AI as a source of expertise?

—

Brille und Text 3

For some questions, there are no robust studies or citable sources available, leaving plausible estimates or reasoned assumptions as the only way to arrive at a useful conclusion. We therefore asked: can AI, in some cases, provide transparent and plausible assessments that come close to a form of “expertise”?

We encountered exactly this challenge while researching benchmarks for customer service staffing levels at a company in the electronics sector. Despite extensive research, we were unable to find any concrete, citable statements or studies indicating how many customer service employees would be appropriate relative to revenue or customer numbers. As an alternative, we turned to Google’s AI Mode. It initially confirmed that no corresponding sources were available. At the same time, however, the AI was able to develop a calculation based on a range of influencing factors and existing, indirectly relevant information. The result appeared internally consistent, logically structured and understandable.

We then tested more specifically whether the AI could distinguish between different industries. It became clear that the calculation incorporated different assumptions and influencing factors. For example, it plausibly derived and calculated that KPIs such as the number of customer service employees per 1,000 customers would vary across industries.

Our conclusion: This capability may indicate that AI systems can derive structured and context-sensitive assessments from existing information in ways that begin to resemble expertise. Human experts, too, do not rely exclusively on directly citable sources in practice, but often draw on experience, contextual knowledge and implicit assumptions. In that sense, it is not inherently implausible that AI-generated assessments without a direct source base can still provide some useful guidance.

At the same time, interpreting such results remains challenging. The limited transparency of individual assumptions and the difficulty of independently verifying the underlying logic are important limitations. A calculation that appears plausible is not necessarily robust.

By Ida Kandler