Jan Balatka on Why Early Supply Chain Signals Live Outside English-Language Media

Most business leaders find out about a supply chain problem the way everyone else does: after it has already hit the news. By then, shipments are stuck, a supplier is scrambling, and the finance team is trying to calculate how much the delay will cost. The data that could have surfaced the warning days or weeks earlier was sitting in a local newspaper somewhere, in a language nobody on the risk team speaks, buried under thousands of unrelated stories. That gap, between what is knowable and what companies actually know, is the space Semantic Visions has been working in for more than a decade.

Jan Balatka is the CEO of Semantic Visions, a Prague-based intelligence company that analyzes global news at scale to flag early signals of supply chain, geopolitical, and third-party risk. The company started out serving government agencies on national security use cases and has since moved into the commercial world, where it now feeds signals into the risk, procurement, and trading workflows of some of the largest companies in the world. We spoke with Jan about where traditional risk assessment falls short, what LLMs can and cannot do with raw news feeds, and how a relatively small team has managed to hold its own against competitors many times its size.

What problem were you originally trying to solve, and how has it changed?

Our archive starts in 2015. The early years of Semantic Visions were spent for initial development of the platform. Originally we served government agencies, we still do, but later we moved also to commercial sectors. What was initially about national security use cases is now about third party / supply chain risk or credit risk and algorithmic trading use cases.

Why do financial ratings and periodic audits miss so many disruptions?

Financial ratings are a very solid tool, especially for large companies. But they are published with a delay and they speak about the company as a whole. We track disruptions that can strike a shipment or a product line and in real time. The very vibrant global business is difficult to describe with annual or quarterly scores. Changes happen too often, disruptions have quick cascading effects and geopolitical issues are torpedoing the whole situation from the back.

How does a small team hold up against much larger competitors?

We focus on our tradecraft which is to provide very precise insights with extremely broad source coverage. We often are integrated as a data provider into larger software solutions because we do not provide end-to-end business process management for let’s say supply chain risk. But where external data feeds this process we have the largest collection with the most accurate signals and we see them first.

What did you deliberately choose not to build?

We did not build a fully fledged application that would support the business processes where external intelligence is needed. We rather integrate our data into existing solutions or directly into the existing environment of our clients.

Why isn’t plugging an LLM into news feeds enough?

LLMs are groundbreaking technology, we use them for a lot of particular tasks. But they are very difficult or extremely expensive to be used in large scale tasks. Imagine you are a global manufacturer with 100.000 direct suppliers. Mapping these to tier2 and tier3 means looking at millions of companies globally. It would be highly impractical to input millions of companies to LLM and then try to digest the output of that.

How does structured data change what an LLM can reliably do?

It enables LLMs do their job much better. By design, public LLMs run top 20 or top 50 queries when they research a topic. That limits completeness at the very beginning because they will most probably consider the largest sources first. They will not get to local non-English sources where important facts might be mentioned.

Another issue is the context window. It is increasing, but still nowhere even close to the task at hand (screening millions of companies over an extended period of time). So by providing the initial pre-processing of the large scale data and tidying it up for LLM, LLM can then do a much better job in summarizing the response or generally making sense of the still large, but consistently categorized data set. Everything is now a combination of human design, build and orchestration and model design, build and orchestration.

Why does language coverage matter so much for spotting risk early?

For multiple reasons. First is the general volume of information. It is then difficult to filter out the noise from what matters to the client. Second is the impact of information. What might be considered low priority for others and thus may not be transferred to English sources might have a large impact on a particular company. Imagine a small producer in Asia who is disrupted. The producer may not be big enough for even nationwide media or more, the international media and the signal stays hidden in the local press. But it might be a critical supplier of a scarce component for larger companies.

Where should leaders start when monitoring supply chain exposure?

With data quality. Large supplier bases are difficult to maintain on their own. Because it is like a never-ending race. Once the record gets updated, it may be subject to another change. We can actually help even with that cleansing as company disambiguation is a core process when we match events to companies globally.

Then it is good to segment the supplier base not only according to geography and total spend, but also by criticality. Under criticality can be understood various other metrics, such as concentration ratio, replaceability, exposure to geopolitical threats, impact on ESG metrics and so on. The segmentation then helps navigating how deep the monitoring should go, frequency of updates and sensitivity to a signal. The more critical the supplier, the more exhaustive should be the n-tier mapping of its chain, the higher the frequency of updates and higher sensitivity to negative (disruptions as an example) or even change (mergers for example) events.

What to measure is then linked to BoM (Bill of Material) line. In order to produce goods, companies maintain BoMs and stock levels for each material and component in that BoM. The combination of stock level, production use in time and disruptive signals from supply chain monitoring allow for this optimisation task to be completed with the right information at hand.

How will AI and open-source intelligence reshape risk management over the next few years?

I believe external intelligence will become an integral part of business processes. Currently the processes rely mainly on internal intelligence (data from ERP systems, logistical systems, etc.). Enhancing the decisioning with external intelligence will assist in decreasing the impacts of disruptions by knowing sooner, hedge or insure properly or find alternatives upfront. The automation and AI greatly helps in completing these tasks as its data heavy and it also allows remediate the scarcity of human resources.