Shipping needs to stop chasing the perfect AI solver

Shipping needs to stop chasing the perfect AI solver
July 16, 2026 https://splash247.com/shipping-needs-to-stop-chasing-the-perfect-ai-solver/

The next evolution of digital shipping is not a universal optimisation engine capable of planning every voyage, writes Natalia Liashenko, the CEO of Marine Solver. It is an architecture that combines specialised mathematical models with transparent, human-led decision-making, she argues in this exclusive for Splash.

As digital technologies continue to reshape shipping, the industry is following a natural evolutionary path. First, it learned to collect data. Then, to analyse it. Today, the focus is shifting toward exploring the decision space – the commercial decision environment where the future deployment of a fleet is determined. This naturally requires a new kind of instrument: a mathematical solver capable of exploring that space.

This is the next logical stage in the evolution of fleet planning intelligence. At this point, a seemingly obvious assumption emerges. If almost every commercial planning problem ultimately produces the same output – a vessel itinerary or a sequence of voyages – perhaps a single universal solver could automatically generate optimal solutions for every operational scenario.

At first glance, the idea appears perfectly logical. In reality, however, the concept of a universal, fully automatic fleet planning solver is a myth. Such a system cannot and should not exist.

Identical inputs and outputs do not mean identical mathematics

Consider a single baseline dataset: open vessel positions, available cargoes, laycans, ports, and distances. On the data layer, the inputs appear completely identical, but the moment the business objective shifts, the entire mathematical formulation changes with it.

A spot voyage requires maximising immediate returns while calculating the value of future positioning. A COA framework demands multi-voyage scheduling and optimising windows between rigid contractual obligations. A trader’s fleet requires managing an entire corporate cargo portfolio alongside open-market arbitrage. Meanwhile, part cargo planning breaks out of routing entirely, transforming into a complex combinatorial puzzle of stowage logic, volume-to-weight ratios, and physical cargo compatibility.

Of course, each of these problems ultimately produces the same practical output: a matching of vessels with one or more cargoes and a resulting voyage itinerary. However, reducing commercial fleet planning to mere routing subtly replaces a highly complex operational reality with a simple problem of geometry. Identical data streams do not imply identical mathematics.

One of the fundamental principles of mathematical modelling is often overlooked. A model does not describe data. It describes operational reality. That operational reality determines the mathematical structure of the optimisation model.

Different planning problems may even share identical objective functions – minimising costs, maximising profit, reducing non-productive time. But identical objectives do not imply identical mathematics, because they optimize fundamentally different processes. Real intelligence does not begin with computation. It begins with choosing – and building – the right model.

Fleet planning intelligence cannot be fully automated

Even with the correct mathematical model, different professionals make different – but equally rational – decisions. Given the same data, two operators may deliberately choose different strategies.

One may allow the model to pursue more aggressive commercial opportunities, operating close to laycan boundaries while expecting that minor deviations can be negotiated if necessary. Another may deliberately maintain larger time buffers to protect the entire downstream voyage chain against unexpected disruptions.

One may extend vessel employment as far into the future as possible, accepting the uncertainty of future market conditions. Another may intentionally release the vessel earlier, expecting stronger opportunities to emerge.

One may sacrifice short-term economics to preserve a more sustainable environmental trajectory for the fleet. Another may reasonably make the opposite choice.

All of these decisions can be equally rational. Because they reflect differences not in mathematics, but in management strategy. This is why Fleet Planning Intelligence systems must provide more than mathematical models. They must allow people to manage those models. Not by changing the mathematics itself, but by enabling the same mathematical framework to support different decision-making strategies that reflect each company’s experience, philosophy, and commercial priorities.

People should control the mathematical model – not the other way around.

Conclusion

Digitalisation has already taught shipping how to collect data. Artificial intelligence is gradually learning how to interpret it. The next step is considerably more challenging. Not to build yet another solver, but to abandon the very idea that one solver can fit every problem.

A single universal, fully automatic solver for fleet planning is neither possible nor desirable. What can be built is the right architecture for fleet planning intelligence – one in which every operational planning problem is represented by its own mathematical model, while every decision remains transparent, explainable, controllable, and ultimately guided by human judgement.