Fijishi Kill-Switch Certificate - Part of Fijishi Certification: certification of AI decisions against a named standard.
Does one named AI system's override or human-review mechanism meet one named override or human-review requirement?
What it answers
A narrow finding on one named AI system, measured against one named override or human-review requirement. At intake we confirm the requirement your system will be measured against.
Two examples:
- A requirement in force: for example, a human-review requirement for decisions based solely on automated processing, where the rule gives the person a right to have the decision reviewed by a staff member.
- A published draft: paragraph 60(ii) of the Reserve Bank of India's draft guidance on model risk management, which calls for override, suspension or deactivation mechanisms, "including kill-switch arrangements" – consultation draft; comments closed 24 July 2026; no date set. Draft text: Reserve Bank of India · Status: Regulatory Deadline Calendar.
Fijishi responded to this consultation on 27 June 2026. The response is published in Institutional Submissions.
Where the requirement is a published draft, or a rule adopted but not yet in application, the finding is a readiness finding and says so; the requirement is labelled exactly as the Calendar labels it.
What it is not
A general AI governance review. It is scoped to one system, one named requirement and one finding.
Process
Fijishi examines the system's override or human-review mechanism – how it is invoked, by whom, and what record it produces – against the requirement named at intake. The finding states whether the mechanism meets the requirement and, where it does not, what would need to change.
The draft criteria behind this Certificate are developed in public: FWD-1 Stopping an AI decision. The Certificate itself measures against the requirement named at intake.
Turnaround and fee
We confirm the turnaround at intake; the Certificate is scoped to be completed in 1-2 weeks once system documentation is provided. The fee is fixed at intake, before any work begins, and does not depend on the finding. Intake has no fee.
Who this is for
Chief risk officers, heads of risk and compliance officers at insurers, reinsurers, banks, non-bank lenders and other regulated firms that face a named override or human-review requirement - in force or published in draft - and want an independent finding for their own governance process. An AI vendor may also commission one on its own system, measured against a requirement its customers face.
What this certificate is not
Fijishi is an independent certification organisation. It is not a government body or a regulator. The Certificate is not regulatory approval, a compliance determination by any regulator, or legal advice. It is an independent, documented finding that the firm owns outright and can keep in its own record of diligence.
A readiness finding measures the system against a draft as published, or against a rule before it applies. It does not predict the final rule or how it will be enforced.
A finding is dated and stands against the text it was measured against. If the requirement is amended, or a draft becomes final, a finding against the new text is a new engagement.
Fijishi does not certify a system it helped design or build.
After the finding
A firm may ask, at the same intake, for a short plain-language summary of the finding for its board or audit committee. It is fixed in scope and fee with the Certificate. It explains the finding; it is not legal advice.
Illustrative finding – not a client engagement
This example is hypothetical. The insurer, its system and every figure below are invented to show what a Kill-Switch Certificate finding looks like. It describes no real firm and no engagement.
The system and the requirement
A motor insurer uses an AI claims-triage system. For claims under a set value, the system either pays the claim or declines it automatically; a declined customer receives a letter, and no person looks at the claim first.
The requirement, named at intake: a rule in force in the insurer's jurisdiction gives a person the right to have a decision based solely on automated processing reviewed by a staff member.
What was examined
- The system's documentation and the claims-handling procedure for automated declines.
- The decline letter sent to customers.
- 50 declined claims, with customers' details removed.
- Three months of review requests and their outcomes.
- Working sessions with the claims operations lead and the model owner.
What met the requirement
- The decline letter tells the customer how to ask for a review by a staff member.
- A claims handler can reverse an automated decline, and the system records who reversed it and when.
- Each review request is logged with its date and outcome.
What did not
- The reviewing handler sees the system's decline code but not the claim details the system used, so the record does not show that the handler could assess the decision rather than confirm it.
- No named role is accountable for the review path; the procedure assigns it to "the claims team". In 7 of the 50 sampled declines, the review request was closed with no reviewer recorded.
- Automated declines can be stopped only by switching off all automated claims decisions; there is no way to suspend declines alone, and the procedure does not say who may switch the system off.
What would need to change
- Give the reviewing handler the claim details and reasons behind each automated decline, and record what the reviewer looked at.
- Name one accountable role for the review path, and record a reviewer for every request.
- Add a way to suspend automated declines on their own, and name who may use it and who may switch off the whole system.
Finding (illustrative): the system's review mechanism does not meet the requirement. It meets it on telling the customer and on recording reversals; it does not on the reviewer's access, on accountability for the review path, or on who could stop the decisions.
This example is hypothetical and describes no real firm or engagement.
Start an engagement
Write to protocol@fijishi.com with the AI system, the requirement it must meet and any deadline. We reply within 5 business days. Start an engagement