Maritime AI Digest — 8 November 2026

Everyone talks about AI agents. Ask ten people in a shipping office what an agent actually is, and most will not be sure. That is normal. The word is new, and most of what we read about agents is a launch, not a working day. This week we finally have a working day. A software company showed what its agent did, day by day, on a real voyage across the Atlantic: what it noticed, what it suggested, and what the operator said yes to. We open with that story because it explains the idea better than any definition. Then a satellite that decides for itself, a ship that will steer itself on a canal, an approved weather router, a bank that runs its AI on its own computers, and ports that use AI but rarely measure it. We finish with simple steps to give an AI agent its first job in your company.

The week's most important developments in shipping & oceans — distilled into a 5-minute read.

🔗 Quick Links

• 🤖 What an AI agent actually did on a 16-day voyage across the Atlantic — Maritime Innovations

  ZeroNorth's agent spotted a 29-mile drift and fuel running low, and prepared new plans. The operator approved every change. No savings figure, and the ship is not named

• 🏗️ Ports use AI widely, but only 11% measure what it gives back — BCG

  A survey of 103 port and terminal experts. 44% are scaling AI. Only 4% can show it in earnings. The top problem is not knowing the return

• 💸 Revolut cuts its AI coding bill by about 90% by running AI on its own computers — Cybernews

  2,000 engineers moved to an open model on 56 company chips. The saving leaves out some costs. The same idea could bring AI on board without a link

• 🚢 Two Chinese canal container ships will try to sail themselves for a third of the route — Smart Maritime Network

  Jiangsu Ocean Shipping sets targets before the trial starts: 90% of objects recognised, at least one third of 80 km sailed without a person steering

• 🌦️ Bureau Veritas approves an AI weather router that must stay inside hard limits — Cyprus Shipping News

  VesselFront's routes can never break a maximum wave height or enter a no-go area. The approval covers how it calculates, not how much fuel it saves

• 🛰️ A satellite spots a ship and decides by itself to look again — Smart Maritime Network

  Ubotica's satellite found a cruise ship, booked a second photo on its own, and two days later saw the ship had gone. No person on the ground told it to

• 🥾 A chatbot said 8 hours. The climb took 16 — Straight Arrow News

  Hikers trusted confident AI answers and needed rescue. The same risk exists when crew ask a chatbot about a procedure or a port

• 🧭 AiatSea: how to give an AI agent its first job — AI at Sea

  Agents booked only four Uber rides in a busy week. Not because they fail, but because most people do not know how to use one yet. Five simple steps to start

🚀 Big Moves This Week

1. What an AI Agent Actually Did on a 16-Day Voyage

Most of what we read about AI agents is a launch. This is the first account we have seen of an agent's working day at sea, so it is worth reading slowly.

First, what is an agent? A normal AI tool answers when you ask it something. An agent keeps watching, notices when something changes, works out what to do, and prepares the next step. Then a person says yes or no.

The voyage. ZeroNorth, a Copenhagen company that makes voyage and fuel optimisation software, showed journalists what its agent, Propel, did on a mid-size bulk carrier sailing from Valletta, Malta, to Bridgetown, Barbados, about 4,500 nautical miles. The ship left on 17 September and the voyage took 16 days.

Day by day. For the first ten days nothing happened. The agent watched position, weather, fuel use, bunker prices and freight rates, and confirmed the plan still held. On day 12 the ship drifted up to 29.1 nautical miles off its planned track. The agent spotted it, calculated a new route and the fuel cost of the detour, and sent it to the operator. Fuel use was also running above forecast and would have eaten into the 5% fuel safety margin, so it suggested a slightly lower speed. On the final approach the trade winds got stronger, and it prepared a new course and speed. The ship arrived safely with its fuel reserve intact. The operator reviewed and approved every route change, and could ask the master what had happened.

What is missing. The ship, owner and charterer are not named. No fuel, time or cost savings are given for this voyage. It is ZeroNorth's own account, presented at its headquarters on 29 October, and nobody outside has checked it. Propel has been in early access since August, with Cargill, Ultrabulk and CMB.TECH named as early partners, but the article does not say any of them were on this voyage.

Why it is still useful. It shows the shape of agent work. Most of the time the agent just watches. When something changes, it does the boring calculation fast and puts a ready answer in front of a person. That is exactly the job a busy operator or superintendent has no time for at 0200.

What to ask if you are offered one. What can it do alone, and what needs a yes? Who gets the alert, and how fast? And after ten voyages, can you show me what it caught that we would have missed?

Disclosure: ZeroNorth says Propel runs on Claude, made by Anthropic. We use Claude to help produce this newsletter. [Maritime Innovations · AiatSea, 18 October]

2. Ports Use AI Widely. Few Measure What It Gives Back.

A port that buys a new crane knows how many moves per hour it adds. A port that adds AI often does not. That is the main finding of a new survey, and it applies to ship managers too.

The survey. In August 2026, the consultancy BCG and TIC4.0, an industry group of terminal operators and suppliers that works on shared data standards, asked 103 experts on five continents about AI. They came from terminal operators, port authorities, container lines and port suppliers.

The numbers. 44% say their organisation has moved past testing and is scaling AI. But only 11% measure, in a regular way, what AI adds to efficiency or margin. And only 4% can point to a gain in earnings. The biggest problem named was not knowing the return (41%), ahead of data quality (32%) and joining systems together (30%). Among 43 uses, reading data out of shipping documents and predicting equipment failures were among the most common. Only 18% said customers had a better experience.

How to read it. This does not say AI fails in ports. It says most ports do not measure it. If you do not count the moves before and after, you cannot show a gain, even when there is one. BCG's advice is simple: prove the value on one job before you spread it everywhere, and fix your data first.

Our caveats. It is a survey of opinions, not audited results. BCG sells consulting on exactly this problem. And the survey was published in late October. We include it now because it fits this week's question so well.

Why a ship manager should care. Ship management has the same gap. Many offices now use AI for emails, reports or document checks, and almost nobody writes down the time it took before.

What to do. Before you start any AI tool, write down one number for the job it will do: hours per week, errors per month, or days to close. Check the same number after three months. That one line is the difference between "we use AI" and "AI saves us this much". [BCG · Smart Maritime Network · WorldCargo News]

3. A Bank Cuts Its AI Bill by About 90% by Running AI on Its Own Computers

For thirteen weeks we have asked who pays for the AI tools used in shipping offices. This week a large company gave a clear answer to a related question: what happens if you stop renting AI and run it yourself?

What Revolut did. Revolut, the European digital bank, moved about 2,000 engineers from paid coding assistants, Anthropic's Claude Code and OpenAI's Codex, to an open model called GLM 5.3. An open model is one anyone can download and run on their own computers. This one is made by the Chinese company Z.ai. Revolut runs it on 56 Nvidia chips in its own setup. It says its AI coding costs fell by about 90% in less than a year, to about 30 US cents per million tokens (a token is a small piece of text, roughly three quarters of a word). It also says coding time improved by 60%. The paid tools are still available, but only for hard jobs.

The honest limits. The 30-cent figure leaves out cloud costs, staff and monitoring. One outside analysis says the real saving is smaller than ten times, and the comparison mixes two different years of use. The model works with text only, so jobs with images still go to paid services. Revolut also had to tune it heavily. And the productivity figures are Revolut's own.

Why this matters at sea. We have written for weeks about AI on board when the satellite link is down. A ship cannot rent AI from a data centre it cannot reach. Revolut shows that a company can run a strong open model on its own hardware, at scale, for daily work. A ship is much smaller than a bank, and a ship's computer is much smaller than 56 chips. But smaller open models are improving fast, and the idea is the same: AI that lives where the work is.

What to ask. For your office: which AI tasks are simple and repeated enough to run on a cheaper open model? For your ship suppliers: does your AI work with no link, on hardware on board, and what does it need?

Disclosure: we use Claude, made by Anthropic, to help produce this newsletter. Revolut moved most of its coding work away from Claude. [Cybernews · AI Chat Daily · AiatSea, 25 October]

4. Two Canal Container Ships Will Try to Sail Themselves, and the Targets Are Public

Most autonomy trials tell you the results afterwards, if at all. This one says what it wants to achieve before it starts. That makes it easy to check later.

The ships and the route. Jiangsu Ocean Shipping in China is modifying two 120 TEU electric container ships, the Jiangyuan Jinling and the Jiangyuan Qinhuai, for trials on an 80 km stretch of the North Jiangsu Canal between Huai'an and Suqian.

What goes on board. Radar, AIS and high-definition cameras to see the canal around the ship. Separate settings for straight sections, crossings, sharp bends and bridges. One central system that sets speed, route and power together, and also manages the engine room and energy use. And a digital twin, a live computer copy of the ship and route, that connects ship and shore for voyage planning, risk alerts, lock bookings and port coordination.

The targets. The cameras and sensors must recognise objects with at least 90% accuracy. And the ships should sail themselves for at least one third of the distance on the trial section.

What is missing. No start date, no partners, no budget. We are not told whether a crew stays on board, or who takes over when the system is not sure. These are targets, not results.

Why it matters. Canals and rivers are where self-steering ships make sense first: short routes, known water, slow speeds and lots of repeated sailing. In September we wrote about a Dutch inland autopilot running on 200 vessels. Now China is testing the same idea with electric ships.

What to watch. One third of the route is a modest goal, and an honest one. The useful questions come later: which third, the easy straight parts or the bends and bridges? And how often did a person have to take over? [Smart Maritime Network · AiatSea, 20 September]

5. An AI Weather Router Gets a Class Approval. Read What It Covers.

Weather routing is one of the oldest jobs where computers help ships. Now the software is starting to use AI agents, and the first approvals are arriving.

The approval. VesselFront, a software company in Piraeus, says its VesselFront.AI system has received Type Approval from Bureau Veritas. The approval covers three things: how it calculates fuel consumption, how it forecasts emissions, and how it supports weather routing decisions. Bureau Veritas also certified the company's quality system in Piraeus.

How it works. Several small AI agents each watch one thing: the weather, fuel use, environmental rules, and how exposed the ship is to bad seas. Every route must stay inside hard safety limits that the AI cannot break: a maximum wave height along the route, northern and southern limits, and no-go areas. The company has applied for a US patent.

What the approval does not cover. It checks how the software calculates. It does not prove how much fuel it saves on real ships. No customers, no fleet size and no savings are published. As with almost every type approval, the company asked for the review and paid for it.

The same idea in leisure sailing. Two weeks ago PredictWind, used by more than a million sailors, added an AI that does not make its own forecast. It reads the routes from eight weather models and explains them in plain words, including where the forecasts disagree. That last part is useful for any ship: an honest router should tell you when it is not sure.

Why the hard limits matter. The most important part of this story is not the AI. It is the rule that the AI must stay inside limits a person sets. That is how AI should work on any bridge.

What to ask your routing supplier. Which limits can the master set, and can the system ever cross them? Does it show how sure it is about each route? And can you show results on ships like ours, measured by someone other than you? [Cyprus Shipping News · Marine Business News]

6. A Satellite Spots a Ship and Decides by Itself to Look Again

Satellites usually work like a camera on a timer. People on the ground tell them where to look, they take the picture, and they send it down. Then people decide what to photograph next. This week a satellite made that decision itself.

What happened. The Irish company Ubotica says its CogniSAT-6 satellite used AI on board to find a cruise ship in a picture over the Pacific Ocean. The satellite then worked out when it could next see the same spot, and booked a second picture on its own. Two days later it took that picture, which showed the ship had left. It did all of this without instructions from the ground. Ubotica calls the idea Autonomous Scheduling.

What it could lead to. One satellite could spot something and tell another satellite to take a closer look, without waiting for people on the ground. Ubotica says this could help watch ships that switch off their AIS, the shadow fleet, and undersea cables.

The honest limits. It went back to the place, not to the ship. This is not tracking a moving vessel. It is one demonstration, reported by the company. Ubotica's main customers are governments and maritime authorities. We covered its $11m funding in June.

Why it is in this issue. It is the same pattern as story 1, in space. The machine notices something, decides what to check next, and acts. The difference is that here no person approved the second picture. For a photo, that is fine. For a route change on a ship, it is exactly the step that should still need a yes.

What it means for ship managers. More and more of your ships' movements will be seen and checked by machines, quickly and without a person asking. Keep your AIS on and your position reports honest. The space above you is getting smarter. [Smart Maritime Network · Splash247]

7. A Chatbot Said 8 Hours. The Climb Took 16.

This story is not from shipping. It is about a mountain. But anyone who has asked a chatbot a quick question at work should read it.

What happened. In early September, three young hikers in California used Google's Gemini chatbot to plan a climb of Mount Shasta, a 14,000-foot volcano. The chatbot said the climb would take about 8 hours, so they took little food and water. It took 16 hours. They reached the top at 7 in the evening, came down in the dark, one of them was injured, and they were rescued the next morning. The sheriff's office told climbers never to rely only on AI to plan a trip.

Not the only case. In July, two hikers in Poland's Tatra mountains followed ChatGPT's advice onto climbing terrain and had to be lifted out by helicopter. Rangers on Mount Shasta report other cases, including a visitor who camped next to a "No Camping" sign because AI said it was allowed.

Why it happens. Researchers point to one simple thing. A person who is not sure hesitates, and we notice. A chatbot has nothing built in that shows doubt. A wrong answer sounds exactly as sure as a right one.

Why it matters at sea. Crew and office staff already ask chatbots about port rules, procedures, weather, equipment and regulations. Most answers will be fine. A few will be confidently wrong, and they will look exactly like the good ones.

What to do. Use chatbots to start the work, not to finish it. For anything about safety, rules or a port, ask the chatbot for its source and check it. Teach crew one habit: if the answer matters, confirm it with a person or an official document. And when you choose AI tools for work, prefer the ones that show where each answer came from. [Straight Arrow News · Engadget · High Country News]

8. AiatSea: How to Give an AI Agent Its First Job

Here is a number from this week that surprised a lot of people. Uber handled about 320 million rides and food orders in Halloween week, its busiest week of October. AI agents booked four of the rides, and about 500 food orders. Its finance chief called it "very early". DoorDash reports only a few hundred orders from agents so far.

Our reading. This does not mean agents do not work. Story 1 shows one doing useful work on a real voyage. We think the reason is simpler: most people do not yet know what an agent is, or how to give one a job. Nobody showed them. That is true in shipping offices too.

The difference between a chatbot and an agent. A chatbot waits for your question. An agent has a job. It watches something, notices a change, prepares the next step, and asks you before anything important happens. ZeroNorth's agent watched a voyage for 16 days and spoke up three times. Ubotica's satellite saw a ship and planned its own next photo. That is the idea.

Five steps to give an AI agent its first job. One: pick one boring job that repeats every day or every week, such as checking noon reports, watching fuel against plan, or comparing a supplier's invoice with the order. Two: write down what the agent may do alone, such as read, compare and warn, and what always needs a person's yes, such as send, order, pay or change a route. Three: write down today's number for that job: hours per week, errors per month, or problems found late. Story 2 shows that almost nobody does this. Four: let the agent run beside the person who does the job now, for one month, without replacing anyone. Five: after the month, sit down with that person and count. What did the agent catch? What did it get wrong? Did it save time or add work?

Two rules from this week. From story 5: give the agent hard limits it can never cross. From story 7: when the answer matters, check where it came from.

Our view. Agents will not arrive in shipping as one big system. They will arrive one small job at a time. The companies that learn how to give an agent a job, and how to check it, will be ready when the bigger tools come. The first job can start this month. [Maritime Innovations · FourWeekMBA · Uber · AiatSea, 4 October]

📊 Why It Matters — Strategic Impact Table

🔭 On Our Radar

• 🗳️ Who Is Actually Paying for Shipping's AI Accounts? — carried for a thirteenth week: Anthropic may start selling its shares to investors from the week of 9 November, and this week Revolut showed that running an open model yourself can cost far less than renting, we monitor whether AI prices change for business customers and whether any ship manager publishes a policy on paid AI tools, and our reader survey is still open (we use Claude, made by Anthropic, to help produce this newsletter).

• 🤖 Will ZeroNorth Name the Ship and Show the Savings? — the voyage account has no ship name and no savings figure, we monitor whether ZeroNorth or a customer publishes results from several voyages, ideally checked by someone outside the company.

• 🏗️ Does Any Port Publish What Its AI Gives Back? — only 11% of port experts say they measure it, we monitor whether a terminal or port authority publishes a before and after number for one AI job.

• 🚢 Do the Canal Ships Hit Their Targets? — Jiangsu Ocean Shipping wants 90% object recognition and one third of the route sailed by the system, we monitor when trials start, which parts of the canal the system handles, and how often a person takes over.

• 🚁 What Does Toei Kisen's Drone Trial Show? — carried from last week: the hull inspection trial is due to start in November, we monitor how long it takes compared with the old way and whether class accepts the findings.

• ⏱️ Does the STEER Project Find the Same Problem With AI? — the first STEER report, from The Nautical Institute and Lloyd's Register Foundation, looked at 471 accidents and found ship systems were more often misused or ignored than broken, we monitor its next phase, which will look at AI and remote monitoring.

• 🇸🇬 What Does Singapore's New Law Say About AI? — the Bill that gives the port authority power over AI, autonomous vessels and drones goes to its second reading in Parliament this month, we monitor what the debate adds and when the detailed rules follow.

• 🔐 Will AI-Powered Attacks Reach Ports? — the UK's National Cyber Security Centre now says an AI-powered ransomware attack on British critical infrastructure is highly likely by 2028, mostly through old weaknesses, we monitor whether port and shipping regulators issue similar warnings or guidance.

📅 Critical Maritime AI Research Areas for Managers

1. 🤖 What Voyage Agents Really Catch: one account shows an agent spotting a drift and fuel running low. Research should compare agent alerts with operator decisions over many voyages, and count what was caught early, what was missed and what was false alarm.
2. 🏗️ Simple Measures for AI Value: few ports measure what AI gives back. Research should define one or two easy before and after measures for common AI jobs in ports and ship management, such as document checks and failure prediction.
3. 💸 Open AI Models on Board: a large company now runs an open model on its own hardware for daily work. Research should test which open models run well on shipboard computers with no link, for which tasks, and at what cost.
4. 🚢 Autonomy Targets on Inland Waterways: canal trials now publish targets. Research should compare system performance on straight sections, bends, bridges and locks, and record every handover to a person.
5. 🌦️ Hard Limits for AI Routing: approved routing AI must stay inside wave and no-go limits. Research should test how routing systems behave close to those limits, and how clearly they show the master how sure they are.
6. 🥾 Chatbot Errors on Maritime Questions: chatbots give confident wrong answers. Research should test common chatbots on port rules, procedures and regulations, and measure how often they are wrong and whether they show sources.
7. 🧭 First Agent Jobs in Ship Management: most people do not know how to give an agent a job. Research should run small trials of agents on repeated office tasks, with a person alongside, and publish what worked and what added work.

📈 Top Investment Opportunities

1. 🤖 Voyage Agents With a Person in Charge — an agent watched a full Atlantic voyage and prepared every route change for approval — the opportunity is in agents for operators and charterers that watch voyages around the clock, explain each alert in plain words, and keep a clear record of what a person approved — ZeroNorth · Propel
2. 🏗️ Measuring What AI Gives Back — only 11% of port experts measure AI's value — the opportunity is in simple tools and services that record a before and after number for each AI job in ports and ship management, so buyers can see the gain — BCG · TIC4.0
3. 💸 Open AI Models That Run Where the Work Is — a bank cut its AI coding bill by about 90% by running an open model itself — the opportunity is in small, tuned open models and hardware that run on board ships and in offices without renting AI from a distant data centre — Revolut · GLM
4. 🚢 Self-Steering Systems for Inland Ships — electric canal container ships are being fitted to sail themselves for part of the route — the opportunity is in sensors, control systems and shore links for short, repeated inland routes, where autonomy makes sense first — Jiangsu Ocean Shipping
5. 🛰️ Satellites That Decide What to Look At — a satellite spotted a ship and booked its own second look — the opportunity is in onboard AI for satellites that cuts the time between seeing something at sea and checking it, for authorities, insurers and cable owners — Ubotica

📅 Top Monthly Picks

1. 🤖 An AI Agent Can Now Book Freight, and the Limits Are Written Down — Flexport opened its booking system to AI agents with clear rules on what they may do alone; this week an agent watched a whole voyage and a person approved every change, which is the same idea at sea — Splash247
2. 💰 The AI You Rent Sits on Very Large Bets — Anthropic's draft filing showed huge fixed costs and a shortage of computing power behind the AI many offices rent; this week Revolut showed what happens when a company stops renting and runs AI itself — VKTR
3. 🧭 An Independent Test for a Robot Navigator, and Nobody Shows the Scores — an outside lab tested a navigation AI on 55 traffic scenarios and kept the results private; this week Chinese canal ships published their targets before the trial even started — Digital Ship
4. 🇰🇷 South Korea Wants a Permanent Test Area for Self-Driving Ships — Korea plans a fixed place to keep testing autonomy off Geoje; this week China started fitting two canal ships for the same kind of trial — Splash247
5. ⏱️ 72% of Seafarers Have No Time to Learn New Systems — a study of 532 seafarers found the missing piece is time to learn, not willingness; this week agents booked almost no Uber rides, and the same lesson applies: people need to be shown how to use new tools — Splash247

Have feedback or a story tip? Email us or connect on LinkedIn.