An issue that lies with AI is related to its black box nature when solving problems. It will not be precise on the indicators driving a solution. In general, Elin Hauge, a self-employed AI and business strategist, thinks we should first look at the problem we’re trying to solve and then only look at the toolbox which includes AI, among many other tools. Solving an issue is sometimes “just about making data flow work properly between two different systems.”

She observed that the decision of many leaders to use AI is “very much” driven by consulting companies. “If McKinsey says it, it must be true, right?“ Hauge commented that understanding the dynamics of data requires a lot of competence in an industry, its data and the value chain. Yet such a background helps when deciding whether mathematics based on stochastic modelling should be used or not.

AI in the financial sector

While discussing the relevance of using AI for predicting some market indicators such as stock or interest rates, AI may be more appropriately used to detect outliers, Hauge suggested during an interview on 27 November 2024, following her presentation at a European Investment Bank conference. “It won't tell you why it’s weird, but it will tell you this data point is weird.”

She remarked that tech experts at financial institutions “use some kind” of large language models when drafting internal code development. “That has nothing to do with the trading.”

Can insurers and banks get something from AI?

She noted that at insurers, machine learning or “good old stochastic modelling” are used for “micro tariffing of individual risk level.” Elsewhere, fraud detection on claims is another sector that benefits from AI. Hauge explained that machine learning helps to identify “patterns of behaviour in known fraud cases,” also called “pattern recognition.” In the financial sector, the credit card industry looks at the issue from the opposite angle, i.e., “a deviation from a typical pattern.” An alarm would go off if your card was used in the Bahamas, for example.

She thinks that “good old classical statistics works really well when you have a predefined data set” to perform regression analysis. On the other hand, Hauge claimed that when one is confronted with a “bigger data set and you’re not really quite sure which data are more important to you and which ones are less important in the pricing, then machine learning is a way of going about that.”

A surprising statement for your correspondent given that there are statistical methods to select relevant variables. Hauge confirmed that the approach or the “underlying mathematics are the same,” but doing it with AI is a faster way to get to your goal. Yet one may lose information about the material indicators along the way.

Identifying the relevant indicators and their impact. All is not lost.

“There is a way of somehow solving that, and that is using heatmaps in segments of your neural network, and AI is one of the like sub-domains of the technology that is currently being developed and researched,” said Hauger. She specified that neural networks with “a lot of layers are called deep learning [systems],” which is the method for most of these models.

She noted that regulations require transparency on data and the use of the algorithms for “higher risk application.” Therefore, heatmaps attempt to assign weights to the nodes to the neural network, “a kind of black box.” However, she admitted that it is not “mature enough… but an area in progress.”

Humans are still needed to back up AI

Hauge commented that “any of those models” provide stochastic predictions. She suggested that an output with a 95% level of likelihood may mean that “you still need to have a human in the loop to actually look at the case.”

Fully relying on the outcome of the model may, indeed, backfire. Hauge reported that the Dutch Tax and Customs Administration four years ago wrongly accused 26,000 families of child benefit fraud. The Dutch government resigned over the scandal.

The consequences of AI being wrong for music selection on Spotify is much lower than using it in the financial sector or after a medical examination.