The parts of correspondence AI shouldn't touch

The approval chain in ship design has a specific structure. Class approves the drawing against the rules. The design engineer certifies that the drawing represents their technical judgment and is correct as issued. These are not the same act. When something fails — a structural connection, a watertight detail, a load assumption that turned out wrong — flag state, P&I, and the courts distinguish between them. Class approval does not transfer liability from the design office. The investigation returns to the engineer's certification, and to the correspondence record that underlies it.

AI is now being applied across the design workflow: structural optimisation routines, rule-based specification generation, document retrieval across project archives, and increasingly, the processing of class comment sheets. Each of these applications sits in a different position relative to that accountability structure. Understanding which side of the line each sits on is the only question that matters for a design office evaluating these tools.

Consider the class comment cycle. A surveyor raises a comment against a drawing. Someone at the design office must read the comment, locate the relevant drawing and its current revision, identify the applicable rule, determine whether the design already complies or whether a change is required, draft a disposition, and issue it formally. This is a sequence of sub-tasks. Most of them are retrieval and synthesis tasks. One of them is a professional judgment issued under the engineer's name.

AI can perform the retrieval and synthesis steps accurately and quickly. It can extract the comment text, retrieve the correct drawing revision from the document management system, cross-reference the applicable society rule, and produce a draft disposition that is technically coherent with the design intent. In straightforward cases — a standard deck penetration, a familiar bracket configuration — the draft will often be substantively correct on the first attempt. The preparation time collapses.

None of that constitutes issuing the disposition. The engineer still reads the draft, evaluates whether the reasoning is sound, confirms the rule reference is correctly applied to this specific configuration in this specific vessel context, and then signs off. That final act cannot be delegated to any system that cannot hold a professional licence, cannot be named on a drawing, and cannot be the subject of a professional misconduct investigation.

This is not a capability argument. The distinction does not rest on whether AI systems are accurate enough to be trusted. It rests on the structure of professional liability in regulated engineering. Classification societies approve drawings on the basis of the responsible engineer's certification. That certification implies a human being made a judgment, accepted responsibility for it, and is traceable. An AI system is none of those things. It is a tool. A very capable tool, but a tool. A torque wrench does not sign off on a bolted connection.

The mistake — and it is a mistake that cuts in both directions — is to conflate these two halves. The first error is refusing AI assistance in the preparation phase on the grounds that AI cannot sign off. That is not what anyone is asking it to do. If a system can reliably retrieve the right drawing and draft a technically grounded response in ten seconds, the engineer who then reviews and approves that draft has more time and cleaner inputs for the judgment they actually need to make. The preparation quality improves. The sign-off is still theirs.

The second error is more serious: allowing AI output to flow into issued correspondence without a genuine review step. This happens not through deliberate policy but through workflow pressure. The draft looks right. The project is behind. The engineer approves it quickly. The problem is that "looks right" and "is right for this vessel, this rule interpretation, this specific configuration" are different things, and only a qualified engineer who has actually read the drawing and the rule can collapse that gap. Shortcuts in the review step do not move liability away from the stamp — they just undermine the basis on which it was given.

Design offices that are thinking clearly about these tools should be drawing the line explicitly, not implicitly. The preparation phase — retrieval, cross-referencing, drafting — is where AI belongs and where it creates genuine value. The sign-off phase is where the engineer belongs and where professional standards require them to remain. That line is not arbitrary and it is not conservative resistance to new technology. It is the correct reading of how regulatory approval actually works in this industry.

There is a clear and defensible position available here. It does not require apologising for using AI, and it does not require pretending that AI output and engineering judgment are the same thing. They are not, and the accountability structure of the design office depends on keeping them separate.