Maritime AI Digest — September 2026

Weekly roundup: CMA CGM starts designing an AI-assisted containership with SDARI and Bureau Veritas and puts the real question in writing, which is how much control stays with the crew; ShipIn's chief executive answers a serving master mariner in print and publishes fleet numbers for the first time, then converts the company to a Public Benefit Corporation with a legal duty to report; Marcura launches a procure-to-pay platform where AI reads the invoice and matches it against the order; Fugro builds an AI marine survey hub in Singapore with government money and claims some analysis runs up to ten times faster; Kaleris buys Newport Systems to automate container repair approvals across 25 data links built over twenty years; GTT Marine turns the voyage data recorder from accident evidence into a weekly bridge monitor; Cydome ships a tool for tracking your suppliers' cyber certificates under IACS UR E27; and we count what shipping actually launched at SMM Hamburg and find plenty of products, almost no published results, and one question nobody wants to answer about who is paying for the AI accounts already in use

Maritime AI Digest — 6 September 2026

Fifty thousand people went to Hamburg this week. Two thousand three hundred companies took a stand. Almost every one of them launched something. We read the announcements, and we counted. Plenty of products. Almost no results. One serious piece of thinking, from CMA CGM, which is asking how much of the ship the machine should be allowed to run and writing the answer into a design brief instead of a brochure. Underneath all of it sits a question we have now carried for four weeks and nobody will touch: the AI accounts already in daily use across this industry, in offices and on ships, are largely being paid for by the people using them. Not by the companies. If that is true, and we think it is, then every adoption figure published so far is measuring the wrong thing.

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

Subscribe & Never Miss an Issue

Take Our 2-Minute AI Adoption Survey

🔗 Quick Links

Also worth a look this week

🚀 Big Moves This Week

  1. CMA CGM Starts Drawing a Ship That Helps the Crew Instead of Replacing It

Most autonomy projects start with a question about the machine. This one starts with a question about the people.

CMA CGM has begun a joint study with SDARI, the Shanghai Merchant Ship Design and Research Institute, and class society Bureau Veritas, to design an AI-assisted containership. The words in the brief are "assistance rather than crew removal". The study will decide which navigation, energy management and operational jobs should get more automation, and, in their own phrasing, how much control should remain with the crew.

Read that second half again. It is written into the scope of work. Not into a press release paragraph at the end.

That is unusual and it matters. Almost every autonomy programme we have covered treats the human as a leftover: the machine does what it can, and whatever is left is the crew's problem. Here the split is a design decision that has to be argued for, costed and defended to class. The deliverables include a concept of operations, a basis of design, technical specifications for approval in principle, capital and running cost estimates, and a development roadmap. The work is deliberately technology-neutral, so no vendor is picked yet, and it is aligned with the IMO's MASS Code as that code takes shape.

If you have kept a watch, you already know why the split is the hard part. A system that does 90% of the work and hands you the other 10% at the worst moment is not help. It is a trap. The officer has been sitting there for three hours with nothing to do, and now he has eight seconds and no picture in his head. Aviation learned this the expensive way over thirty years. Shipping is about to learn it too, and the companies that write the boundary down early will learn it cheaper.

For a manager the useful part is the shopping list. When a builder or a supplier next offers you an automated function, ask for the same four things CMA CGM is asking for: what does it do without asking, what does it do only with confirmation, what does it hand back and how much warning does the officer get, and how is the handover recorded afterwards. If the answer is a brochure, it is not an answer.

What we do not have: no vessel, no yard slot, no budget, no timeline, no target date. This is a study, and studies die quietly all the time. Bureau Veritas being in the room from day one is the strongest signal in the announcement, because class involvement early usually means somebody intends to build. [Splash247]

  1. The Camera Argument Gets an Answer, and the Answer Has Numbers in It

Last week we led with a serving Maersk master mariner, Alex Byelyavtsev, arguing that an AI camera detects behaviour but cannot measure safety culture. On 2 September Osher Perry, founder and chief executive of ShipIn, published a reply on the same site and named the article he was answering.

He concedes the camera point. "A camera is only a sensor," he writes. His argument is that the value appears when visual data is combined with operational data, vessel status and history, so you see a pattern rather than an event.

Then he does the thing nobody in this category had done. He publishes figures. Across roughly 600 vessels, 29% of the ships accounted for 70% of the accidents. Ships scoring below 75 on ShipIn's safety score had 5.7 times the accident rate of ships scoring above 85. And deployments show safety deviations falling from around 230 to 8 per vessel per month, which he puts at a 96% reduction.

In the same week the company converted to a Public Benefit Corporation. That means the safety mission is written into the corporate charter, and the company now has a legal duty to publish an annual report on how it performed against it.

Now the honest part, because we asked for this data and we do not get to only like it.

The 96% figure is deviations detected. That number falls when crews adapt to the camera. It is exactly the measurement Byelyavtsev warned about. The 5.7 times figure uses ShipIn's own score to predict accidents, which is the company grading its own grading system. None of it is audited and no method has been published. And a Public Benefit Corporation is a promise about reporting, not a result.

So the fair summary is this. A vendor was challenged in public and answered in public with real numbers, which is more than the rest of this industry has managed. The numbers are not yet the kind you can rely on. Ask for the next annual report. That is the one that counts. [Splash247 · Smart Maritime Network]

  1. Marcura Points AI at the Invoice Pile, Which Is Where the Money Actually Leaks

Nobody will make a conference keynote out of this one. That is roughly the point.

Marcura has launched Procure-to-Pay, which joins up ordering, supplier setup, invoice handling, compliance and payment in one chain. The AI reads the invoice, pulls the numbers out, and matches them four ways against the purchase order, the delivery note and the receipt. It flags anything that looks wrong. Supplier bank details are verified. Sanctions screening runs daily, not once at onboarding. Payments go out in about 130 currencies, and the company says 99% land first time. Every step is timestamped from order to settlement.

The plumbing behind it is Marcura's payments arm MarTrust, regulated by the UK financial regulator, plus the ShipServ supplier network Marcura bought in 2023.

Anyone who has run a purchasing department knows where the money goes. Not to fraud, usually. To a spare that was ordered twice, an invoice that quietly does not match the order, a currency conversion nobody checked, a payment that bounced because the bank details were retyped by hand. It is a hundred small leaks and no single one is worth chasing.

Two things to watch. Daily sanctions screening is the sleeper feature, because a supplier who was clean when you set them up may not be clean when you pay them. And "anomaly detection" means the machine flags things a human then has to judge, which only works if somebody is actually given the time to judge them. A flag nobody reads is worse than no flag, because now it is documented that you were told.

No customers named, no vessel numbers, no savings figure. The 99% payment success rate is Marcura's own. [Smart Maritime Network]

  1. Fugro Builds an AI Survey Hub in Singapore, With Government Money Behind It

Fugro is expanding its Singapore technology hub into an AI-driven marine survey platform, backed by grant funding from the Singapore Economic Development Board.

What it does is pull together things that normally sit in separate boxes: geophysical survey, geotechnical boreholes, laboratory results and imagery, all into one model of the seabed. Machine learning does the joining up and much of the first-pass analysis, which is work currently done by hand and takes weeks. Fugro says its wider GeoAI framework can deliver some hazard, habitat and infrastructure analysis up to ten times faster than the conventional route. The company is hiring locally in AI, data science and digital engineering, and plans to scale the platform across Asia Pacific.

"As offshore projects become increasingly complex, the ability to turn data into actionable insights is more important than ever," said Safri Drahman, regional line director for marine site characterisation.

Why this reaches beyond survey work: seabed data is the input to cable routes, anchor design, jack-up positioning and wind farm layout. If the ground model arrives weeks earlier, every decision downstream of it moves earlier too. That is a real schedule saving, and for anyone in offshore it is worth more than another dashboard.

No investment figure, no grant amount, no timeline, no customers. "Up to ten times faster" is the company's own number for some tasks, not all of them, and it is speed rather than accuracy. Faster analysis of the wrong ground model is not progress. [Splash247]

  1. Kaleris Buys Twenty Years of Plumbing

Kaleris has acquired Newport Systems. Price not disclosed.

Newport writes maintenance and repair software for terminals and container depots. Pre-trip inspections, reefer monitoring, genset management, temperature tracking. The plan is to automate the paperwork chain that surrounds a container repair: estimate, approval, exception, invoice. Today that is done by people typing. Kaleris also says the joined-up dataset will support AI-based predictive maintenance and automated approvals later.

Note the word later. There is no AI in this today. What there is, and what Kaleris actually paid for, is 25 live data connections to shipping lines, leasing companies and equipment owners, built up over more than twenty years.

That is the honest lesson of the deal and it applies well beyond containers. The hard part of maritime AI is almost never the model. It is that the data lives in forty places, in forty formats, owned by forty parties who have no reason to help you. Somebody spent two decades wiring that up. You cannot train your way past it.

Kaleris says it serves more than 680 companies across 105 countries. [Splash247]

  1. The Black Box Stops Waiting for an Accident

Every ship carries a voyage data recorder. The VDR. It records the radar picture, the chart display, position, heading, speed, rudder, engine orders, alarms and the microphones on the bridge. Voices included.

Normally nobody opens it. It sits there for years. You pull the data after a collision, a grounding or a death.

GTT Marine, which owns Danelec, has now launched BOQA, short for Bridge Operations Quality Assurance, part of its Safety Insights product. It reads the VDR continuously and looks for passing another vessel too close, excessive rate of turn, excessive rudder angle, odd heading changes, sustained heel, too much speed in shallow water and severe wind. It weighs each event against context: closest point of approach, vessel length, speed, water depth. It filters over a period so one odd moment does not become an alarm.

"By leveraging that same foundation to provide safety insight while operations are still going right, BOQA ensures that every voyage becomes an opportunity to improve the next one," said Christian Kock, executive vice president for safety at GTT Marine.

Be clear about one thing: as described, this is rules, not AI. It checks recorded values against thresholds. We are running it anyway, because the interesting part is not the technology. It is what changes when the box changes job.

The black box was accepted by crews on a specific understanding. It is there for accidents. Nobody listens unless something terrible happens. Turn it into a weekly report and you have changed the deal without asking anyone. That is the camera argument again, only this time the sensor is already fitted, already recording, and already listening to what the master says to the pilot.

And before any of this is worth building on, there is a question we would put to every superintendent reading this. Does your VDR actually work? Not on the annual test certificate. In practice. Are all channels mapped after the last radar change. Is the bridge microphone still where it should be. Did the last download complete. Anyone who has tried to pull VDR data in anger knows the answer is often no. Clever analysis on top of a half-recording ship is worse than nothing, because it produces a confident report about a vessel the system cannot really see.

No customers named, no figures, no false-alarm rate, and no statement of who inside the company gets to read the output. [Smart Maritime Network]

  1. Somebody Finally Built the Spreadsheet Replacement for E27

No AI here either, and we are including it on purpose, because this is landing on real desks right now.

IACS UR E27 requires cyber resilience for computer-based systems on board, and it pushes the requirement down into the supply chain. Which means somebody in your office is tracking cyber certificates for every supplier who touches a ship system, on a spreadsheet, by hand, with reminders in Outlook.

Cydome has launched a platform to do it properly: one place for the vendor documents, dashboards by vessel, site, supplier and system, automatic flags for missing or unverified information, alerts before certificates expire, and a compliance score.

"E27 introduced important rules for mandatory supply-chain cyber resilience. However, this also brings operational complexity for managing compliance," said Alon Ayalon, Cydome's vice president of research and development and co-founder.

We have not covered cyber since 12 July, and this is the unglamorous end of it. It is also where most owners will actually meet the rule: not as a threat model, but as a filing job that grows every year.

No customer numbers, no pricing, no market figures. [Smart Maritime Network]

  1. AiatSea: Everybody Launched. Almost Nobody Measured. And Somebody Else Is Paying.

Fifty thousand people. Two thousand three hundred exhibitors from seventy countries. Four days in Hamburg. We went through the week's announcements looking for one thing: a number showing what a deployed system actually did.

Here is the count. A ship design study with no vessel. A survey platform with no investment figure. An acquisition with no price. A stowage tool with no saving. A performance integration with no customers. A compliance tool with no users. One safety product with real figures, published by the vendor about its own product, in reply to criticism.

That is not a scandal. It is a launch show, and launch shows are for launching. But it is now the fourth week running that we have written the same sentence in different words, and at some point a pattern stops being a coincidence and becomes the story.

Here is the part that should bother the industry more.

The published research says 81% of maritime companies are running AI pilots and only 11% have policies that let them scale. Separate work found only 14.4% of AI agents go live with full security and IT approval, and 31% of organisations cannot say whether they have had an AI breach at all.

Put that next to what you can see with your own eyes in any office or any messroom. The chief engineer drafting a defect report. The operator checking a charterparty clause. The superintendent turning three pages of notes into a job specification. It is happening now, every day, at scale, and almost none of it appears in a corporate adoption figure.

Because in most cases the person is paying for it themselves.

We have asked this question for four weeks and no manager, union or class society has published an answer. So say it plainly: if your staff are buying their own AI accounts to do your work, you are not a company with low AI adoption. You are a company with unmanaged AI adoption, and you are getting it for free while carrying all of the risk.

Company information is going into an account you do not control. There is no log, so after an incident you cannot show what the machine advised or who acted on it. When the person leaves, everything they built leaves with them. And in Europe the liability stays with you regardless of who paid the twenty euros.

The fix is not complicated. If it is company work, the company buys the account, and the company keeps the log.

And the measurement problem is not hard either. Pick five jobs that eat time. Port call paperwork, defect reports, PMS job descriptions, charterparty questions, supplier chasing. Time them now with a stopwatch and a spreadsheet. Give the tool to half the team, paid for and logged properly. Time them again in four weeks. Publish the difference. The reason nobody does this is not difficulty. It is that nobody wants the answer written down.

Our own caveat, and it is a real one. We do not have the numbers either. This is judgement built on what the surveys show, what the vendors do not publish, and what we see in our own working life in ship management. Our reader survey is open and collecting, and when we have enough responses we will publish what we find, including if it contradicts everything above. [AiatSea reader survey · AiatSea, 9 August]

📊 Why It Matters — Strategic Impact Table

Development ⇒ What It MeansWhat Ship Managers Should Do
CMA CGM Writes the Human Boundary Into a Design Brief ⇒ How Much Stays With the Crew Is Now a Specification ItemFor every automated function offered to you, get four answers in writing: what it does without asking, what needs confirmation, what it hands back and with how much warning, and how the handover is recorded. A brochure is not an answer. Put it in the specification, not the meeting notes.
A Vendor Answers Criticism With Its Own Figures ⇒ Published Numbers Beat No Numbers, But Detection Is Not SafetyTreat a falling deviation count as evidence that behaviour near the camera changed, nothing more. Ask what happens to the number when crews adapt. Diary the first Public Benefit Corporation annual report, because a legal duty to publish is the only part of this that a buyer can hold anyone to.
AI Reaches the Invoice Chain ⇒ The Return Is Real, Small and EverywhereCheck whether your sanctions screening runs daily or only at supplier onboarding, because a supplier clean at setup may not be clean at payment. Then make sure somebody is actually given time to judge the flagged invoices. A flag nobody reads is worse than no flag.
Kaleris Pays for Twenty Years of Data Connections ⇒ The Bottleneck Is Plumbing, Not ModelsWhen any supplier promises AI, ask first where the data comes from and who already owns those connections. If the answer involves building new interfaces to third parties, add a year and a large number to whatever they quoted you.
The VDR Becomes a Weekly Monitor ⇒ A Box Fitted for Accidents Is Being Repurposed for OversightBefore buying any VDR analysis, verify the recorder itself: channel mapping after equipment changes, microphone position, and whether the last download actually completed. Then decide and write down who may read the output, how often, and whether the master sees it before the office does.
E27 Supply-Chain Cyber Becomes an Administrative Load ⇒ The Rule Arrives as Filing, Not as a ThreatFind out today who in your organisation tracks supplier cyber certificates and what they use. If the answer is a spreadsheet and one person's memory, you have a single point of failure attached to a class requirement. Price the tool against the cost of one missed certificate at a survey.
Staff Are Buying Their Own AI Accounts ⇒ You Do Not Have Low Adoption, You Have Unmanaged AdoptionAsk your people, without blame, what they are using and who pays. If it is them, buy the accounts, log the usage, and keep the liability where it already sits, with you. Then run the four-week measurement: five time-consuming tasks, timed before and after, half the team, and publish the difference internally.
A Launch Show Produces Products Without Results ⇒ Evidence Is Now the Scarce Thing, Not TechnologyAsk every supplier for one deployed customer, one measured outcome and the method behind it. If none exists, that is fine, but price the contract as a trial with an exit, not as a solution. Write down now what the system must show in twelve months for you to call it a success.

🔭 On Our Radar

  • 🗳️ Who Is Actually Paying for Shipping's AI Accounts? — carried for a fourth week and now the central question of this newsletter: every published adoption figure counts organisations rather than people, our reader survey is open and collecting, we monitor whether any manager, union or class society publishes provisioning guidance instead of general policy language, and we will publish our own findings when we have enough responses, including if they prove us wrong.

  • 🚢 Where Does CMA CGM Draw the Line Between Machine and Crew? — the study asks in writing how much control stays with the crew, we monitor whether that boundary appears in the concept of operations and the approval in principle rather than in marketing, track whether Bureau Veritas publishes how it assesses such a split, and assess whether any other owner copies the question into its own newbuild specifications.

  • 📷 Does the First Public Benefit Corporation Report Contain a Real Outcome Number? — ShipIn now has a legal duty to report annually against its safety mission, we monitor whether that report carries fleet-level incident or claims data with a stated method, track whether any insurer or owner publishes independent before-and-after figures, and assess whether detection counts are still being offered where outcome data is asked for.

  • 🎙️ Who Is Allowed to Read the Black Box, and How Often? — VDR data is moving from accident evidence to routine monitoring, we monitor whether any owner, union or flag State publishes rules on access, frequency and retention for continuous VDR review, track whether masters see the output before the office does, and assess whether bridge audio is included or excluded in these products.

  • 🔧 How Many VDRs Would Actually Survive Being Used? — analysis products assume a healthy recorder, we monitor whether any service provider, class society or Port State Control body publishes failure and deficiency rates for voyage data recorders, track how often channel mapping breaks after equipment changes, and ask readers to tell us what they find on their own ships.

  • 🔒 Does E27 Compliance Become a Cost Anyone Publishes? — supply-chain cyber requirements are now administrative work spread across every supplier, we monitor whether any owner states what E27 compliance actually costs in hours or money, track whether the first non-conformities appear at survey, and assess whether small suppliers start being excluded because they cannot produce the paperwork.

  • 📦 Does Anyone Price the Plumbing Honestly? — Kaleris has just paid for twenty years of data connections rather than a model, we monitor whether maritime AI proposals begin stating integration cost and time separately from licence cost, track how many pilots stall at the data stage rather than the model stage, and assess whether owners start asking for the integration plan before the demonstration.

📅 Critical Maritime AI Research Areas for Managers

  1. 👤 Who Pays for AI at Sea and Ashore, and What That Costs the Owner: every adoption survey in this industry counts companies, and the actual usage is happening on personal accounts nobody has counted. Research should establish, by rank and by department, what proportion of maritime AI use is personally funded, what company data passes through those accounts, and what the legal and data-protection exposure is for the employer, so that owners can see the difference between low adoption and unmanaged adoption.
  2. 🚢 Where the Handover Boundary Should Sit in Assisted Navigation: designers are now being asked how much control stays with the crew, and there is no evidence base to answer with. Research should measure what happens to an officer's situational awareness and reaction time across a full watch of supervising an assisted system, and identify at what level of automation the handover becomes more dangerous than the manual task, giving class and flag something better than vendor assurance.
  3. 🔧 The Real Condition of the Installed VDR Fleet: safety analytics products assume the voyage data recorder is complete and correctly configured, and long experience says otherwise. Research should sample recorders across a real fleet and report channel completeness, mapping errors after equipment changes, audio quality, download success and time to retrieve, producing the baseline that any VDR analytics business case actually depends on.
  4. 🧾 Measured Return in Back-Office AI: the clearest current returns are in invoices, reporting and procurement, and they are described entirely in vendor input metrics such as time saved. Research should follow a defined set of back-office tasks through before-and-after measurement at several operators, publish the method openly so others can repeat it, and convert time saved into money with the assumptions stated.
  5. 🔒 The Administrative Cost of Supply-Chain Cyber Compliance: E27 pushes cyber requirements into every supplier relationship, and nobody has published what compliance costs in hours or in headcount. Research should quantify that load across fleet sizes, identify where small suppliers are being squeezed out for paperwork reasons rather than security ones, and test whether the requirement is improving actual resilience or only documentation.

📈 Top Investment Opportunities

  1. 🧾 AI in the Back Office, Where the Returns Are Small, Real and Repeatable — a procure-to-pay platform now reads invoices, matches them four ways against orders and delivery notes, screens suppliers daily for sanctions and pays in around 130 currencies with 99% success on first attempt, and this is the least glamorous and most provable corner of maritime AI — the opportunity is in invoice and document automation, continuous counterparty screening, anomaly review tooling that a human can actually work through, and the independent measurement services that would let a buyer verify any of it — Marcura Procure-to-Pay
  2. 🔗 Data Plumbing, Which Is What Buyers Are Really Paying For — Kaleris bought a company whose core asset is 25 live data connections to lines, lessors and equipment owners built over two decades, not a model, and that is the honest shape of this market — the opportunity is in integration layers, schema translation, EDI and API modernisation, and the specialist firms that connect maritime systems nobody else can reach, because every AI project in shipping stalls at this step before it stalls at any other — Kaleris + Newport Systems
  3. 🌊 AI on Offshore Ground Data, With State Money Behind It — a survey major is building an AI marine site characterisation hub in Singapore on government grant funding, claiming some hazard and infrastructure analysis up to ten times faster, in a market where seabed data gates cable routes, anchor design and wind farm layout — the opportunity is in automated interpretation of geophysical and geotechnical data, ground-model quality assurance, and the independent verification of speed claims that offshore developers will need before they compress a schedule on the strength of one — Fugro Singapore hub
  4. 🎙️ Getting Value Out of the Black Box, and the Governance That Must Come With It — every commercial ship already carries a voyage data recorder, a product now reads it continuously instead of after an accident, and the installed base is enormous while the recorders' real condition is unknown — the opportunity runs on both sides: analytics that turn existing recordings into safety insight, and the far less crowded business of recorder health verification, channel mapping assurance, retrieval services, and the access-governance tooling that decides who may listen to a bridge and when — GTT Marine BOQA
  5. 🔒 Supply-Chain Cyber Compliance as an Administrative Product — IACS UR E27 has pushed cyber requirements down into every supplier relationship, and the current method across most of the industry is a spreadsheet and a person's memory — the opportunity is in certificate tracking and expiry management, vendor assurance workflows shared between owners, class and suppliers, and the audit trails that will be demanded the first time a non-conformity is written against supply-chain cyber at a survey — Cydome vendor risk platform

📅 Top Monthly Picks

  1. 🚢 Japan Marine United Puts Supervised Autonomy on a Working Limestone Carrier — not a demonstrator, not a converted research vessel, but a ship that earns money carrying limestone, with a human supervising a system that navigates; it has aged well against this week's CMA CGM study, because it is the same question asked in steel instead of in a design brief, and it is still the clearest example of the boundary being drawn on a vessel that has a schedule to keep — JMU + NS United
  2. 🌊 ESVAGT Gives the Officer on Watch a Few Minutes of the Future — motion radar and software on a wind farm support vessel, predicting what the ship will do over the coming minutes so the officer's go/no-go call on a personnel transfer is made on a prediction rather than on feel; read next to the black box story in this issue, it is the cleanest illustration of the difference that matters — one system gives the officer more time to decide, the other gives the office more to review — ESVAGT NJORD + NextOcean
  3. 🪪 Dualog Gives Every Seafarer an Identity — individual accounts tied to vessel, rank and rotation instead of a shared bridge login, with message history that passes to the next person in the role; it looked like a messaging story at the time and it reads differently now, because you cannot give a seafarer an AI account, log what it was asked or prove who acted on the answer until the ship knows who is who, which makes this the quiet precondition for everything in our lead analysis this week — Dualog Workspace
  4. 🤝 MACN Turns 70,000 Corruption Reports Into Something the Ship Receives Before Arrival — two decades of anonymous reports about demands at the gangway, turned into a warning the master gets before the vessel berths rather than a report published after the fact; it remains the best example we have covered of data being pointed at the person who has to face the problem, and almost nobody in maritime software has copied the idea — MACN operational intelligence
  5. ✅ Evergreen Signs a Fuel-Saving Trial and Asks Class to Check the Number — a container line testing AI voyage optimisation with Samsung Heavy Industries and Weathernews, with ClassNK brought in to verify the result rather than to bless the product; in a week where almost every announcement arrived without a measured outcome, the habit of asking an independent party to check the figure before publishing it is the practice the rest of this industry should copy — Evergreen + ClassNK

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

Follow AI at Sea
Maritime AI Digest — 6 September 2026 | AI at Sea | AI at Sea