Toward artificial governance? The role of artificial intelligence in shaping the future of corporate governance

Although national laws differ in terms of the roles and responsibilities assigned to directors, Cossin and Metayer (2014) identify three generic key roles of BoDs across all jurisdictions: supervisor, co-creator and supporter, hereby extending the traditional dualistic perspective of direction and control. We shall base our role definition on Cossin and Metayer (2014) and label them as follows:

As we focus on the impact of AI on the BoD’s decision making, we need to identify the key decision types a BoD usually deals with (there are other non-routine tasks such as crisis management or communication which we ignore for the purpose of this article):

To determine which types of decisions benefit most from being driven by AI, it is important to recognize an inherent feature of decision making. Decision making is always about consciously choosing between two or more options. The options can be either binary, e.g. yes or no, or multifaceted, e.g. options 1, 2 and 3. The choice always depends on the criteria chosen. An informed decision usually follows a similar pattern as outlined by Still et al. (1958). They distinguish between three phases, i.e. conceptualization, information, and prediction (Fig. 1). We apply this taxonomy and add a second level to distinguish the sub-processes even further (For simplification, we do not consider the decision feedback loop usually associated with decisions at this point).

In order to assess whether it is desirable to use AI in business decisions, we propose to apply the taxonomy of the different decision types according to Stacey (1992) which distinguishes between four types of decisions depending on the degree of certainty and agreement (Fig. 2):

How can this logic be applied to board-level decision making? The table below summaries which decisions are common, complicated, complex, and chaotic from a governance perspective (Table 1).

We have seen that board decisions are of varied natures when it comes to certainty and agreement. How can AI be utilized to facilitate or even drive board decisions?

AI is a technology in the making, as explained above. Although we are currently seeing important advances in AI, this is not the first and will not be the last wave of development. It is therefore important to look at AI from a dynamic perspective, i.e. consider the “trajectory of AI” (Armour and Eidenmueller 2019). As Tegmark (2017) notes, we are currently in the early phase of AI development.

At the highest level, we can distinguish between rule-based systems and machine learning (ML). The former requires that humans fully understand a given context and define the rules that the machine should execute, while the latter enables the machine to learn and derive conclusions based on a set of data and learning algorithms without requiring context understanding (in the end, ML could also be considered rule-based, if only for the rules underlying the learning algorithms).

Within ML, deep learning (DL) is currently the most popular approach. It owes its name to its architecture of “deep (or multi-layered) artificial neural networks—software that roughly emulates the way neurons operate in the brain” (Ford 2018, p. 10). Earlier, more traditional approaches to ML still require the extraction of features by humans. Within DL, three approaches are usually distinguished, i.e. supervised (SL), reinforcement (RL), and unsupervised learning (UL). Supervised learning, currently the most commonly used approach, requires well-structured and labeled training data to train the algorithms to improve AI-driven applications such as image recognition or translation. In contrast, the philosophy behind reinforcement learning is best described by trial and error, which is often used in board game simulations or. The challenge of RL lies in the large number of trial rounds required to achieve good results. Currently, the most promising, but also most challenging approach is unsupervised learning, where the algorithms are designed to “learn directly from unstructured data coming from their environments” (Ford 2018, p. 12).

All these approaches assume a separation of the machine from the mind. The next wave of AI could overcome this separation and be based on the assumption that machines and minds can be connected. Whether we will see the emergence of “neuromorphic chips” (Marsh 2019) or other ways to connect the mind to the machine, this type of AI is likely to unleash new potentials. We will call this futuristic category Mind Machine Learning (MML) (Fig. 3).

In this article, we focus on the impact of the technologies currently used and developed, i.e. SL, RL and UL, and the technologies currently being researched, i.e. MML. For simplicity, we will continue to use the term AI to refer to implications common to all the approaches discussed.

To determine the strengths of the different approaches to artificial intelligence, it is important to be clear about the scale, i.e. human intelligence, and how the process of intelligence development, i.e. learning, is compared between the mind and the machine.

To address the latter question, we will briefly compare human and machine learning cycles based on the decision model presented above. Both approaches assume that decisions are based on predictions of possible outcomes, whereas prediction is defined as “the process of filling in missing information. Prediction takes information you have, often called “data,” and uses it to generate information you don’t have” (Agrawal et al. 2018, p. 24). In both approaches, predictions are based on input data.

However, in contrast to human learning, ML assumes three types of data, which play a different role in SL, RL and UL: Input data (direct input to the algorithm leading to predictions), training data (used primarily to generate the algorithm), and feedback data (used to improve the algorithm over time) (Agrawal et al. 2018). While all three approaches to ML rely on input data, training data is crucial for SL, while feedback data is crucial for RL and UL (Fig. 4).

This brings us back to the first question: the power of AI. Here, it is important to discuss which learning approach is best suited for which decision type, as discussed earlier. While SL can be used for common decisions, it becomes much less effective for other decision types, i.e. complicated, complex or chaotic, given the need for relevant training data. RL, on the other hand, can be effective in automating complicated decisions based on past routines. However, because it relies heavily on trial and error and thus on feedback data, it is not very effective in handling complex or even chaotic decisions. For complex decisions, UL could provide clues, while for chaotic decisions, it will be even more difficult to rely on any known ML approach. Any machine that would succeed in mastering chaotic decisions would pass the final Turing test, i.e. possess artificial general intelligence as proposed by Alan Turing in his 1950 paper “Computing Machinery and Intelligence” (Levesque 2017). Or to put it differently, the current stage of AI faces a number of serious obstacles as summarized by Marcus (2018): There are a number of problems arising from the lack of transparency or transferability, the inherent inability to distinguish between causality and correlation, and inefficiency in terms of the amount of data required for valid predictions.

Here comes the hope for alternative approaches to ML, such as MML, as indicated above. However, the true potential and limitations of this type of approach have yet to be seen. The question of whether and when artificial general intelligence will become a reality is still up for debate.

The BoD will be guided by five key legal considerations in its decision to incorporate and integrate AI into its processes and procedures.

In addition to a number of legal questions, there are even more serious ethical considerations to take into account (Simonite 2018). The five key ethical questions can be summarized as follows:

Applied to the context of corporate governance, the application of assisted intelligence does not influence the basic principles of board practice but rather provides board members with additional tools to support their decision making. The human decision-maker, i.e. the board member, remains at the center of power. The implications along the three key functions of the board of directors are as follows:

The use of augmented intelligence neither calls into question the fundamental nature of corporate governance, nor does it deprive board members of power. Rather, it promises to make decision-making not only more efficient but also more effective, as intelligent technologies are used to deliver better results. Along with the three key functions of the board of directors, the following implications arise:

The trend towards amplified intelligence is beginning to challenge some of the basic business and legal assumptions of corporate governance. As amplified intelligence assumes an evenly distributed contribution and accountability of man and machine, the human-centered approach to governance is being challenged. There is a need to discuss how to deal with the basic assumption of corporate law, i.e. personal accountability for corporate liabilities, when effective decision making is shared. Who should be held accountable for what? In this scenario, the implications along the three dimensions are as follows:

In the autonomous intelligence scenario, all or some board members are replaced by machines, either by a governance robot (a robot representing the board of directors) or by robo-directors (various robots that simulate contributions from individual directors), depending on how much the legal framework changes. At the same time, it is still assumed that people will retain responsibility for determining the scope of corporate governance. There is already a real case study in Hong Kong where Vital, an AI system, is one of many board members in an investment company (Burridge 2017). As highlighted by Armour and Eidenmueller (2019), this scenario is most likely realized in subsidiaries first. The effects on the three key roles in this scenario are as follows:

In a scenario where autopoietic intelligence dominates, the governance of a business entity is not only fully automated to ensure proper governance within a certain framework, e.g. the representation of clearly defined shareholder interests, but also capable of driving its future development and interacting proactively with other business entities. In this scenario, the impact on the three key roles is as follows:

Artificial governance in Horizon 1 is very much driven by advances in supervised and, to a lesser extent, reinforcement learning, which mainly leads to assisted and augmented intelligence. In both cases, correct and sufficient data is key to performance, either as training or feedback data.

The governance of data is the key focus when it comes to governance of AI in Horizon 1. Companies increasingly understand the strategic value of data as training, but even more importantly, as feedback data (Iansiti and Lakhani 2020). For many companies, data is becoming the most valuable asset (Mayer-Schönberger and Ramge 2018) in line with “the rise of the intangible economy” (Haskel and Westlake 2018). It is essential to understand the economic characteristics of data, in particular, that the data is not competitive, has a high option value, often has high up-front and low marginal costs, and requires complementary investments such as technology and qualified staff (Coyle, Diepeveen, Wdowin, Kay, and Tennison 2020). Therefore, the governance of these key assets is of strategic importance and needs to be addressed at the board level (Hilb 2019). At the same time, data-related issues pose major risks and increasingly determine corporate culture and the relationship between companies and society (Mantas 2019).

At the same time, it is important that the board recognizes that AI does not only affect the business but also the board itself, i.e. the governance with AI. This awareness is necessary to make the BoD more data-driven (Libert et al. 2017) to take full advantage of the benefits of what Davenport and Ronanki (2018) call the “future cognitive company”. A key focus in Horizon 1 is on reporting and controlling, as AI enables sample-based audits, now data and stronger predictive power of data (Mantas 2019). This will eventually lead to a reduction in the cost of internal control, but also to increased supervision and liability risks beyond the company (Armour and Eidenmueller 2019).

The impact of artificial intelligence on corporate governance will reach a new level in Horizon 2, mainly through the development of reinforcement and unsupervised learning, which will extend the impact of artificial intelligence lending support to augmented, amplified and autonomous intelligence. For the first time, the vision of a self-propelled organization is more than just science fiction. Systems will be able to exceed human learning capabilities not only in narrow areas but also in broader, interrelated decisions, as described above.

In such a context, the algorithm becomes the focus of attention of the governance of AI. Those organizations that are able to develop and apply an algorithm that enables reinforcement and, more importantly, unsupervised learning, will be best positioned to benefit from it. Since no single organization will be able to develop or use such algorithms in isolation to achieve their full potential, collaboration between organizations in ecosystems will be critical (Hilb 2017). This results in some legal issues, such as liability for the algorithm, i.e. should those who use or build it be held liable for its result, antitrust law, i.e. can a superior algorithm lead to anti-competitive behavior, especially when companies join forces to develop and use it, or related regulatory issues such as tax law, i.e. how should the value-added of algorithms be taxed? At the same time, in-house procedures and processes must be put in place to ensure that all these questions are addressed properly, or as Armour and Eidenmueller (2019, p. 36) put it: “(T)raditional agency problems in corporations will increasingly lose importance as an analytical issue: algorithms optimize a given goal function. At the same time, new oversight challenges and, as a consequence, ‘strategic’ agency costs are created at the top of a firm’s hierarchy.”

Horizon 3 of artificial governance still holds many unknowns, since we do not know how AI technology will evolve, nor what rules society will impose on those who develop and use it. What we do know, however, is the direction in which we are heading. There will be approaches to artificial intelligence that go beyond neural networks and include direct connections between the mind and the machine (Marsh 2019). This approach is likely to be disruptive in nature as defined by Christensen (1997), i.e. to achieve a new level of efficiency and effectiveness, whereby old approaches eventually become obsolete. These approaches are likely to bring us closer to what is known as general intelligence (Marsh 2019) that is, what we call autonomous and autopoietic intelligence.

This will also shift the governance with AI to new levels. AI will enable the organization to control and direct itself more efficiently and effectively than any human being could. There will, therefore, be economic pressure from the capital market to adapt to AI-driven forms of governance. This model is likely to challenge the basic tenant of today’s corporate law: The personal liability of the directors of limited liability companies (Moeslein 2018) (Table 2).