How AI-led SKU categorisation strengthened ERP migration for a global hospitality group

As part of a major ERP transformation, a global hospitality company was faced with the challenge of migrating a highly complex SKU catalogue. A manual approach risked creating a major resource drain, slowing down the programme, and introducing errors. While AI offered a way to address the scale of the challenge, success depended on accurately interpreting the organisation’s business-specific taxonomy and category rules. To achieve a high-quality migration, the organisation partnered with Efficio to combine its procurement and AI expertise. 

The challenge 

The organisation was migrating its entire SKU catalogue into Dynamics 365, requiring each item to be categorised through a five-level taxonomy. With the catalogue spanning millions of SKUs, manual classification was not a feasible solution. To further complicate the task, the categories were based on a taxonomy that demanded detailed interpretation of business documentation, accounting dictionaries, and product attributes.  

Efficio’s approach 

Efficio developed a custom AI-enabled SKU categorisation solution based on a five-layer large language model (LLM) architecture. Efficio utilised their deep procurement category knowledge to refine the client’s taxonomy to enable the LLM to accurately interpret it.  

Bringing together procurement category knowledge, a strong understanding of the client’s business context, and technology expertise, the Efficio team was able to ensure accurate interpretation of the client’s taxonomy from the outset. This meant the solution was built in weeks rather than months, with the number of feedback loops with the client kept to a minimum.  

The solution translated SKUs to English where necessary and categorised items for all levels, enabling a structured final review. The result was a scalable, automated categorisation tool that enabled the team to categorise each hotel’s historic catalogue items in under three hours with strong accuracy: 5–10 times faster than traditional, non-AI approaches.