This scenario is a composite created to illustrate the workflow. The slang and company names are invented.

1. The Incident

An investigative journalist receives a tip that a tanker with its Automatic Identification System (AIS) transponder switched off is transferring oil off the coast of Southeast Asia, possibly to evade international sanctions.

The challenge: Thousands of satellite images and hundreds of open Telegram posts are generated in the region every week, far too much unstructured data for one investigator to review manually.

2. The Investigation Timeline

Day 1, 09:00: AI Data Sweep
The investigator runs a computer vision model over Sentinel-1 radar imagery, which works through cloud and at night, to detect vessels. Detections are compared against AIS records to isolate ships broadcasting no signal. In parallel, a Large Language Model (LLM) scans multilingual public chat channels, translating and tagging mentions of fuel transfers or rendezvous coordinates.

Day 1, 11:30: The AI Flag
The model flags two vessels side by side in open water with no AIS signal. The LLM also surfaces a Telegram post boasting about an upcoming “cash-for-fuel” handover near the same coordinates.

Day 1, 13:15: Critical Human Interventions

• Mistake caught: The AI linked one vessel to Company X based on a broad hull-dimension match. The analyst checks historical port logs and the ship’s permanent IMO number in international registries, and finds that Company X sold the vessel two months earlier. Registered ownership has moved to an opaque shell company in Panama.
• Context applied: The AI scored the Telegram post as low relevance because of local slang (“the big whale is drinking”). The analyst, familiar with regional maritime jargon, recognizes “drinking” as a reference to ship-to-ship (STS) oil transfer.

Day 2, 16:00: Ground-Truth Verification
The analyst orders high-resolution commercial satellite imagery to inspect the vessels’ deck layouts and confirm the match. Image metadata and sun-shadow angles are used to confirm the capture time, and regional wave and weather data are checked to confirm sea conditions allowed a transfer. Panama registry filings identify the shell company’s registered agent and directors. Tracing the real owners behind it will take further work.

3. Key Takeaways

• AI finds the needle; humans prove it’s the right needle. Models can sift satellite imagery and chat logs in hours rather than weeks.
• Context requires human knowledge. AI struggles with local slang, sarcasm, and non-standard terminology.
• AI errors are an operational liability. Outdated or faulty entity matches, such as linking a ship to a former owner, can destroy an investigation’s credibility if uncorrected.
• Verification is non-negotiable. IMO checks, timestamps, weather data, and official registries all require human judgment and documented methods.

The AI processed weeks of data in hours. But without a human catching the outdated ownership link and recognizing the slang, the story would have been missed or published with critical errors.