AiatSea Analysis: The Ten Questions That Belong in Every Maritime AI Contract

Three institutions wrote down what good looks like this week. A vendor published the boundary its agent will not cross and what it does not do with your data. A classification society published the five capabilities it verified. Another published what a digital model must contain before it will be reviewed. Not one of them published the list a ship manager needs on the desk when the contract arrives — so here is ours. Why a list rather than a principle. Telling people to do due diligence on AI vendors is the same category of advice as telling staff to be careful with confidential information: it moves a judgement onto somebody who has neither the time nor the authority to make it. What works is a short set of questions with a right to a written answer, asked before signature, when you still have leverage. Everything below comes from stories this publication has covered in the last twelve months and from the failure patterns that repeat in them. On evidence — the first three, and the ones most often skipped. One: how many vessels is this running on today, and how many of those are paying? Deployed fleet counts and pilot counts are routinely reported as the same number. Two: show me an outcome, not an installation. Research covering 420 organisations found four in five running pilots and roughly one in ten holding the governance to scale them; the interesting vendors are the ones that can produce a measured result from a named fleet over a stated period, as the Orca AI and NorthStandard study did with 139 vessels and 10.8 million nautical miles. Three: how did this perform through a real disruption, live and timestamped, not in backtest? A model trained on history cannot price a closed strait, and the honest vendors will say so. On data — where the exposure actually lives. Four: is our content used to train your models, and does that answer change by subscription tier? It very often does, and the marketing page describes the enterprise tier while your staff are logged into something else. Five: where is it processed and how long is it retained? For a European operator handling seafarer personal data this is not a technical footnote. Six: what happens to our data if we leave, and in what format do we get it back? Ask this in month one, not in year three when the renewal is being negotiated. On control — what happens when it is wrong, because it will be. Seven: what does the system do on its own, what does it prepare for a human, and who exactly is that human here? Write the name in the procedure before go-live. Eight: when the output is wrong, whose error is it, and can we reconstruct how it got there? Logging is a contractual question, not a feature. Nine: what does it do when its inputs are missing or stale — fail loudly, or produce a confident number from bad data? The second behaviour is the dangerous one, and it is rarely in the demonstration. And one about you. Ten: what is the process this replaces, and have we actually mapped it? A top-twelve carrier spent real money this month building a live model of its own workflows specifically because enterprise AI without operational context underdelivers. Automating a process nobody has traced produces the existing mess faster, with less human sight of it. Why we think this is the right week for it. Because the pattern in this issue is that the guardrail is becoming the product. A vendor now leads with what its agent will not do. A class society now publishes what it did and did not certify. That only happens when buyers start asking, and buyers only start asking when somebody hands them the questions. What to do now: take these ten into your next vendor meeting and ask for written answers. You will learn as much from which ones the vendor cannot answer as from the answers themselves. The honest caveat, applied to ourselves: this is our editorial judgement, not a standard, and it is not legal advice — a checklist compiled by a newsletter is no substitute for your own counsel and your own IT and insurance advisers reviewing an actual contract. It is also incomplete by design; a list somebody will use is more valuable than a list that covers everything. If you think we have missed something that belongs on it, tell us and we will publish the additions with credit. And we are still collecting the underlying data. Our short reader survey on who holds and pays for the AI accounts in shipping, ashore and at sea, is open at aiatsea.com/survey — six questions, two minutes, fully anonymous. We will publish the aggregate free and ungated, and we will say plainly that a self-selecting survey of a maritime AI newsletter's readership over-represents people already interested in AI. [AiatSea, 9 August · AiatSea, 26 July · AiatSea reader survey]

This story is part of the Maritime AI Digest of 16 August 2026.