Fijishi Certification
Certification of AI decisions against a named standard.
What Fijishi certifies
Fijishi Certification asks one question of one named AI system: can the firm show which rule the system's decisions met, who was accountable and who could stop it? The answer comes from the firm's own records, with a live test of the stop where one can be agreed, and is measured against a named standard, not opinion.
Decisions we examine
- Underwriting – whether to offer cover, and on what terms.
- Pricing – the premium a customer is quoted.
- Reserving – the amounts set aside for claims not yet paid.
- Claims – whether a claim is paid, how much, and how fast.
- Fraud detection – which claims or applications are flagged for investigation.
- Customer contact – what customers are told and offered, including by chatbots and other AI agents.
- Credit – whether a loan or credit limit is approved, and on what terms.
- Financial crime – which customers or payments are flagged in KYC and anti-money-laundering checks.
How this differs from existing audit practice
Compliance audits examine whether processes are documented and followed. Responsible-AI assessments examine whether an organisation's stated principles are reflected in its governance. Technical assurance tests whether a system works. Each has real value. Fijishi asks a narrower question that comes after them: for one named system and one named rule, does the record show that the system's decisions meet that rule, who was accountable and who could stop them, in a form that can be put before a regulator, a court or an insurer?
It is offered where that question sits outside the scope of the audits already in place – not because those audits have failed.
The two certifications
| Feature | Kill-Switch Certificate | Baseline Review |
|---|---|---|
| Scope | One named AI system, one named override or human-review requirement | One named AI system, one named standard in force |
| Question answered | Does the system's override or human-review mechanism meet this requirement? | Does the system meet the named standard, criterion by criterion? |
| What it covers | How the override or human review is invoked, by whom, and what record it produces | Override authority, audit-trail completeness and data-handling boundaries, as far as the named standard addresses them |
| Finding | Whether the mechanism meets the requirement and, where it does not, what would need to change | A criterion-by-criterion finding: where the system stands against the standard, and what is missing |
| Turnaround | Scoped for 1-2 weeks once system documentation is provided; confirmed at intake | Typically three to six weeks from agreed scope to finding; confirmed at intake |
| Fee | Fixed at intake, before any work begins | Fixed at intake, before any work begins |
| Best suited to | One named requirement that applies now - for example, a human-review requirement in force for decisions based solely on automated processing - or a published draft, or a rule adopted but not yet in application, as a readiness finding. | A broader review ahead of a regulatory examination, a legal proceeding or an insurer's underwriting review |
Not sure which applies? Describe the system and the rule when you write in, and we confirm the right scope at intake.
What the finding answers
For one named AI system, against the requirement named at intake, the finding states in words:
- which rule the system’s decisions have to meet;
- who the record shows is accountable for them;
- who can stop or override them, how quickly, and, where a live test of the stop was agreed, whether it worked and how long it took;
- whether a single decision can be reconstructed from the record;
- where a vendor runs the system, whether the firm keeps the right and the means to stop it;
- what met the requirement, what did not, and what would need to change.
Each finding carries its date and number. See an illustrative finding and the example certificates (hypothetical; not client engagements).
Who can commission it
- The firm that runs the system, for its own governance, ahead of an examination, a renewal or a dispute.
- Its board or audit committee. The engagement letter, with the firm, names the board, the audit committee or its chair as the party Fijishi reports to. A short plain-language summary for the board can be added at intake.
- A reinsurer or capacity provider, on a system a cedant or coverholder uses in business it reinsures or backs, with that firm’s written agreement. The same finding can be used at renewal in place of a questionnaire the firm completes about itself.
- An AI vendor, on its own system, measured against a requirement its customers face. It is not an endorsement of the vendor and not a mark the vendor may display, and Fijishi never reviews, for a buyer, the contract for a product it has certified.
- An investor, on a company it is considering, with that company’s written agreement; the fee does not depend on whether the investment goes ahead (Fijishi Deal Review).
- Brokers and advisers may point a client to Fijishi; the client engages Fijishi directly. No referral fee is paid or accepted, either way.
Whoever commissions the finding is the client and owns it. Where the system belongs to another firm, that firm receives a copy and counts as a Fijishi client, so Fijishi Juris never acts against it (Governance).
Who does the work, and how your material is protected
- Both founders examine the evidence and sign every finding. There is no hand-off to junior staff and no sales layer (Founders).
- Before an engagement is accepted, it is checked against the independence rules and against a register of the systems each founder worked on (How a finding is decided). Fijishi does not certify a system it helped design or build.
- A non-disclosure agreement is signed before any material is shared. Material is used only for that engagement, with customers’ personal details removed wherever the work allows (How we handle your information).
- We do not ask for a model’s source code or weights.
Fee and validity
One fee, fixed in writing at intake, before any work begins. It does not depend on the finding. Intake has no fee. Fijishi does not bill by the hour and offers no open-ended retainers: each engagement produces one finding, and there is no ongoing dependency on Fijishi to interpret or maintain it. A certificate is valid for one year, or until the system or the requirement changes materially, whichever comes first.
Sharing your finding
The client owns the finding and may share it with anyone it chooses – for example its board, auditor, broker, reinsurer or capacity provider, or its underwriters at renewal, including for directors’ and officers’ cover. If the client quotes the finding publicly, it quotes it whole, with its date and the requirement it was measured against. Fijishi does not share a finding with anyone else without the client’s written agreement.
The standard
Each engagement measures against a standard named at intake: a rule or standard that applies to the system, not Fijishi's own view. The Kill-Switch Certificate may also measure against a published draft, or a rule adopted but not yet in application (for example the EU AI Act's high-risk duties, which apply from 2 December 2027), as a readiness finding that says so; the Baseline Review measures only against a standard in force. Fijishi's draft criteria, in Standards in development, are not the standard it certifies against.
A finding is Fijishi's independent assessment: not a certification by, or approval from, any regulator, and not legal advice.
Start an engagement
Write to protocol@fijishi.com with the AI system, the rule it must meet and any deadline. We reply within 5 business days. Start an engagement