AI-Supported Decision-Making and Trust
It is no longer news that artificial intelligence, in the form of Large Language Models (LLM), is being used in businesses. Currently, this applies in particular to the increasing automation of routine tasks, especially in the service industry. In law firms, for example, this means that tasks - such as summarizing the latest case law on specific legal issues - can not only be completed more quickly, but also that the need for human review of such results tends to be reduced. This automatically raises the question of how much trust can actually be placed in such LLM results. After all, it is, in a sense, part of the “DNA” of LLM to generate new content on its own without properly labeling it. While Retrieval Augmented Generation can significantly reduce the associated error rate, it cannot eliminate it entirely.
With regard to general management tasks, an LLM-based approach is currently being promoted under the term Intelligent Choice Architecture, based on a comparatively simple idea. Agent-based AI applications are intended to handle data analysis, pattern recognition, and the provision of purpose-optimized data, while managers focus on “higher value” decision-making tasks.
What sounds good at this abstract level is, however, fraught with pitfalls in practice. How complete or consistent is the data used in this process? Under what assumptions and for what purposes was it collected, and over what time period? Which angle should guide the analysis—profitability, cash generation, inventory efficiency, customer focus...? What should be considered relevant in the event of (foreseeable) conflicting objectives? And who takes responsibility for this?
Anyone who cannot answer these questions but is simultaneously pushing forward AI-supported decision-making processes should expect that, in the future, certain unforeseeable events will hit them all the harder. It’s the story of the old mechanical loading crane — experts immediately see the relationship between the jib, the load, suitable suspension points, and the crane’s stability. Since this relationship is free from subjective judgments, distortions, and time-related factors, it can be clearly calculated and translated into highly automated solutions that are technically superior and economically more efficient. However, this does not mean that one no longer needs to understand problem-specific interrelationships.
This applies just as much to AI-supported decision-making systems. At their core, they too address the question of how specific interactions should be structured, who assumes responsibility for the consequences, and what exactly underpins trust in the AI-generated results. coopartner also addresses these aspects of goal-oriented action and decision-making.
