Two Indian financial regulators have now named the same control.

The Reserve Bank’s draft model risk guidance, published in June, asks for human-in-command arrangements with “override, suspension, or deactivation mechanisms, including kill-switch arrangements” (para 60(ii)), and for continuity planning with fallback mechanisms for when a model is unavailable, degraded or fails (para 43). In August, SEBI’s chairman told a capital markets conference in Mumbai that the regulator’s coming framework for AI will require “kill-switch and humans-in-the-loop controls along with data controls”. This month in New Delhi he said again that the detailed guidelines will be issued soon.

Both want the switch to exist. Neither yet says that anyone has to pull it.

We made this point to RBI in our comments in July, and it applies equally to what SEBI is preparing. This post is the longer version: why a kill switch that has never been used should not count as a control, five questions that tell the two apart, and what the record of a drill looks like, including what a small drill does not prove.

Why existing is not enough

A kill switch usually exists in one of three forms: a sentence in a policy (“the Chief Risk Officer may suspend any model”), a configuration flag that an engineer added some time ago, or a clause in a vendor contract. All three are statements about what would happen. None is evidence that it does.

Banks already know this pattern from disaster recovery. Nobody accepts a recovery site on the strength of its design document. It is switched over on a schedule, the result is written down, and the write-up is what the auditor reads. A stop mechanism for an AI agent deserves the same treatment, for the same reason: a path through software that is never exercised stops working, and nobody finds out until the day it is needed.

Agents raise the stakes because of what they do while you are deciding. A scoring model that should have been withdrawn produces a bad number that a person may still catch. An agent that should have been stopped keeps approving, paying and replying until something makes it stop.

Five questions

1. Who can pull it?

A role in a policy is not a person at a keyboard. Ask who, by name, can stop this agent tonight, from which screen, and whether they have ever done it. Ask what happens at 2 a.m. on a Sunday, and whether stopping a vendor’s agent means raising a ticket with the vendor.

For SEBI-regulated entities this is already more than good practice. Regulation 16C, in force since February 2025, makes the entity solely responsible for the output of the AI tools it uses, whether it built them or bought them. Responsibility that cannot be exercised without someone else’s help is a weak position to hold.

2. Where does the stop take effect, and how long does it take?

This is the question that is hardest to answer from a document. If the stop works by setting a flag that the agent reads, you have the agent’s cooperation, not control over it. And the agent reads that flag on its own schedule: before every action, every few minutes, or only when it restarts.

Figure 1How long an agent keeps acting after the order to stop
The agent checks whether it has been stopped:
Actions completed after the order5
The agent kept acting for2 min 50 s
An illustration, not a measurement. The agent acts once every 30 seconds and the stop is ordered at 2 min 10 s. Nothing about the agent changes between the four cases except where the check sits.

The check belongs in the path an action has to take, somewhere the agent cannot skip: the gateway its tool calls pass through, or the permission it must be given before money moves. The time between the order and the first refused action is the number that matters, and there is only one way to learn it.

Our own product connects to an agent in two ways, and they are not equal on this question. With the SDK, the agent’s code asks before each action, and fails closed if it cannot get an answer. That is a check before every action the code was written to ask about. An action nobody wrapped never asks, so the guarantee is as good as the wiring and the review of it. With the proxy in its enforcing mode, the agent’s tool calls pass through a separate process, and a call that is refused never reaches the tool. That is a stop the agent cannot skip, for every tool it reaches that way. Ask any vendor, us included, which of the two you are being shown.

3. What happens to work in flight?

A stop refuses the next action. It does not undo the last one. Some work will be half finished: a loan approved and not yet disbursed, a claim paid and the customer not yet told. Someone has to own those items, and the list of them has to exist. A stop that leaves nobody knowing what was in progress has traded one incident for another.

4. What does the business fall back to?

Paragraph 43 asks for fallback mechanisms, and the reason is practical. If stopping the agent stops the business, people will hesitate to stop it. Off should not be the only setting. Europe’s AI Act has a useful phrase for the alternative: a stop that lets the system “come to a halt in a safe state” (Article 14(4)(e)).

Figure 2Off is not the only setting
A lending agent, in each of three states:
Answers a customer's question about a loan's statusActs on its own
Approves a personal loan of ₹40,000Acts, then a person reviews
Approves a loan of ₹6,00,000Waits for a person to decide
Its limits decide, action by action, how much it may do alone.
Safe mode keeps the work moving at the speed of the people reviewing it. A halt stops it. Which of the two an agent goes to should be settled before the day it is needed.

Our own default is that an agent which moves money or speaks to customers halts outright, because there should be no doubt about what it may still do, and a low-stakes internal agent drops to safe mode. The less visible half of the answer is the queue those actions land in. It needs people, and a drill is how you find out whether the people on it can carry what the agent was doing alone.

5. What record proves it worked?

Afterwards you need to show who gave the order, when, on which agent, what was refused as a result, who released it and why. You also need that record to be one nobody could tidy up later. A stop that is a setting can be set back without a trace. A stop that is an entry in an append-only log is different: the agent’s state is whatever the latest entry says, so the state and the evidence are the same thing.

One more property is worth insisting on. Automatic triggers should be able to pull the switch: reviewers overriding the agent unusually often, a spike in volume, or a model that changed without anyone declaring it. Only a person should be able to release it.

What a drill looks like

Here is one from our demo environment, exactly as it was recorded on 22 November. The agent is the demo’s fraud triage agent: the one with the most autonomy, and one that touches both customers and money.

Figure 3A kill-switch drill, as it was recorded
Agent fraud-triage-agentDate 22 Nov 2026Started 04:05:33.487 UTCType drill
+0 s
Halt ordered
Actor on record: priya.sharma, governance officer. Reason given: "Kill-switch drill: confirm the halt reaches the enforcement point and holds for 60 seconds".
event 5b3d5cd0 · action activate · response halt · drill true
+19 ms
The agent's next action is refused
Releasing a hold on a payment of ₹85,000. Outcome: block. Rule matched: kill-switch.
evaluation 746ba710 · ref DRILL-20261122-040533-1 · "target halted by event 5b3d5cd0 (drill)"
+15 sto +45 s
Three more actions are refused
A hold release of ₹12,500, a card unblock, and a hold release of ₹2,40,000, fifteen seconds apart. Each one: block, kill-switch.
evaluations 139382f4, 32d35d39, 3da13309 · refs DRILL-20261122-040533-2 to -4
+60.1 s
Halt released
Actor on record: priya.sharma. Note: "Drill complete: 4 of 4 actions refused while halted".
event b9bac034 · action release · releases 5b3d5cd0
+60.1 s
Drill attested
A statement assembled from the two events and the refusals: "4 action(s) blocked at the enforcement point during the 60.1s halt window".
evidence kill_switch_events:5b3d5cd0,b9bac034
From our demo environment. A script ran this drill under a demo persona's login; nobody pulled it by hand. Identifiers are shortened.

Read it against the five questions and it answers the fifth, and half of the second.

It answers the fifth. The order, the refusals and the release are separate entries that refer to one another, and the attestation that closes the drill cites them. That attestation goes into the evidence pack for paragraphs 43 and 60, and the pack is signed.

It answers half of the second. The order took effect at the checkpoint where the agent’s actions are evaluated: the first action after it was refused 19 milliseconds later, the halt held for a minute, and all four actions sent during that minute were refused. Each refusal names the order that caused it. What the drill does not show is that an agent could not have gone around the checkpoint, because the drill sent its actions straight to it.

It does not answer the other three, and we would not want it mistaken for more than it is. A script ran it under a demo persona’s login, so no person pulled anything. The four actions were the drill’s own, so nothing was in flight. And it was a halt, not safe mode, so no fallback queue took any work.

The same day we ran one through the console and filmed it: the order, the halt, the release and the attestation, in under two minutes.

demo.rotavision.com
Recorded on the demo environment. No voice-over: the words are captions on screen, over music.

That one adds a little to the first question: it shows the screen an officer would use, and what the order and the release ask of them. It is no more a rehearsal than the first. The recorder that films our videos pressed the buttons, under the same demo persona’s login, and a script asked for the three actions that were refused. No person pulled anything here either.

We also know the underlying failure from the inside. Drilling our demo used to be a manual step, and until this month nobody had done it. The product noticed before we did: its readiness check for paragraphs 43 and 60 read “0 attested drills”, and the control failed. A manual drill is a drill that gets skipped, so the demo now drills its agents on a rota, one a week. That fixed the evidence gap. It did not turn a scripted test into a rehearsal, and a bank should not accept the first as a substitute for the second.

A drill inside a regulated institution has to be larger: pulled by the people who would be on call, during business hours, on an agent that is doing real work, and held for long enough that the fallback queue is tested too.

What we would ask of the final rules

One sentence in each instrument: that kill-switch arrangements be exercised periodically, and that each exercise leave a record of the time it took to take effect, the scope it covered, and the fallback that carried the work. Banks already do this for disaster recovery. It costs little, and it is the difference between a control and a claim.

What to do before the rules arrive

  • Pick the agent whose failure would cost the most.
  • Answer the five questions for it in writing. The blanks are the finding.
  • Pull the switch during business hours, with the people who would be on call.
  • Keep the record: who, when, how long, what was refused, what the fallback queue did, who released it.
  • Put the next drill in the calendar.

When the final directions and guidelines arrive, “we have a kill switch” will be an easy sentence to write. The institutions in a strong position will be the ones that can say when they last pulled it, and what happened.

Sources: RBI, draft Guidance on Regulatory Principles for Model Risk Management, 2026 (Press Release 2026-2027/528), paras 43 and 60. Remarks of the SEBI chairman at the 23rd FICCI Capital Markets Conference, Mumbai, 19 August 2026, as reported by ANI, and at the India International Trade Fair, New Delhi, November 2026, as reported by IANS. SEBI (Intermediaries) (Amendment) Regulations, 2025, Regulation 16C. Regulation (EU) 2024/1689, Article 14(4)(e).

← All posts