How Are Large Language Models Shaping the Future of Freight?

Supply chain pressures rose marginally in January . The good news is tensions in the Middle East have not yet created widespread disruptions. The not-so-good news is that geopolitical risks remain threats—especially in Europe, where rerouting means longer shipping lanes than in the US.

Yet, even in an ideal, frictionless supply chain, the global supply chain would still face challenges. That’s because data, like the universe, accelerates and expands, and it’s endless. In fact, companies create 328.77 million terabytes of it across all sectors. Daily. While in the logistics space alone, 7.3 billion datasets are exchanged yearly.?

But volume isn't the problem here. It’s what that data looks like. Unstructured data such as spreadsheets, PDFs, and scanned images in emails are gibberish to traditional analytics tools (which almost everyone in the supply chain is still using.)?

And while general-purpose Large Language Models (LLMs) like ChatGPT can’t solve the global supply chain’s data problems. Domain-specific LLMs trained specifically on industry data can be used. More about that in a moment…

Five reasons why LLMs are vital to smart supply chains

Today’s supply chain is a high-stakes race against time and uncertainty. Every hour a shipment sits idle costs money. Every mistake in a document can cause delays and fines. How do LLMs eliminate these risks? They analyze vast amounts of data precisely, uncovering the answers supply chain stakeholders need to take swift, informed action.

Automation and process optimization

If you’re a freight forwarder, an LLM can automatically extract critical details from your shipping documents, such as bills of lading and commercial invoices. By reducing the need for manual data entry, it also reduces human errors (30%). It can also generate standard operating procedures, checklists, and templates based on natural language inputs, streamlining document creation processes.

Predictive capabilities

For shippers, an LLM trained on internal historical shipment data, market trends, and customer order patterns can provide accurate demand forecasts for different products and regions. Thanks to LLMs, shippers can optimize inventory levels, production planning, and transportation needs.

Cognitive automation

Let’s say a logistics provider integrated LLM with existing RPA systems to handle complex claims processing tasks. The LLM would extract data from claims forms AND cross-reference with historical logistics data to identify patterns in claims fraud, enhancing the accuracy and efficiency of investigations. The game-changer is the LLM learns and improves, continuously adapting to changes in customer preferences, carrier rates, and market conditions.

Adaptability and flexibility

Just as ChatGPT has released updates, LLMs can be fine-tuned and retrained on new data sets, keeping them updated with evolving industry trends, regulations, and best practices. For example, supply chain managers can retrain their LLMs to comply with the new rules in response to sudden regulatory changes affecting customs documentation.?

The limitations of LLMS in the supply chain?

While LLMs offer incredible potential, general-purpose models aren’t ideal within complex, specialized industries like freight, logistics and marine insurance.?

Lack of domain-specific knowledge?

Imagine a freight forwarder needs to generate a commercial invoice for a shipment of hazardous materials. A general LLM lacking specific training on supply chain regulations and documentation could generate inaccurate or incomplete information, leading to compliance issues and potential delays or penalties.

That’s because the LLM might not know the specific data fields required for hazardous material shipments, such as the UN/NA number, proper shipping name, hazard class, and packing group. It may also lack the knowledge to correctly apply the relevant regulations, such as the International Maritime Dangerous Goods (IMDG) Code or the Dangerous Goods Regulations (DGR) of the International Air Transport Association (IATA).

Consequences of inaccurate documentation?

While marine insurance companies can benefit significantly from LLMs, relying solely on the technology to streamline document analysis without oversight introduces risks. A general LLM might misinterpret the terms and conditions due to outdated data or a lack of specific training. This could result in insurance policies falling short of regulatory standards, exposing the company to financial risk.

Limitations in complex problem-solving

General LLMs may not have the context to handle complex problem-solving scenarios in the supply chain. In my experience, if a critical component for a US manufacturing plant is stuck aboard a container ship held at a congested Chinese port, a general LLM might miss these crucial nuances:

• It may suggest rebooking on the next vessel, unaware of the frequent transshipments needed on that route, adding weeks to the delay.
• It might propose finding a different supplier globally, not factoring in the time to qualify a new vendor, and likely affecting their lead times.
• It could recommend air freight as the fastest option without understanding that, due to the component's size, it’s a more expensive option.

Data privacy and security risks

LLMs often require large datasets for training and fine-tuning. Supplying a general-purpose LLM with sensitive supply chain data raises privacy and security concerns, as malicious actors can exploit LLMs.?

That could happen if confidential client information or internal business strategies are included in the training set.? By training on such sensitive data, the LLM could inadvertently memorize and potentially leak private information through its outputs.?

This raises serious data privacy concerns, as confidential details about shipments, clients, and business operations could be exposed, violating non-disclosure agreements and compromising competitive advantages.

In addition, LLMs can be vulnerable to adversarial attacks, where malicious actors attempt to extract or manipulate the model's knowledge. Accessing the trained LLM could extract sensitive information embedded within the model's parameters.

How do domain-specific LLMs solve this?

As we’ve seen, while generic large language models like ChatGPT can unlock impressive capabilities, supply chain data is complex, and compliance demands precision, blunting their impact within this specialized sector.?

Domain-specific LLMs are designed to address this gap. Anticipating the need for tailored solutions, Stargo's LLM development began four years before the ChatGPT surge. Unlike ChatGPT's broad focus, Stargo's LLM approach reflects four years of specialized development within the logistics and supply chain sectors, resulting in a deep understanding of industry-specific challenges.

The first LLM fluent in freight

Stargo Large Language Model – the world's first LLM trained exclusively on freight and logistics data to understand the nuances of freight terminology. Terms like "demurrage," handwritten UN/NA numbers or the intricacies of LTL shipping won't leave it confused.?

SLLM is also constantly updated on changing compliance requirements, minimizing the risk of costly document errors. It also thinks strategically. Facing a delayed shipment? It analyzes real-world carrier routes and port congestion and suggests viable alternative suppliers, ensuring you always have a plan B, C, and D.

How SLLM learns (and why that matters)

The Stargo Large Language Model (SLLM) is trained exclusively on over 1 million samples from real-world freight and logistics-related emails and data sources. This extensive dataset provides the SLLM with deep domain knowledge that is context-specific to the freight industry.

Crucially, the data is thoroughly anonymized and secured to protect sensitive business information. By self-hosting the SLLM and implementing robust encryption, access controls, and monitoring mechanisms, your sensitive data to train the model remains secure—and under your complete control (not Stargo’s).?

SLLM excels at complex problem-solving?

Remember our use case where a general-purpose LLM struggled to understand the nuance of a critical component stuck aboard a container ship held at a congested port on its way to a US plant?

When seeking alternative suppliers, SLLM doesn't blindly spit out search results. It factors in compatibility variables and the time needed for qualification. Additionally, while air freight might be expensive, SLLM can calculate whether speed should become the top priority.?

The road ahead?

Despite their advantages, LLMs are not magic wands. The accuracy of their outputs heavily depends on the data quality on which they are trained. Poor data quality or biased datasets can lead to errors that might amplify if not checked. Furthermore, the reliance on large datasets poses inherent privacy and security risks, necessitating robust measures to safeguard sensitive information.

Integrating domain-specific LLMs offers a promising path forward as the logistics sector evolves and evolves again. These advanced models can transform data into a strategic asset, driving innovation and resilience in supply chain operations.