Using AI to simplify parcel movements after EU Exit

The challenge they faced

From 1 May 2025, a new requirement under the Windsor Framework came into effect, mandating that parcel operators provide accurate 6 to 8 digit commodity codes for all parcels moving from Great Britain to Northern Ireland. With over 120 million parcels a year affected and an industry dependent on speed of movement. HMRC needed a high-performance, resilient solution that ensured compliance while keeping parcel movements as simple as sending them from London to Manchester.

The challenge: enable trade to continue at pace — seamlessly and compliantly.
More specifically:

• Scale and accuracy: Billions of goods descriptions required classification against a vast and complex tariff schedule.
• Balancing compliance and facilitation: The UK Government needed to uphold international commitments under the Windsor Framework while maintaining trade and facilitating wherever possible.
• Operational pressure: The system had to go live quickly and operate at national scale without compromising reliability.
• Human dependency: Traditional manual or rule-based approaches couldn’t keep up with volume or policy change.

How we helped

Experimenting to find a solution
Partnering closely with HMRC, we ran a series of rapid experiments to explore how emerging AI models could support classification at scale. This included bespoke GPT-style models, improved search techniques, and the use of Large Language Models (LLMs) built on sentence transformers — deep learning architectures that capture meaning rather than just keywords.

Matching text to code
Our solution automatically takes parcel descriptions and identifies the most likely commodity codes in order of statistical probability by transforming text into high-dimensional vectors that capture semantic meaning. The model was trained on millions of real trade description–code pairs, learning to identify product names, brands, and variations — even in messy, user-submitted data.

Data processing and model training
We built automated data-cleaning and assurance pipelines to remove unverified records and continuously improve training data. Using DevOps engineering practices, we enabled automated training, testing, and deployment so the system evolves as new data becomes available.

Keeping the human in the loop
Classifying goods remains complex. We designed a verification tool through which HMRC staff validate the model’s classifications, allowing the system to learn and improve accuracy over time. This ensures a balance between automation and accountability.