How does an AI think about taxonomy?
Emerging technologies are ‘all the rage’ in procurement today – especially for rules-based tasks that have to be quickly and reliably executed at scale. Transaction categorization is no exception from the trend. There are a number of ways to leverage technology for categorizing procurement activity, but they are not all created equal. Some employ top-down business rules ‘under the hood’ while others leverage AI, and while the end result might appear the same at first glance, this could not be further from the truth.
When taxonomy information is assigned to a transaction, the goal is to represent an accurate and actionable truth about the business activity. Although users most comfortably interact with a taxonomy from the top down, that is not the most effective way to assign taxonomy at the transaction level. Top-down taxonomy assignments rely upon broad details borrowed from supplier profiles or general ledger codes. While these pieces of information are related, they are inferred rather than being actual descriptors. They lack the granularity modern procurement organizations require. Over time, these assumed assignments create compounding problems for the enterprise and lead to mounting data inaccuracies and a lack of trust in the data and resulting business recommendations.
Historically, the rationale for taking a top-down, business rules-driven approach was the challenge of categorizing detailed transactions quickly and at scale. Fortunately, neither of these is a problem for AI. If we can better understand how AI ‘thinks’ about taxonomy, and how it approaches spend, suppliers, and transactions, then we can achieve a high level of categorization accuracy – not for its own sake, but for the sake of business strategies and decisions.
AI thinks detailed data is best
AI can assign standardized taxonomy information based upon the PO line description for individual transactions. It does not need to start high and drill down through the tiers like a human might or a business rule has to. AI does not have to assign all lines in a PO to the code associated with the supplier. Not even cryptic codes or multi-lingual terminology present a challenge. AI can work with an unstructured, user-defined description such as “Schraube DIN 933 M60x10” and derive enough information to categorize the transaction accurately as a hexagon head screw.
AI is able to convert non-standard human descriptions into descriptions that align with a standard taxonomy because it ‘learns’ from multiple sources of information. Chief among them is the largest multilingual knowledge database of industrial components. Layer on top of this the fact that AI has the experience gained from analyzing 100 million descriptions, and it is no longer necessary to choose between granular accuracy and speed.
AI thinks your custom taxonomy is fine
There are several commonly applied standard taxonomies that companies use: UNSPSC and eCl@ss are two examples. It is important to remember, however, that taxonomies are not ‘function agnostic.’ For example, UNSPSC was designed for eCommerce, while eCl@ss is ERP-based. When a company looks at their transactions categorized into a standard taxonomy and realizes it doesn’t fit their purpose, the problem is not the taxonomy; it is the intent of that taxonomy. A common solution to this problem is to take a standard taxonomy and modify it to fit both the company and the need. For instance, a general ledger-driven taxonomy emphasizes a financial point of view, while bills of material (BOMs) categorize materials as they will be used in the production process. Neither is wrong. They just have different objectives, and AI thinks either (or both) is fine.
AI doesn’t think much of complexity
As with transaction volume and complexity, AI is undeterred by the size and scope of the taxonomy itself. It will perform just as well with a 4, 5, or 6-tier, 150,000 commodity taxonomy as one with just 2-3 tiers and 5,000 commodities. While standardized taxonomies offer a great deal of depth (UNSPSC offers 5 levels, while eCl@ss offers 4), most companies do not need the same degree of granularity in all categories. AI is not fazed by this, thinking only about the information available, not the differences in taxonomy.
The value of having well-categorized spend is not a high automated categorization percentage, although that is good to have. It is more important for the categorized spend to empower the company to act appropriately and decisively because they can trust their data, from the most granular level to the aggregated top level. Procurement’s best bet is to understand how AI thinks about taxonomy and work with that, not hampering the AI by assigning it human perspectives.