ClassNK Stops Auditing AI Long Enough to Use Some

ClassNK has launched Survey Compass, an AI assistant that answers natural-language questions about classification survey requirements — and after a week in which the same society appeared three times as an auditor of other people's AI, it is worth noticing that it has now put a model inside its own core process. What it does: users ask a question in ordinary language; the system searches survey-related documentation and returns a summary with links back to the source material. The initial release covers Part B of the Rules for the Survey and Construction of Steel Ships, the Guidance for Undergoing Surveys, and Technical Information. During beta ClassNK plans to extend it to international conventions and flag State requirements, and to integrate vessel-specific data. It is reached through the ClassNK Customer Hub – Ship in Service portal, available to shipowners and management companies with existing NK-SHIPS accounts. The society's stated reason is the one every superintendent will recognise: the growing volume and complexity of requirements, particularly in environmental compliance. The design decision worth copying is the citation. A summary with a link to the underlying rule is a fundamentally different product from a summary alone, because it lets the user do the one thing that makes an AI answer safe to act on — check it. Retrieval with sources is not a novel technique, but it is still the exception rather than the rule in maritime software, and its absence is why most shipboard AI assistants cannot be trusted with anything consequential. Why this matters for ship managers: survey and certification knowledge is concentrated in a small number of experienced people, most of them close to retirement, and the volume of applicable requirement has grown faster than the population who can navigate it. A retrieval assistant does not fix that, but it changes the cost of a first-pass answer from a phone call to a query. What to do now: if you hold NK-SHIPS accounts, put it in front of a superintendent with five questions you already know the answer to. That is the only useful acceptance test for a system like this — not whether it answers, but whether it is wrong in ways you can catch. Then look at whether your own procedures, circulars and PMS documentation could be retrieved the same way, because the technique transfers and the licence is not the hard part. The honest caveat: this is a beta with a deliberately narrow corpus — three document sets, one society, one portal. No accuracy rate, error rate, evaluation methodology, user numbers or pricing beyond existing accounts has been published. Retrieval systems fail in a specific and dangerous way: they return a confident summary of a rule that does not apply to your vessel, and the link at the bottom is only protective if somebody clicks it. Treat it as a faster way to find the rule, never as a substitute for reading it. [Smart Maritime Network]

This story is part of the Maritime AI Digest of 30 August 2026.