The UKGoverment are keen to encourage every industry to use AI so the country can achieve an increase in productivity but – does a Generative anything belong in an industry that relies on accuracy and integrity?
To evaluate this, we looked at three potential use cases for frieght and customs, based on products that are on the market for freight forwarders to use.
The Generative Customs Audit Usecase
Forget the naysayers that preach this is the wrong tool for the job and that machine learning approaches are more efficient, it’s easy to see the logic. AI is a quick and easy fix to a problem in the industry. You could realistically throw togeher a couple of agents and a workflow to pull this together in a couple of days – maybe even a couple of hours if you’re persistent.
- Extract Customs data (PDF document or HMRC’s new TRE service)
- Extract evidence data from things like invoice and packing list. Rarely clean and simple but practically do-able with OCR and structured data prompts.
- Match the commodity using UKGov API’s and pull the requirements for that code (controls, duty, vat).
- Throw in a few other end points.
- Either put it all into an LLM for analysis and request a readable data format back or, spend time doing a bit of traditional programming checks
It’s likely the company providing an AI customs audit system pushes their data points into an LLM to get to market quick. Known as a gpt-api-wrapper. It does the job of a low-performance worker. The skeptasim should be focused entirely on hallucinations and context. An LLM will fill in blanks with it’s training data – muddying your audit data with fabricated information and, LLMs lack of context-awareness – the ability to understand the reasoning for the data existing in it’s structure.
In October 2025, BBC found that LLM’s misinterpret the news 45% of the time and a further study reported by The Register showed LLM’s failed at up to 70% of office tasks – the latter being the most relevant as sellers of AI customs audits use an off-the-shelf LLM such as gpt or claude to perform analysis.
The Data Extraction Usecase
Data extraction – this is where data is pulled from a set of trade documents, normally PDF (if pdf is digital, or using OCR or a visual LLM) and then it is sent to an LLM to structure the data. Typically this would be for invoices, packing lists, export documents etc. We employ data extraction tools at Kinnes Shipping for select customers and to be fair, it works from time to time although it does not nessesarily speed up the workflow as corrections and checking take as much time as manually extracting data.
The issues with data extraction methods:
- Multi document PDF files – international trade often scan all documents into a single non-digital PDF. This throws the document extraction out the window.
- Generative by nature – even with low temperature figures, an LLM will fill in the blanks with random values if it is unable to find an appropriate match.
- Data security – while there are agreements in place for API user data not to be used in AI training, you’re relying on big tech companies “trust me” ideaology.
Data extraction provides a small boost to productivity in certain use cases however this is negated by the time it takes to manage incompatable documents structures and conjoured illusions injected into the end data.
The Chat Usecase
The chat usecase is an unusual way to implement AI. It’s unorganic and from a data management point of view, chaotic.
Take a moment to ask an LLM what is the CPC for re-import of returned goods, commonly refered to as RGR and it’ll fabricate an answer – you can be more specific and you may eventually get the answer but, you’ve now wasted 15 minutes trying to explain to a chatbot specifically what you mean when you could have pulled the informaton from UKGov website within 2 minutes.
Chat as an interface could work providing embeddings (fine-tune training) are used, or there is a large reference library to search and use as context – this reduces hallucinations. Tack on multiple tools that pull in data for example, UKGov data end points and this would remove the LLM needing to guess, which limits the lies. However, it will still fabricate information as the training data bleeds in and that is the real problem.
The problem with this: The next point of cognitive surrender and a decline in problem solving skills is relevant to The Chat Usecase, however the main issue of using The Chat Use case is the liability in the returned data, and the reliance of the data within an environment of responsibility.
Problem solving skills suffer and our ability to seperate fact from fiction erodes.
Cognitive Surrender
As detailed on ArsTechnica. Reliance on LLMs results in the surrender of our cognitive function. Our willingness to use LLMs as a first step in research rather than a last is causing cognitive decline as we begin to lose our logical thinking. Problem solving skills suffer and our ability to seperate fact from fiction erodes. In an industry such as freight and customs, this would severely hurt clients and the industry.
Conclusion
Many freight companies have jumped into the AI pool with two left feet although there is few reports of productivity gains or benefits to international traders. For us, the water looks murky and the apparent benefits are questionable. From reviewing the available systems and processes in the market, the use of AI in the freight and customs industry appears to be a risk to international traders.
If your freight forwarder has adopted generative AI as a mechanism in their business – can you accept their work is accurate? can you rely on them to problem solve effectively?
We, as a company, are keen to leverage any tool that improves results for our clients. However as a compliance first freight forwarder, we will not employ systems that may impact the accuracy of our data nor a process that could have a negative impact on our clients and/or their shipments. We will continue to focus on people, training, traditional automation, the human connection with our clients and improving our problem solving skills as we truly believe this is the best approach for our customers.
