Why Fijishi

The question

When an AI system decides a claim, a price, an application or a fraud flag, someone may later ask three things about that one decision: which rule did it meet, who was accountable for it, and who could have stopped it. The person asking may be a regulator, a court, a customer's lawyer, an auditor or a reinsurer. More and more, the answer has to be evidence about the decision itself, not a policy document.

Why now

The Fijishi Jurisdiction Index 2026 examines AI in insurance decisions in 38 entries across 34 jurisdictions. In 12 entries, AI-specific rules are already specified and enforced. In 14 entries, the binding rule reaches reinsurers directly. Each entry ends with one sourced question that the public record does not yet answer. The Regulatory Deadline Calendar lists, with sources and dates, the rules in force, the rules coming and the drafts still pending.

The market is moving too. Since January 2026, the standard US general liability policy forms have included optional endorsements that exclude generative-AI exposures (trade association report). In the courts, class actions over AI used in health-insurance claim denials are proceeding, and in March 2026 a US federal court ordered an insurer to produce documents about its algorithm (legal report). In the European Union, the new Product Liability Directive treats software, including AI, as a product for products placed on the market after 9 December 2026, so its makers can be held liable for damage it causes without proof of fault (Directive (EU) 2024/2853). These are the three questions a Fijishi finding answers - which rule applied, who was accountable, and who could stop the decision - with dated, independent evidence.

What already exists, and what each covers

A great deal of serious work already exists, and a firm may need several pieces of it. Each covers something real, and each describes its own scope:

Management-system standards – ISO/IEC 42001, with ISO/IEC 23894 on AI risk management and ISO/IEC 42005 on AI impact assessment, certified by accredited bodies under ISO/IEC 42006. A certificate shows that an organisation manages AI through a documented system that operates as documented.
Risk frameworks and principles – the NIST AI Risk Management Framework and the OECD AI Principles guide how organisations identify and manage AI risk. They are voluntary and are not certifications.
Harmonised standards – CEN and CENELEC are developing the European standards that will support the requirements of the EU AI Act.
AI system and agent assurance – AIUC-1 for AI agents, Singapore's AI Verify testing framework, the IEEE 7000 series and IEEE CertifAIEd, and testing by certification bodies examine how a system behaves or how it was designed, against their own criteria.
Algorithmic and ethics audits – schemes such as ForHumanity's, and audit firms such as BABL AI, audit systems against defined criteria, often for a specific law such as New York City's bias-audit rule.
Governance software – platforms such as Credo AI and Holistic AI help firms run AI inventories, policies and assessments.
Public AI safety and security institutes – such as the UK AI Security Institute and the US Center for AI Standards and Innovation – evaluate the capabilities and risks of advanced models.
AI insurance – insurers that cover AI performance and liability take on part of the financial risk.

What Fijishi answers

Fijishi answers the question that comes after these: for one named AI decision, in one jurisdiction, does it meet the rule that binds it – and can the firm show who was accountable and who could stop it, in a way that would hold up before a regulator, a court or an insurer?

A firm can hold a management-system certificate, run a risk programme and test its models, and still be asked that question about a single decision. A Fijishi finding is built to sit alongside that work. It uses the firm's existing records as evidence, names the rule it measures against, and states the result in words.

Fijishi does not replace any of the work above, and does not compete with a firm's own auditors or counsel.

What a firm gets from a Fijishi finding

Independent evidence of whether one AI decision meets the rule that binds it, from someone with nothing to gain:

Time	A Kill-Switch Certificate is scoped for 1-2 weeks once system documentation is provided. We do not ask for a model's source code or weights. A readiness finding can be made against a rule that is adopted but not yet in application.
Cost	One fee, fixed in writing before any work begins, for the decision examined. No subscription, and no follow-on work tied to the finding. Intake has no fee, and the Index and the Calendar are free to read.
The rule, and its date	The Index and the Calendar show which rule reaches a decision, and from when. The finding is measured against that rule, named at intake.
Who can stop it	Who can stop this decision, and can the firm show it: a named role, a live test, and one decision reconstructed from the firm's own record.
If a decision is challenged	Every finding points to the evidence behind it. Fijishi Juris supports counsel when an AI decision is disputed.
Something to show others	The client owns the finding and may share it whole - with its board, its reinsurer, its auditors or its customers. Anyone holding the certificate number can check that it is current.
Nothing to gain	Fijishi sells no insurance, builds or sells no AI, and does no consulting work, so it has nothing to gain from what a finding says.
Clear limits	Each certificate says what it does not cover, so no one reads it as more than it is.

How certification works · Example certificate · Who a Fijishi finding is for

How we work

Every finding names the rule or standard it measures against.
The fee is fixed at intake, before any work begins, and does not depend on the finding.
Where an engagement produces a finding, the client owns it.
Fijishi does not certify a system it helped design or build, and does not accept a fee from the vendor whose contract it reads.
Our Index method and our draft criteria are published, corrections are public and dated, and companies named in the Index have a right of reply.
See an illustrative finding (hypothetical; not a client engagement).

Who does the work

Fijishi is led by its two founders, Akhil Sharma and Preethi Sharma, and every finding is examined and signed by both of them. Both have worked on decision systems in regulated financial services, Akhil since 2010 and Preethi since 2015, including systems that used AI models, and on the controls, compliance and audit around them (Founders); Fijishi never examines a system that either founder worked on as a consultant. There is no hand-off to junior staff and no sales layer between a firm and the people doing the work; the fee pays for their time. Fijishi does not design, build, sell or implement AI systems, so no other part of its business depends on what a finding says. We take a limited number of engagements at a time, so that both founders work on each one, and we aim to start within two weeks of agreed scope.

How our draft criteria are developed

Certification against named rules comes first: it is how criteria are tested on real systems. Alongside it, we publish our own draft criteria and develop them in public – a working draft, then public comment, then a numbered version that can be tested in a certification, then use. Early testing laboratories and rating agencies became references others rely on in the same way: by measuring, publishing and correcting in public, until their work was used.

Our drafts are working drafts, and none is a standard. The first, on who can stop an AI decision, is open for public comment until 15 January 2027. A draft becomes a standard only when the people who would use it, and those who comment on it, find that it holds - not because we say so. Anyone may comment, and no one can pay to change a draft.