How AI agents are changing the way we analyze AIS data

For thirty years, studying AIS data has meant more or less the same thing: write a spatial query, geofence a region or a vessel, pull the tracks, and stare at a dashboard until a pattern emerges. The data got bigger — well over a billion AIS position signals a day, from thousands of terrestrial and satellite receivers — but the workflow did not. The analyst was still the query engine and the reasoning engine.

That is the part agents change.

From query-and-stare to plan-and-investigate

An AIS dashboard shows you pre-computed views and leaves the investigation to you. An agentic workflow performs the investigation. Given a question — “did this tanker do a ship-to-ship transfer in the last 30 days?” — an agent will:

1. Decompose the question into steps: find the vessel’s track, detect gaps, look for a second vessel loitering nearby during a gap, check both vessels’ registry and sanctions status.
2. Call tools to execute each step — a track query, a gap detector, an encounter detector, a registry lookup — rather than generating an answer from memory.
3. Fuse the results across feeds, resolve the vessel identity across MMSI and IMO, and weigh the evidence.
4. Answer in natural language, with the track, the gap window, the co-located vessel and the registry hit attached as citations.

The analyst has stopped being the query engine. They ask in plain language and review an evidence trail. The mechanical work that used to take an afternoon collapses to minutes.

Why AIS is unusually well suited to agents

Three properties of AIS make it a natural fit for tool-using agents:

• It is a means, not an end. Nobody wants “the AIS tracks”; they want a decision — is this vessel risky, where is the cargo going, will it arrive on time. That gap between raw feed and decision is exactly the reasoning layer an agent fills.
• The questions are compositional. Real maritime questions chain many small operations — gap → encounter → identity → sanctions. That is precisely the multi-step tool-use that agents are good at and that a single SQL query is bad at.
• The answer needs corroboration. A single feed is never enough; the credible answer fuses AIS with SAR, weather and registries. An agent orchestrating several APIs does the fusion that a human would otherwise do by hand across five tabs.

What the research shows

This is not speculative. Research systems already unify the pieces: AIS-LLM, for example, combines vessel-trajectory prediction, anomaly detection and collision-risk assessment behind a single natural-language interface, so an operator can interrogate a situation in words and get an explained result rather than a raw score. It is a template for the explainable maritime copilot.

Underneath the agent, the models are changing too. Time-series foundation models (such as TimesFM and Chronos) bring zero- and few-shot forecasting to ETA, congestion and demand across many ports at once, and geospatial foundation models (Prithvi-EO, TerraMind) can be fine-tuned for SAR and optical vessel detection. Agents are the orchestration layer that puts these models to work against a live question — see our data & methods for how they fit together.

The interoperability shift: tools, not exports

The older integration pattern was export and ingest — dump a vendor’s AIS into your warehouse and build on it. The emerging pattern is give the agent a tool. Standard interfaces (the Model Context Protocol and similar) let an agent call a data source directly, on demand, alongside your own systems. That is a quiet but important change: it favours vendor-neutral orchestration over single-platform lock-in, because the agent can reach whichever feed answers the question rather than only the one it lives inside. It is the same argument we make in why an independent analytics layer wins.

What stays hard

Agents move the work; they do not remove the judgement. Four things stay firmly human:

• Grounding over generation. An agent that reasons about tool results is trustworthy; one that generates vessel facts from a language model is dangerous. The whole design has to force tool-use and cite sources, or it will confidently invent a port call. Retrieval and tool-calls are the guardrail.
• Coverage normalisation. An AIS gap only means something once it is normalised against expected satellite coverage. An agent that treats every silence as evasion is a false-positive machine. The hard maritime knowledge still has to be encoded.
• Explainability and lineage. In sanctions, underwriting and enforcement, an unexplained score is unusable. Every agent output needs the evidence and the data lineage a human can defend to a regulator.
• Analyst-in-the-loop accountability. The agent does the legwork and shows its work; a person makes the call and owns it. That division of labour is the point, not a limitation.

The net effect

Agents do not replace the AIS analyst — they replace the analyst’s worst hours: the repetitive querying, the tab-juggling across feeds, the manual write-up. What is left is the part that was always the job — asking the right question and judging the evidence — now at a scale and speed that a lone analyst could never reach.

That is how we build for maritime domain awareness and sanctions screening: agentic pipelines that investigate, fuse and explain, with a human holding the decision.