for a signer should stay held. Nothing should be half written into a spread, a grade field or a borrower record.
People keep working. Analysts and officers can continue manually while AI is stopped, and the record should show which work was done by people during the stop. When AI restarts, it should pick up from the recorded state, not replay steps that were in flight.
How do you test an AI kill switch?
A control that has never been used is a guess. Test the kill switch the way the bank tests disaster recovery: on a schedule, with a written result.
- 1Choose a test window and tell the affected teams.
- 2Start AI work on a set of test or low-risk files.
- 3Pull the stop-all switch and confirm that no further AI step takes effect.
- 4Check that held steps stayed held and nothing was partly written.
- 5Repeat with one narrow stop, such as a single workflow.
- 6Restart with the required approvals and confirm work resumes from the recorded state.
- 7Record the results, the time taken and any fixes, and report them to the risk committee.
What should be recorded when the kill switch is used?
The record of a stop is as important as the stop itself. It shows that the bank was in control, and it is what the bank learns from afterward.
- Who pulled the switch, when and at what scope.
- The reason given at the time.
- Which AI steps were in progress, held or blocked at that moment.
- What people did manually while AI was stopped.
- What was found, what was changed and who approved the restart.
- When AI resumed and at what scope.
How Aarvion helps
In Aarvion Risk OS, every AI step in credit work is checked against the bank's own rules before it takes effect, with four outcomes: allow, hold for the right signer, block, or stop all activity. Stop all halts AI work across the platform at once, and each step, hold and stop is recorded with who, what, which rule and when.
The bank decides who holds that authority. Banks, credit unions and lenders typically start with a 90-day pilot on one workflow and test the stop control as part of it.
