.png)
The case for curated OSINT and what you lose when you skip it.
Picture this scenario.
Your risk team is asked to screen a key supplier before signing a three-year contract. An analyst opens an AI-powered tool. Within five minutes, there is a clean, well-structured report on the screen: financial stability — solid; reputation — no red flags; regulatory compliance — clear.
The contract is signed.
Ninety days later, the supplier files for insolvency protection. The warning signals were there, employee complaints about delayed payroll, a series of small creditor disputes in regional court filings, a quiet management reshuffle. None of those signals made it into the AI tool's training data. Several were not in English.
This is not a hypothetical. It is a pattern.
What generative AI actually knows and when It stops knowing
Large language models are genuinely impressive technology. Trained on vast corpora of text, they excel at synthesis, summarisation, and generating answers that sound authoritative. In many contexts, that is exactly what you want.
In risk monitoring, it is precisely what makes them dangerous as a standalone solution.
The core problem is temporal. Every generative AI model has a training cutoff, a date after which the model has no knowledge of what has happened in the world. Ask it about a company's current financial health, and it will synthesise from data that may be six, twelve, or eighteen months out of date. It will not tell you what it does not know. It will construct a plausible-sounding answer from what it has.
In the language of AI research, this is called hallucination. In the language of enterprise risk management, it is called a liability.
Three blind spots of the AI-only approach
The gap between what generative AI appears to know and what it actually knows becomes especially acute across three dimensions:
Blind spot 1: Weak signals do not look like risk, until they do.
Real warning data does not arrive as a press release announcing a company's collapse. It arrives as a sequence of seemingly unrelated events: a first legal notice from a minor creditor, a CFO departure, negative employee reviews on a local jobs platform, a delivery lag that deviates from historical averages.
None of these individually triggers an alarm. Their correlation over time does.
Generative AI without access to verified, structured, real-time data streams cannot perform this correlation. It works with what it has, which is typically what was large enough to be reported at scale. By that point, the risk has already materialised.
Blind spot 2: Volume without classification is noise, not intelligence.
Today’s global news ecosystem generates millions of articles every day from hundreds of thousands of sources. For a risk team relying on an AI-only approach, this is not an opportunity, it is a flood.
A standard language model can read and summarize text, but it fails to transform this massive volume into actionable intelligence. At Semantic Visions, we bridge this gap by processing data across 12 languages and over 220,000 sources. What an LLM cannot do reliably is classify this scale consistently into a structured taxonomy of 680+ event types across jurisdictions. Without that precise classification layer, you don't have risk intelligence; you have a very fast news reader.
Blind spot 3: Compliance environments require auditability, not fluency.
A regulator, an auditor, or a risk committee does not need an output that sounds confident. They need to know: what was the source, at what time was the signal captured, what was the relevance score, and how was the event classified.
Generative AI without a structured data foundation cannot provide this chain of evidence. And without the chain, you do not have compliance, you have a convincing story.
A real-world illustration: when the signals were there
The collapse of Czech solar energy company Energetický Holding Malina in spring 2023 offers a useful case study in what structured early warning monitoring can and cannot catch.
The trajectory looked like this:
• July 2022 — The company signed a CZK 1 billion purchase agreement for lithium batteries, generating positive press coverage
• January 2023 — employee walkouts over unpaid wages begin to surface in regional media
• March 2023 — customer complaint volumes surge; communication breakdowns reported
• April 2023 — company struggles to fulfil orders, seeks investor; management reshuffle announced
• Late April 2023 — first legal enforcement actions, creditor complaints, reports of suspicious asset transfers
https://www.novinky.cz/clanek/domaci-zalobci-resi-uz-tri-oznameni-kvuli-solarni-firme-40429796
• May 2023 — insolvency filing
For anyone monitoring structured, real-time signals in Czech-language media, the warning was available more than 30 days before the insolvency filing. Each individual signal was ambiguous. The sequence was not.
A generative AI tool with no access to real-time, multilingual, structured data feeds would have looked at the July 2022 contract win and the early 2023 media profile and returned a broadly positive assessment. The signals that mattered were too granular, too local, and too early-stage to register any other way.
What curated OSINT adds where LLMs fall short
Open Source Intelligence (OSINT) is not simply 'searching the internet'. It is the systematic collection, filtering, classification, and correlation of information from verified sources in real time.
The operative word is curated. Not all data is created equal. A regional Czech outlet that first reported on a solar company's payroll difficulties carries more risk-relevant signal than the same company's annual report translated into English three months later.
Effective curated OSINT for risk monitoring has four components that generative AI alone cannot replicate:
• Real-time ingestion: signals captured and classified within hours, not weeks
• Structured taxonomy: consistent categorisation across hundreds of event types, enabling meaningful comparison over time
• Multilingual coverage: signals captured in the language in which they first appear, before translation or editorial selection erases context
• Correlation and scoring: the ability to identify when a combination of signals crosses a risk threshold, not just when a single event is severe
When language models are applied on top of this foundation, for summarisation, translation, pattern detection, and analyst-facing output, they become genuinely powerful. They are answering the right question with the right data. That is a fundamentally different proposition from asking an LLM to reason about a company it last 'knew' six months ago.
The architecture of actual risk intelligence
A well-designed risk monitoring stack looks like this:
At the input layer: a structured, continuous data feed from verified global sources, news outlets, regulatory databases, corporate registries, social signals, classified in real time against a predefined event taxonomy.
At the intelligence layer: language models applied not as the source of truth, but as tools for summarisation, cross-language synthesis, anomaly detection, and analyst-ready output generation.
At the delivery layer: interactive dashboards, audit-ready event logs with source attribution, and data feeds compatible with the client's existing infrastructure, cloud storage, data warehouses, SIEM systems.
The result is not a faster way to read news. It is a system that can say: this entity is exhibiting a combination of signals that historically preceded financial distress in the majority of comparable cases. Here is the timeline. Here are the sources. Here is the relevance score.
Three questions to ask about your current Setup
If your organisation is using AI tools for risk monitoring and an increasing number are three questions are worth asking:
• Are the data sources your tool draws on updated in real time, or is the model working from training data that is months old?
• Can you audit why a given entity was flagged as high-risk or do you receive a conclusion without a traceable chain of evidence?
• Does your coverage extend to the languages and regions where your suppliers, counterparties, and customers actually operate?
If the answer to any of these is not a clear yes, you have a blind spot. The question is not whether a risk event will eventually find it, but how much it will cost when it does.
The bottom line
Generative AI is neither a solution nor a threat to risk intelligence. It is a component of it, a powerful one, when properly positioned. The mistake is in treating it as the entire stack.
The organisations that will manage risk most effectively over the next decade are not the ones with the most sophisticated language models. They are the ones that understand what those models need to be useful: clean, structured, real-time, multilingual data and the analytical layer to make sense of it.
That is the gap curated OSINT fills. Not instead of AI. Before it.
Access Premium Insights
.png)
See Everything. Focus on What Matters.
svEye™ filters the noise to uncover meaningful patterns and insights. Gain clarity, stay informed, and drive smarter decisions with a comprehensive overview.
