Artificial intelligence: challenges for the financial sector

A task force was therefore established by the ACPR in early 2018. It brought together professionals from the financial industry (business associations, banks, insurers, Fintechs) and public authorities to discuss current and potential uses of artificial intelligence in the industry, the associated opportunities and risks, as well as the challenges faced by supervisors. The purpose of this discussion paper, based on these discussions as well as on exchanges abroad or with other French players, is to present a first diagnosis of the situation and to submit to consultation the reflections that deserve further study to enable the development of these new technologies within a secure environment.

Artificial intelligence is a polysemous notion that tends to cover different realities as algorithmic techniques evolve: the report followed a relatively broad definition of artificial intelligence, including all machine learning techniques, but generally excluding robotic processes that automate repetitive cognitive tasks.

The first finding by the Task Force is that projects based on artificial intelligence are at uneven levels of progress and their development is often less advanced in the processes that a supervisor would tend to consider to be most sensitive. However, all conditions are met for a rapid and widespread development of artificial intelligence techniques in the financial sector: a growing awareness of the possibilities of exploiting data, which are increasingly numerous and varied; development of available technology offers (open source libraries, new specialised players, major technology providers, notably through the cloud…); multiplication of tests and projects.

There are many uses – in production, tested or just planned - covering most of the banking and insurance activities: from customer relationship (with the already very advanced rollout of chatbots and also opportunities in advice or explanation to customers), back office management (e.g. insurance claim management) to personalised pricing, risk management and compliance (fraud detection, anti-money laundering, cyber security, internal risk modelling for regulatory capital requirements). The development of these technologies is naturally not without risk: those inherent in the techniques used and those associated with their “disruptive” power. The first category relates to the risk of algorithm bias, increased by their complexity and the effects induced by the combination of the different underlying statistical and heuristic methods, as well as cyber risks. In the second category are risks related to the possible emergence of a small number of key players in the use of these techniques and the power relations - possibly systemic effects – that such a phenomenon would induce.

Against this background, supervisors have to deal with issues with strong differences in statement and time horizon.

In the short term, it seems important that the development of artificial intelligence in the banking and insurance sectors be accompanied by practical reflection on the minimum criteria for governance and control of these new technologies. This should allow for progress, among other things, on techniques to prove the reliability of the algorithms used (for both internal and external auditability), their “explainability” and the interactions between humans (clients, advisers, supervisors, etc.) and smart algorithms. It also needs to clarify, more generally, what good “governance of algorithms” might look like in the financial sector.

At the same time, supervisors need to remain alert to the medium and long-term impact of artificial intelligence developments on the market structure in order to anticipate the necessary changes in the performance of their mission.

Finally, the discussion paper discusses the need for increased expertise and cooperation of supervisory authorities to address these two types of issues.