AiatSea: What May the Agent Do Without Asking?

For two years, AI in shipping mostly gave advice. It wrote a draft, suggested a route, summarised a report. A person then decided. That is changing. This week an AI agent could book freight. In the Marcura survey, 72% said their company uses, tests or plans AI that takes action. Only 15% said their rules are ready.

Shipping already knows how to handle this. We do it for people every day. A purchaser can approve an order up to a set amount. Above that, the superintendent signs. Above that, the technical manager. A junior officer keeps the watch, and calls the master when the standing orders say so. Nobody finds this strange. It is simply written down.

Almost no company has written the same thing down for AI. So here is a start. It is three short lists.

List one: the agent may do this alone. Read, search, sort and summarise. Write a draft. Prepare a form. Remind a person. Nothing leaves the company and nothing is changed in any system.

List two: the agent may do this after a named person says yes. Send an email outside the company. Place an order or make a booking under a set amount. Change a record in the maintenance or crewing system. The person who says yes is responsible, the same as if they had done it by hand.

List three: the agent may never do this. Anything that changes how the ship is navigated or operated. Signing or accepting a contract. A payment above the limit. Sending crew personal or medical data to an outside tool.

One rule under all three. Every action leaves a record: what the agent did, what data it used, and who approved it. If you cannot find that record a month later, the tool is not ready.

This matches what people asked for this week. In the survey, 80% want a named person to stay responsible. On the chartering panel, Fednav's Jesper Klarup said the tool can never become your alibi. And Flexport published the same shape of limits for its booking agent.

What to do on Monday. Take one tool your office already uses. Write its three lists on one page. Show it to the people who use the tool, and ask what is missing. You will have a better AI policy in one hour than most companies have after a year of meetings.

Disclosure: this newsletter is produced with the help of Claude, an AI model made by Anthropic. It sits in list one. It reads and drafts. A person checks the sources and decides what is published. [Splash247 · survey · Smart Maritime Network · Flexport]

This story is part of the Maritime AI Digest of 4 October 2026.