Artificial intelligence is always touted as the solution when problems are highly complex. Fueled by the impressive capabilities of large language models (LLM) such as ChatGPT, Copilot, and Gemini, depending on the problem at hand, you encounter a mixture of absurd rejection and unrealistic expectations.
There are undoubtedly interesting approaches to supporting strategic purchasing control with AI, but its use is not a sure-fire success. Much of what is currently passed off as AI is conventional, rule-based programming, whereby data input is converted into results via software rules. So far, nothing new.
Artificial intelligence comes into play where no rules are known. There is a lot of data available, but how are data input and results related? Are there data patterns and relationships? The idea behind AI models is to simulate known results from data input until a trained model can make predictions. An LLM, for example, calculates its answers from fed mass data, and what is presented in such a knowledgeable manner is merely probability linked to keywords. As long as topic-related, cleaned mass data is available that can be used to train a model, the use of artificial intelligence for purchasing controlling and purchasing management offers inspiration, insights, and proposed solutions.
Product group and industry structure
The machine-based review and improvement of product group assignments is ideal for the use of an AI model, as this task tends to be unpopular when performed manually. A prerequisite for machine support is a learning model with product group assignments that are as accurate as possible, combined with suitable attributes in the material master. Since product groups only work where there are orders, automatic industry assignment to vendors as an additional categorization has proven to be useful. In both cases, the cleaner and more meaningful the learning model, the more promising the use of the AI model for future automatic recommendations. Investigations in our WebCIS modules have shown that automatic assignments and correction recommendations with a quality score of over 75% are possible. The increase in quality achieved through AI-generated categorization enables more homogeneous clusters, which allow meaningful price benchmarks within product groups and to external price indices. The classic purchasing manager question, “In which product groups and with which suppliers is there potential for renegotiation?” can thus be checked automatically. How can technically similar parts be identified in order to check for standardization and highlight unjustified price differences? AI-based similarity scores derived from technical characteristics in the material master provide helpful data patterns for new ideas for action in every case. Quantity-dependent prices, known as graduated prices, are common in many areas of purchasing. However, the question of how to use them sensibly in scheduling arises time and again. How do increases in quantity affect invoice prices? Expected correlation effects between volume increases and price reductions are usually not as clear-cut as often assumed. AI-based data patterns reinforce the assumption that there is fact-based room for negotiation that can lead to significant savings.
Objectification of suppliers’ price demands
The question of the extent to which prices and price change demands from suppliers are in line with the market is as old as purchasing itself. Cost breakdown is considered a common analytical but time-consuming approach. AI models, on the other hand, enable mass analysis. Price components are features of a product that make up a total price with different weightings. But what is the impact of these features? AI models simulate the characteristics that influence the price until mathematical patterns become apparent. These multi-regressions, which are based on statistical correlations rather than technical effects, have the advantage of providing insights automatically and quickly, but – as with all AI topics – they depend on a good learning model.
Global risk assessment
What has long been standard practice in banking and insurance—namely, using AI to classify borrowers’ creditworthiness and risk assessment—is also possible in purchasing as an additional approach. A purchasing department has sufficient features – delivery reliability, complaints, external scores, audit results, country scores, product group benchmarks, price indices, sourcing information, and so on – to feed a learning model that highlights warnings about the supply chain. In addition to an analytical risk assessment with weighted influencing factors, impending surprises in the supply chain quickly become visible. A powerful information cube such as WebCIS AI offers analysis, information, and reporting features in structured dashboards, customized favorites, and on-the-fly reports for just about any purchasing topic. The key question is which feature fits my question. The integration of an LLM into the WebCIS interface will provide the appropriate suggestion for prompting. The extended prompting analysis also passes filters on to the respective features. The advantage for the user: greater breadth and depth of information while simplifying operation, as continuous text is intelligently processed.
Werner Güntner, Managing Director, Softcon CIS