Maritime AI Digest — 15 November 2026

Two weeks ago we asked whether a robot would ever board a ship. This week one did. A small humanoid robot stood on the deck of an anchor handler in Brazil, next to a crew member, because Petrobras wants to know if robots can one day do the deck jobs that hurt people. It is a first look and nothing more, but it is a real ship and a real client. The rest of the week asks the same question from other sides. An airline shut AI agents out of its bookings. A Korean authority let a self-navigation system onto a passenger ferry. A warehouse company said yes to robots, but only if they pass the tests. And when Microsoft's services failed, the AI assistant went down with the email. Every story is about a door, and about who decides to open it.

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

🔗 Quick Links

• 🤖 A humanoid robot on the deck of an anchor handler working for Petrobras — Solstad

  Petrobras is looking at robots for tensioned cables, Kenter links and shackles. Solstad's Normand Sigma hosted the visit. No robot maker is named and no work is shown yet

• 📦 A warehouse company plans up to 1,000 humanoid robots, but only if they pass the tests — GXO

  GXO and Apptronik will judge each job against agreed safety and output targets before the next step. First real deployment is planned for late 2027

• ✈️ United Airlines bans outside AI agents from booking flights — Skift

  Even an agent that a person supervises is covered. Bookings made by agents can be cancelled without notice. Ports and booking portals may decide the same

• 📧 Microsoft 365 fails, and the AI assistant goes down with the email — Digital Ship

  20 services were hit from 12 November, including Outlook mail, Teams calendars and Copilot. No shipping company has reported damage, but most offices run on the same account

• ⛴️ Korea approves self-navigation on a passenger ferry for the first time — Smart Maritime Network

  Samsung's system can steer and control the engines on the Nami Island ferry Tamnara. We are not told how many crew stay or how much a person supervises

• 🌦️ Weathernews adds an offline mode after masters report links too slow to use — Digital Ship

  The forecast and route stay on screen when the link drops, with 7 to 14 days of weather stored on board. The AI agent still needs a connection

• 🏗️ Shanghai's port group builds a place to test AI before it goes to work — Smart Maritime Network

  54 AI agents are registered for testing in cargo handling and inland transport. No budget, no partners and no finish date are given

• 🧭 AiatSea: who decides when a machine may come in? — AI at Sea

  Petrobras, United, a Korean safety authority and a warehouse company each gave a different answer this week. Five questions to answer before a robot or an agent reaches you

🚀 Big Moves This Week

1. A Humanoid Robot Stands on the Deck of an Anchor Handler

Anyone who has worked on an anchor handler knows where the danger is. It is on the aft deck, close to a wire or chain under tension, with a hammer in your hand and a Kenter link in front of you. This week a robot stood on that deck.

What happened. The Norwegian owner Solstad said its anchor handler Normand Sigma recently supported a Petrobras initiative that is exploring humanoid robots, which means robots shaped like a person with two arms and two legs, for future anchor-handling work offshore. Solstad lists the jobs being looked at: handling tensioned cables, opening Kenter links and shackles (a Kenter link is the joining link that connects two lengths of anchor chain), and also inspection, monitoring and maintenance tasks. The stated aim is better safety and efficiency. Solstad's photo shows a small humanoid robot, about waist high, standing on deck next to a crew member.

The ship. Normand Sigma is an 87-metre anchor handling tug supply vessel built in 2014. It is one of two Solstad anchor handlers that started four-year contracts with Petrobras in early 2026, together worth about $210 million.

What we do not know. Almost everything else. The robot maker is not named. We are not told what the robot did on board, or whether it did anything more than stand on deck. The photo looks like it was taken alongside, not at sea. No next step and no date are given. The only source is Solstad's own short post, and we found no press report. So read this as a first look, not as a robot at work.

Why it still matters. Petrobras does not just talk about robots. In 2023 it spent US$4 million on six four-legged inspection robots for its production ships, after testing one on the FPSO P-74. Two weeks ago we asked whether a robot dog would ever board a ship. An oil company with that record has now put a humanoid on a working anchor handler and named the deck jobs it wants to study.

Our view from the deck. The jobs on that list are the hardest ones for a robot. A wet, moving steel deck. Heavy gear. Strong and exact hand work. A small robot like the one in the photo cannot open a Kenter link today. The honest first jobs are probably looking, measuring and carrying. That is still useful if it keeps one person out of the snap-back zone, the area a broken wire sweeps through.

What to ask. If a client brings a robot to your ship: who is responsible for it on deck? What may it do, and who stops it? Has it been tested on a moving, wet deck? And what happens to the crew's job description when it stays? [Solstad · Baird Maritime · ANYbotics · AiatSea, 1 November]

2. Up to 1,000 Humanoid Robots for Warehouses, but Only If They Pass the Tests

Story 1 shows a robot arriving with no plan we can see. This story shows the opposite: a plan with conditions written down before the robots arrive. It is from warehouses, not ships, but it is the best example this week of how to let a robot in.

What was agreed. On 12 November, GXO Logistics, a large company that runs warehouses for other companies, signed a letter of intent with the robot maker Apptronik. A letter of intent is a written plan to do a deal, not the final contract. It sets a path to up to 1,000 of Apptronik's Apollo humanoid robots over four years. GXO would not buy them. It would pay for them as a service, like renting.

The conditions. This is the interesting part. Each job the robot does will be judged against targets the two companies agree together, for output, safety and day-to-day running. A job moves to the next stage only on measured results. The trials must also prove the safety, security and technical needs for a larger roll-out. Trials in working warehouses start in 2027, and the first real deployment is planned for late 2027 in North America.

What the robots have done so far. The two companies have worked together since 2024. Earlier this year Apollo finished a pilot in the United States, picking small consumer goods by itself. GXO called it a success. Next come picking, moving boxes and crates, packing and sorting.

What is missing. No target numbers are published. No price. "Up to 1,000" is a ceiling, not an order. And the pilot result is GXO's own word, with no figures.

Why a ship manager should read it. A warehouse is the easy case: a flat, dry floor, the same task all day, and a robot that can stop without danger. Even there, a company with years of pilots says first deployment is a year away and every step depends on a test. A ship's deck is much harder. So when someone shows you a robot on board, the right question is not "when does it start?" but "what must it pass first?"

What to take from it. Write the test before the trial. One task. One number for output, one for safety. Who measures it. And what happens if it fails. [GXO · Unite.AI]

3. United Airlines Shuts the Door on Outside AI Agents

Last week we wrote that AI agents booked only four Uber rides in a busy week. This week a large airline decided it does not want them booking at all.

What United did. On 11 November, United Airlines changed its rules. AI agents that United has not approved may no longer book, change or buy travel on its website, its app, or through travel agencies and booking sites. An AI agent is software that does a job for you by itself, such as logging in and booking a ticket. United's new contract text is wide. It counts software that acts by itself even when a person supervises or approves each step. Bookings made or changed by an agent can be cancelled without notice. United can also cancel tickets, remove air miles and close the customer's account. Bookings made before 11 November are safe.

United's reasons. Giving your login to an AI assistant can expose private data. The agent may buy the wrong thing, change your account or spend your air miles. And some steps must be done by a person. At check-in a traveller must confirm that the bags hold nothing dangerous, and an agent clicking "yes" for you is not the same thing.

What is still allowed. Using AI to plan and compare, then booking yourself. Tools for people with disabilities. United says it works with no outside agent today, and wants secure, agreed connections first.

The other view. Critics say United mainly wants to keep customers on its own website and keep out a new middleman. CBS News notes that Southwest Airlines went the other way and now shows its flights inside ChatGPT. Both readings can be true.

Why this matters in shipping. Our industry runs on portals: carrier booking sites, terminal and port systems, vetting and inspection databases, class and flag portals. Most have terms of use nobody reads. If your office lets an AI agent log in with a staff password, you may be breaking those terms, and the account may be closed. Some confirmations in shipping also belong to a person, like a declaration of dangerous goods.

What to do. Never give an AI agent a staff password for an outside portal. Make a short list of the portals your office depends on, and check what their terms say about automated tools. Ask your main suppliers one question: do you allow AI agents, and how should they connect? [Skift · CBS News · AiatSea, 8 November]

4. Microsoft 365 Fails, and the AI Assistant Goes Down With the Email

Most shipping offices run on Microsoft: Outlook for email, Teams for calls, SharePoint for files. Many have now added Copilot, Microsoft's AI assistant, on top. This week all of it stopped together.

What happened. Around midday on 12 November in New York, which was the early hours of Friday in Manila, a failure in Exchange Online, Microsoft's email service, spread across 20 listed services. People could not open mailboxes in the browser, the desktop program or on phones, and some email arrived late. Teams calendars, Copilot, file sharing with people outside the company, device management and the email security filters that check links and attachments were all hit. Microsoft's cloud computing service Azure was not affected. On 13 November Microsoft reported progress, but also warned that servers it had restored could stop again once traffic came back. Users were still reporting problems.

Was shipping hit? Digital Ship found no shipping company that had reported damage when it published. But it makes the right point: many ship managers' shore offices depend on exactly these services, and many have put their AI assistant in the same Microsoft account as their email.

Why that matters. Two weeks ago we wrote about a bank that runs AI on its own computers. This is the other side. If your email, your calendar, your files and your AI all sit with one supplier, they fail on the same day. And a ship that needs an answer from the office at that moment does not care why.

The security detail. The filters that check links and attachments were also down. During an outage, people move to private email and messaging apps to keep working. That is when a false payment instruction gets through.

What to do this week. Check that every ship has a second way to reach the office that does not use Microsoft: a phone number, a satellite phone, a second email service. Keep a printed list of emergency contacts on board and in the office. Write down which of your AI tools stop when Microsoft stops. And agree one rule for outage days: any change to bank details or cargo release is confirmed by phone to a known number. [Digital Ship · AiatSea, 8 November]

5. Korea Approves Self-Navigation on a Passenger Ferry for the First Time

Last week we wrote about two Chinese canal ships that will try to steer themselves with containers on board. This week Korea took the next step: an approval for a ferry that carries ordinary passengers.

What was approved. The Korean maritime safety authority, KOMSA, approved the safety assessment for Samsung Heavy Industries' self-navigation system, called SAS, on the passenger ferry Tamnara. The ferry serves Nami Island, a tourist island on a river with no road to the mainland. More than 90% of its visitors arrive by boat. Smart Maritime Network reports this is the first approval of its kind for a passenger ferry in South Korea.

What the system does. It uses radar, GPS and AIS, the system ships use to report their position to each other, to find other boats and obstacles. It judges the risk of collision and suggests a route around it. It can also control the engines and the steering to take the ferry to a set point.

The path to this approval. In December 2025 the same system passed a safety check on the island's car ferry, Nami 51. So Korea started with vehicles and moved to passengers eleven months later. The system will also go on a new ferry, Sequoia, now being built. The island's operator wants to use the data from both ferries as a model for other tourist ferries on Korean rivers and lakes.

What is missing. The most important things. How many crew stay on board. How much of the trip the system steers and how much a person does. Whether the approval has conditions or an end date. And any results from eleven months on the car ferry.

Why it matters. Self-steering is arriving first on short, repeated routes on calm water, exactly as we wrote last week. But this is the first time we see an authority put its name on it for paying passengers. Approvals like this become the pattern other countries copy.

What to watch. Whether KOMSA or the operator publishes how often a person had to take over. That one number says more than the approval itself. [Smart Maritime Network · AiatSea, 8 November]

6. Masters Said the Link Was Too Slow. Weathernews Added an Offline Mode.

We keep writing that AI on a ship is only as good as the link. This week a large weather company showed what happens when you ask masters.

What masters reported. Weathernews, the Japanese weather company, tested its navigation support tool for masters, SeaNavigator for Master, between February and April 2026. The masters reported weak satellite coverage in parts of the Southern Hemisphere. They reported delays of more than 1,000 milliseconds, which means every click waits over a second for an answer. Routes failed to load. And in some coastal waters satellite use was restricted.

What Weathernews did. It launched an offline mode. A helper program downloads the weather for a chosen sea area, with a forecast of about 7 to 14 days, and keeps it on the ship's computer. Downloads can be planned up to four times a day, for the hours when the link is good. When the link drops, the forecast and the planned route stay on screen. The program can be installed from a shore connection or from a USB stick. It is available now.

The limit, and it is the point of this story. Two things still need a live link: testing a new route, and the tool's AI agent. So the map and the forecast work offline. The AI does not.

Why this is worth reading. It is a small product update. But it is one of the few times a supplier says in public what crews told it and what it changed. And it shows where the industry really is. The data can now live on board. The AI still lives ashore. Last week a bank showed that AI can run on a company's own computers. Nobody has yet shown that for a ship.

What to ask any supplier of AI for ships. What works when the link is down for a day? What works when the link is slow? Which parts need the AI ashore, and what does the master see when it cannot answer? Ask them to show you with the cable pulled out. [Digital Ship · Smart Maritime Network · AiatSea, 8 November]

7. Shanghai Builds a Place to Test AI Before It Goes to Work in the Port

Most AI in ports is tried on live work and judged later, if at all. Last week's survey showed that only 11% of port experts measure what AI gives back. Shanghai is trying a different order: test first.

What is being built. Shanghai International Port Group, the company that runs the port of Shanghai, has started building an AI test base for shipping and logistics, with university and industry partners. It will cover three areas: cargo handling in the port, transport between the port and inland, and transport that combines ship, rail and road. The base will have its own computing power, data storage and security, sets of real industry data, AI models, and a platform to manage AI agents. The aim is to move AI from development into daily work.

The number. 54 AI agents built for business tasks are already registered for testing and use.

The bigger plan. The news came with a new action plan from Shanghai's city transport authority on using AI in transport. It names three priorities: large AI models and agents for transport, good industry data, and computing power. A regional centre for an alliance that works on large AI models for transport was also set up. It covers Shanghai and three neighbouring provinces.

What is missing. No budget. No names of partners. No finish date. We are not told what the 54 agents do, who built them, or what an agent must show before it is allowed onto live work.

Why it matters. The idea is right, and it is the same idea as story 2. Give AI a safe place with real data, test it against something, and only then let it near ships and cargo. The world's busiest container port doing this in a formal way will be noticed by other ports.

What a ship manager can copy. You do not need a test base. You need one old voyage, one closed month of invoices or one finished dry dock. Give the AI tool that real, past data and compare its answer with what really happened. If a supplier will not do that test with your data, that tells you something too. [Smart Maritime Network · AiatSea, 8 November]

8. AiatSea: Who Decides When a Machine May Come In?

Read this week's stories again as one question. A machine stood at a door, and somebody had to decide.

Four doors, four answers. Petrobras opened the door a little: come on board, stand on deck, let us look. United Airlines closed it: no outside agents until there is a safe, agreed way in. Korea's safety authority opened it with a stamp: tested on a car ferry first, then approved for passengers. And GXO opened it with conditions: every job must pass agreed targets before the next step.

None of these is wrong. They fit different risks. A robot standing on a deck in port hurts nobody. An agent holding your password can spend your money. A ferry carries people. What is wrong is the fifth answer, the one most companies give without noticing: nobody decides. The tool simply arrives. A crew member installs an app. An office worker gives an assistant a password. A supplier switches on an AI feature in an update.

The Microsoft outage shows the cost of not deciding. Many offices never chose to put their AI assistant on the same account as their email. It came with the licence. This week both stopped together.

Five questions to answer before a robot or an agent reaches you. One: who in the company can say yes? Give it a name, not a department. Two: what exactly may it do, and what may it never do without a person? Write both lists. Three: what must it pass first? One task, one number, as in story 2. Four: what happens when it stops working, or when the link is down, as in stories 4 and 6? Five: whose door is it really? If the machine will log in to someone else's portal, as in story 3, the answer is theirs to give, not yours.

Our view. We like what Petrobras did. It is better to look early than to be surprised later. But a look is only a start. The companies that will use robots and agents well are not the first to let them in. They are the ones that know who decided, what was allowed and how it was tested. That takes one page of paper, and you can write it before the machine arrives. [Solstad · Skift · AiatSea, 8 November]

📊 Why It Matters — Strategic Impact Table

🔭 On Our Radar

• 🗳️ Who Is Actually Paying for Shipping's AI Accounts? — carried for a fourteenth week: Anthropic started selling its shares to investors on 10 November, with 825 million shares at an expected $70 to $78 each and a first trading day reported for the days before 26 November, we monitor whether AI prices or limits change for business customers after the listing, and our reader survey is still open (we use Claude, made by Anthropic, to help produce this newsletter).

• 🤖 Who Made the Robot on Normand Sigma, and What Comes Next? — Solstad's post names no robot maker and shows no task, we monitor whether Petrobras or Solstad says what the robot did on board, whether a trial at sea follows, and whether another offshore owner reports the same kind of visit.

• 📦 Which Targets Must the Warehouse Robots Pass? — GXO and Apptronik agreed targets but did not publish them, we monitor whether any number is made public when the trials start in 2027, and whether a port or terminal operator signs a similar deal.

• ✈️ Does a Port or a Carrier Follow United? — United banned outside AI agents from its bookings, we monitor whether any container line, terminal or port system changes its terms of use to block or to allow AI agents.

• 📧 What Caused the Microsoft Failure? — Microsoft had not fully restored service when we wrote this, we monitor its report on the cause, and whether any ship manager says how its ships and offices were affected.

• ⛴️ How Often Does a Person Take Over on the Korean Ferry? — the approval for Tamnara gives no crew numbers and no results, we monitor whether the safety authority or the operator publishes figures from the car ferry that has used the system since December 2025.

• 🚁 What Does Toei Kisen's Drone Trial Show? — carried for a third week: the hull inspection trial was due to start this month and no result is public yet, we monitor how long the inspection takes compared with the old way and whether class accepts the findings.

• 🇸🇬 What Does Singapore's New Law Say About AI? — carried from last week: the second reading in Parliament is still listed for November and we have seen no report of the debate, we monitor what it adds on AI, autonomous vessels and drones.

📅 Critical Maritime AI Research Areas for Managers

1. 🤖 Humanoid Robots on a Working Deck: a humanoid has visited an anchor handler. Research should test what such robots can really do on a wet, moving deck, starting with looking, measuring and carrying, and record grip, balance and battery life in different sea states.
2. 📦 Pass and Fail Tests for Robots and AI: warehouse robots must now pass agreed targets. Research should define simple output and safety tests for common ship and port tasks, so that owners can write the test before a trial.
3. ✈️ AI Agents and Portal Rules: an airline has banned outside agents. Research should review the terms of use of the main shipping, port and class portals, and set out safe ways for an agent to connect without using a person's password.
4. 📧 One Supplier for Email and AI: a single failure stopped email and an AI assistant together. Research should map which office and ship tasks stop during such a failure, and test simple backup channels between ship and shore.
5. ⛴️ Approval of Self-Navigation for Passenger Ships: Korea has approved a system on a ferry. Research should compare how authorities assess such systems, what tests they ask for, and what results operators must report afterwards.
6. 🌦️ AI That Works With No Link: weather data can now be stored on board while the AI agent cannot. Research should test which AI tasks can run on a ship's own computer, on what hardware, and how they behave on a slow link.
7. 🧭 Who Approves a Machine on Board: machines arrive through crew, offices and supplier updates. Research should study how shipping companies decide today, and propose a one-page approval that names the person, the limits and the test.

📈 Top Investment Opportunities

1. 🤖 Robots for Dangerous Deck Work — an oil company is looking at humanoid robots for anchor handling — the opportunity is in robots and tools built for wet, moving decks, starting with inspection and carrying, and in the safety systems that keep people out of the danger zone — Solstad · Petrobras
2. 📦 Robots Paid for as a Service, With Tests Attached — a warehouse company will pay for humanoids as a service and judge each job against agreed targets — the opportunity is in robot makers and service firms that accept measured targets, and in tools that record whether those targets are met — GXO · Apptronik
3. ✈️ Safe Connections Between AI Agents and Business Portals — an airline banned agents until a secure, agreed connection exists — the opportunity is in software that lets an agent prove who it works for and what it is allowed to do, without holding a person's password — United Airlines
4. ⛴️ Self-Navigation for Short Passenger Routes — Korea approved a system on a river ferry — the opportunity is in navigation systems for ferries on short, repeated routes, and in the testing and reporting services authorities will ask for — Samsung Heavy Industries · Nami Island
5. 🌦️ Ship Software That Keeps Working With No Link — a weather company added an offline mode after masters reported slow links — the opportunity is in tools that store data on board and, next, in AI that runs on the ship's own computer — Weathernews

📅 Top Monthly Picks

1. 🐕 Robot Dogs Start Real Inspection Rounds — ANYbotics began delivering a four-legged inspection robot with Equinor among the first buyers, and we asked whether one would ever board a ship; this week a humanoid robot stood on the deck of an anchor handler working for Petrobras — Splash247
2. 🛳️ What an AI Agent Actually Did on a 16-Day Voyage — ZeroNorth's agent spotted a drift and low fuel, and a person approved every change; this week United Airlines showed the other side, an owner of a system deciding that agents may not act at all — Maritime Innovations
3. 🏭 Ports Use AI Widely, but Only 11% Measure What It Gives Back — a survey of 103 port experts found most cannot show the gain; this week a warehouse company made agreed targets the condition for its robots, and Shanghai began building a place to test AI first — BCG
4. 💸 A Bank Cuts Its AI Bill by About 90% by Running AI on Its Own Computers — Revolut moved 2,000 engineers to an AI model it runs itself; this week a Microsoft failure took email and an AI assistant down together, which is the risk of renting everything from one supplier — Cybernews
5. 🚢 Two Canal Container Ships Will Try to Sail Themselves — a Chinese owner published its targets before the trial; this week Korea approved a self-navigation system on a passenger ferry, without saying how much a person supervises — Smart Maritime Network

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