Artificial intelligence applied to maritime operations, fleets, and shoreside businesses.
Maritime AI is the use of machine learning, large language models, and decision-support systems in shipping, superyachting, marinas, defence marine, and the wider blue economy. It spans voyage optimisation, predictive maintenance, vessel traffic analytics, document automation, crewing, charter operations, and shoreside admin. The common thread is that the AI is tuned to maritime data, regulation, and workflow rather than treated as a generic office tool.
Software that recommends routing, speed, and trim to reduce fuel burn and arrival risk.
Voyage optimisation systems combine weather, current, bathymetry, vessel performance, ECA boundaries, and commercial constraints to recommend a route and speed profile. Modern systems use machine learning on historical noon reports and high-frequency sensor data to predict fuel consumption with greater accuracy than naval-architecture curves alone. The output is a plan a master can accept, adjust, or override.
Using sensor data and models to predict component failure before it happens.
Predictive maintenance applies anomaly detection and time-series models to engine, generator, HVAC, and auxiliary data to flag emerging faults. For maritime operators it shifts the maintenance pattern from calendar-based or run-hours-based intervals to condition-based interventions, reducing both unplanned downtime and unnecessary overhauls.
Computer-assisted approach and mooring that combines sensors, thrusters, and control software.
Automated and assisted berthing systems use lidar, radar, GNSS, and computer vision to measure approach geometry and apply thrust automatically or guide the bridge team. Fully autonomous berthing remains rare in commercial service; the more common deployment is decision support that reduces contact damage and pilot workload in confined waters.
A large language model whose responses are grounded in maritime regulation, doctrine, and operational context.
A maritime LLM is not necessarily a new model trained from scratch — it is more often a general-purpose model (Claude, GPT, Gemini) configured with maritime-specific retrieval, prompts, and policy guardrails. The value comes from grounding answers in MARPOL, SOLAS, flag-state circulars, class rules, charter party language, and the operator's own SMS, rather than from the model's open-web training data.
A pattern where the model retrieves authoritative documents at query time and answers from them.
RAG keeps the model's reasoning ability while constraining its answers to a controlled corpus — SMS manuals, class rules, charter parties, technical drawings. For maritime operators this is the most common way to deploy LLMs safely: answers are sourced from documents the organisation already trusts, with citations the user can verify.
The policies, controls, and oversight that determine how AI is selected, deployed, and monitored.
AI governance covers procurement standards, data-handling rules, human-in-the-loop requirements, incident reporting, model evaluation, and disclosure to clients and regulators. ISO/IEC 42001 is the emerging international standard for an AI management system. For maritime businesses, governance also has to interoperate with ISM Code, flag-state requirements, and cyber rules such as IACS UR E26/E27.
The international standard for an AI Management System (AIMS).
ISO/IEC 42001 specifies how an organisation establishes, implements, maintains, and continually improves a management system for the responsible use of AI. It mirrors the ISO 27001 structure familiar to maritime cyber teams. Maritime operators are increasingly asked about AIMS conformance during due diligence, particularly by flag states, class, and large charterers.
A control pattern where a qualified person reviews or approves AI output before it takes effect.
For safety-of-life and commercially material decisions, human-in-the-loop is the default expectation. The AI proposes; a master, chief engineer, or shore manager disposes. Good governance is explicit about which decisions are human-in-the-loop, which are human-on-the-loop (monitoring only), and which are automated outright.
When a generative model produces a confident answer that is not supported by evidence.
Hallucinations are a structural property of generative models, not a bug. The mitigation is not 'a better model' but better grounding: RAG over trusted documents, citations the user must check, narrow task scoping, and refusal patterns when the model lacks evidence. For maritime use, the cost of an unchecked hallucination — a wrong port-state regulation, an incorrect class requirement — is high enough that grounding is mandatory.
An AI system that takes multi-step actions toward a goal, calling tools and other services.
Agents combine an LLM with tools (search, code, APIs, internal systems) and a loop that lets them plan and execute. Useful maritime examples include charter-enquiry triage, document collation for vetting inspections, and shoreside admin chains. The governance question for agents is bounded authority: which systems they can read, which they can write to, and what triggers human review.
Adjusting a base model's weights on a specific dataset to specialise its behaviour.
Fine-tuning is one way to make a general model more reliable on a narrow task — house style, terminology, formatting. It is not the right answer for keeping the model current with regulation (use RAG) and it locks the operator into a specific base model version. Most maritime deployments get further with prompt engineering and retrieval than with fine-tuning.
Numerical representations of text or images that let software compare meaning, not just words.
Embeddings turn a document into a vector. Two documents about the same topic — even in different languages or with different terminology — produce similar vectors. This is what makes RAG work: a query is embedded, the system retrieves the closest documents from the corpus, and the LLM answers from them.
AI that interprets images and video — object detection, classification, tracking.
Maritime computer-vision use cases include collision avoidance, dock approach, marine-mammal detection, hull and tank inspection from drone footage, and security monitoring. The performance limits to understand are camera-physics limits at night and in heavy weather, and the cost of false negatives in safety contexts.
Analysis of Automatic Identification System tracks to derive port, fleet, and market intelligence.
AIS analytics turn raw vessel position broadcasts into derived data: port congestion, voyage times, dark-vessel detection, charterer behaviour, sanctioned-vessel screening. This is the most mature 'AI in maritime' category and is sold widely by Windward, Spire, MarineTraffic, and others.
A live computational model of a vessel, system, or operation kept in sync with real-world sensor data.
A digital twin can simulate 'what-if' scenarios — load cases, performance under weather, intervention timing. For maritime operators the practical value is concentrated in newbuild design, performance management, and incident replay. Twins are infrastructure: they only pay back if there is a discipline of using them to make decisions.
An open protocol that lets AI assistants connect securely to tools and data sources.
MCP, introduced by Anthropic in 2024, is becoming the standard way to plug an AI assistant into an organisation's own systems — calendars, file stores, internal databases, custom tools. For maritime operators it is the path to bringing Claude or other assistants safely into vessel-management, charter-management, and ERP systems without bespoke integration each time.
AI features added to tools you already use, often without explicit opt-in.
Most maritime businesses encounter AI first through their existing tools — Microsoft 365, Google Workspace, Adobe, Zoom, Slack, vessel-management software. Each vendor has its own data-handling and model-training defaults. A practical governance step is to inventory which vendors have added AI features, and to set the model-training, retention, and regional-residency settings on each.
The craft of structuring instructions to a model so it produces reliable, useful output.
Prompt engineering is the closest thing to 'programming' a general-purpose model. For maritime applications, well-designed prompts encode role (e.g. 'You are a SOLAS compliance officer'), source constraints ('use only the attached SMS'), output format, and refusal conditions. Good prompts often outperform fine-tuning at a fraction of the cost.
How an LLM breaks text into the units it actually processes.
Models read text as tokens — roughly word-fragments. Token counts drive both cost and the size of context an operator can work with. For long maritime documents (SMS, charter parties, class rules), token economics shape which model is practical for which task.
How much text a model can consider in a single request.
Context windows have grown from a few thousand tokens to over a million in the latest models. For maritime work this changes what is possible: a single prompt can now hold an entire SMS, a year of noon reports, or a full vetting questionnaire — reducing the need for retrieval in narrowly-scoped tasks.
