- The real gains are unglamorous: fuel, maintenance, arrival accuracy and document processing.
- Better ETAs are the change that reaches cargo owners fastest — you can plan trucks and warehouse slots against a prediction that holds.
- Every one of these systems depends on data quality. Bad sensor data produces confident, wrong answers.
- Models propose; masters, operators and agents still decide. Accountability has not moved.
Shipping has always generated enormous amounts of data — positions, weather, engine telemetry, port line-ups, cargo manifests — and for most of its history wrote it down and forgot it. What has changed is not that the data exists but that it can now be used continuously, at scale, while the voyage is still happening.
Cutting through the marketing, here is where machine learning genuinely earns its place aboard, and where it does not.
1. Voyage optimisation: routing that reads the weather
Traditional weather routing avoided the worst of the ocean. Modern voyage-optimisation systems do something more specific: they weigh forecast wind, swell and current against this particular vessel's performance curve — how this hull, at this draft, with this fouling condition, actually behaves — and continuously re-solve for the best speed and track.
The gains are a few percent of fuel per voyage. That sounds modest until you multiply it by a fleet across a year, and it comes with two side benefits: fewer heavy-weather cargo claims, and arrival windows that terminals can plan around.
Arriving early to wait at anchor burns fuel to achieve nothing. Optimisation frequently recommends reducing speed to arrive exactly when a berth is free — the single largest saving available on many voyages, and it needs no new hardware at all.
2. Predictive maintenance: engines that flag themselves early
Vibration, temperature, pressure and consumption signatures drift before components fail. Models trained on those signatures across many sister vessels can flag a bearing, a turbocharger or a fuel pump moving toward failure while there is still time to plan.
The value is not the repair — it is where the repair happens. A part changed at a scheduled port call is an inconvenience. The same part failing mid-ocean is off-hire days, a possible tow, and a chain of missed berths behind it.
3. ETAs you can plan a truck against
This is the change that matters most to cargo owners. A published schedule is a plan made weeks ago. A machine-learned ETA blends the vessel's live behaviour with berth availability, terminal productivity and queue depth at the destination — and updates continuously.
The difference is practical: warehouse labour booked for the right day, trucks scheduled without idle waiting, customs filed in the right window, and a delivery date you can give a buyer without hedging.
4. Documents that check themselves
A single shipment can generate dozens of documents that must agree with one another: invoice, packing list, bill of lading, certificate of origin, manifest, declarations. Most customs holds trace to a mismatch between two of them.
Document-extraction models read these in seconds and flag inconsistencies — a weight that disagrees, an HS code that does not match the description, a consignee name spelled two ways — before a human submits the file. The clearance gain is not glamorous; it is simply fewer queries.
Where each of these actually stands
| Application | Maturity | Main constraint |
|---|---|---|
| Weather routing / speed optimisation | Mature, widely deployed | Accurate hull performance data |
| Predictive maintenance | Maturing on newer tonnage | Sensor coverage on older ships |
| Predicted ETAs | Mature and improving fast | Port-side data sharing |
| Document automation | Mature | Document quality and format variety |
| Autonomous navigation | Experimental | Regulation, liability, edge cases |
What it does not change
Three things are worth stating plainly, because the marketing tends to skip them.
Data quality governs everything. A model fed a mis-calibrated flow meter will produce a confident, precise, wrong answer — and confident wrong answers are more dangerous than obvious gaps.
Prediction is not control. Knowing a vessel will arrive late does not create a berth. The value of a good forecast is only realised if the plan downstream actually changes in response, which is an organisational problem, not a technical one.
Judgement stays human. A model proposes a route; the master decides. A model flags a bearing; the chief engineer decides. A model predicts an ETA; the agent decides whether to hold the gang. Liability has not moved, and neither has responsibility.
The winners this decade will not be the operators with the most models. They will be the ones whose people know which outputs to trust, and what to do about them.
The data problem nobody puts in the brochure
Every application above rests on data that shipping has historically collected loosely. Noon reports filled in by hand, sensors that drift out of calibration between dry-dockings, fuel figures rounded to a convenient number, port timestamps entered hours after the event.
Models trained on that produce outputs with false precision. A predicted fuel saving of 4.2% carries an implied accuracy the underlying data cannot support, and the danger is that the number gets believed because it has a decimal point.
The operators seeing genuine returns almost all did the same unglamorous thing first: they fixed measurement. Calibrated flow meters, automated noon reporting, consistent event timestamping, one authoritative record of each vessel's particulars. It is a year of dull work that makes everything afterwards possible, and it is routinely skipped in favour of buying a platform.
What this means for a cargo owner
If you ship goods rather than operate vessels, most of this is somebody else's capital expenditure. Three consequences still reach you directly.
Your ETAs should be getting better. If your forwarder still quotes the published schedule and nothing else, they are not using information that exists. Ask whether the arrival window they give you is the carrier's plan or a live prediction — the two differ, often by days.
You should hear about problems sooner. The value of predictive systems is warning time. A partner who knows on Monday that Friday's connection is at risk and tells you immediately is worth considerably more than one who confirms the miss afterwards.
Fewer of your shipments should be held at customs. Automated document checking catches the mismatches that cause queries. If your consignments are still being held for paperwork inconsistencies, the tooling is not being used on your file.
How to adopt this without wasting money
- Fix the data first Calibrated sensors, consistent noon reports, clean master data. Everything downstream inherits these errors.
- Pick one measurable problem Fuel per voyage, off-hire days, or ETA accuracy — one, with a baseline you can defend.
- Run it against the status quo Compare the model's recommendation with what the team would have done anyway. If they agree, you have learned the model is safe. If they differ, you have learned something more valuable.
- Change the workflow, not just the dashboard A better ETA that nobody plans against has delivered nothing.
- Keep the human in the loop Every recommendation should be explainable enough for the person accountable to accept or reject it.
Two more places it is quietly working
Beyond the four headline applications, two less-discussed uses are already delivering value.
Stowage planning. Deciding where each container sits on a vessel is a constrained optimisation problem: weight distribution and stability, discharge sequence across multiple ports, reefer plug positions, dangerous-goods segregation rules, and the need to avoid restowage. Planners have always done this well by hand; optimisation tools do it faster and explore more alternatives, which matters most on complex multi-port rotations where a good plan removes restowage moves entirely.
Empty container repositioning. Trade imbalance leaves empties piling up where they are not needed and absent where they are. Forecasting demand by depot several weeks out lets a line reposition on a scheduled service rather than in a panic, and the saving is substantial because repositioning empties is pure cost with no revenue attached.
The limits worth being blunt about
Three claims should be treated sceptically whenever you encounter them.
"Fully autonomous vessels are imminent." Crewless deep-sea shipping faces regulatory, liability and insurance obstacles that are further from resolution than the technology. Reduced-crew and remotely supported operations are plausible this decade; unmanned ocean crossings at scale are not.
"The system will predict disruptions." Models extrapolate from patterns they have seen. Genuinely novel events — a canal blocked, a sudden regulatory change, a conflict closing a corridor — are precisely the ones with no precedent to learn from, and they are the ones that hurt most.
"It removes the need for expertise." The opposite is true in practice. Interpreting a recommendation, recognising when it is being fed bad inputs, and knowing when to override it all require more domain knowledge than following a fixed procedure ever did.
What we do with it
At S J Logistics the interest is narrow and practical: knowing sooner, and telling customers sooner. Live tracking on every shipment, predicted arrival windows that inform the road and warehouse plan, and document checks that catch mismatches before filing rather than after a customs query.
The technology is the boring part. The point is that when something moves, someone tells you — with enough notice to do something about it. Tell us what you are shipping and we will show you what that visibility looks like on your lane.
