AI in the Boardroom: How Much Oversight Remains Human?

Let us take a hypothetical situation: A medium-sized company wants to enter a new market. Management has calculated demand, investment requirements and staffing needs. The supervisory board receives a coherent proposal.

An AI system summarises the documents. That may save time. It would be more interesting, however, if it found an overlooked supply dependency in accessible earlier documents and compared it with current information. Whether that dependency still exists, and which market assumption has changed, would then become questions for targeted review.

That would not be a decision. But it would provide a different basis for the conversation.

AI could do more than make oversight easier: it could shift who identifies, challenges and exercises judgement within a company. Whether that leads to better decisions is a different question, and one for which the evidence remains limited.

The situation is familiar even without a supervisory board: A management team or a group of founders reviews an investment and has to question its own reasoning. This is not the same institutional role. The shared question is: Which aspects of identifying, challenging and exercising judgement can AI change, and what remains unresolved?

Not just overseeing AI, but working with it

People who lead, advise or oversee a company need a sound understanding of it. Reports and metrics are part of that, but so are the connections between them: What did we assume previously? What has changed? Which alternative did we not pursue? Which question do we keep asking without following up on the answer?

AI can contribute at different points: retrieving information, condensing documents, comparing scenarios, preparing critical questions or making earlier decisions easier to find. These possibilities extend beyond the control of AI systems. They concern the work of the people who steer the company.

A practice report by the Australian Institute of Company Directors and the Human Technology Institute describes early applications before, during and after board meetings. It also documents how immature some experiments remain: The Real Estate Institute of New South Wales tested an AI board adviser using internal documents and governance information. At the time of publication, it was not yet a significant contributor to meetings.

That limitation is precisely what makes the case interesting. A digital participant at the table is not evidence of better oversight. The report is an interview-based account with case examples, not a controlled evaluation of impact. Its Australian legal guidance cannot simply be applied to Germany.

The real change may happen between meetings

An INSEAD author team including Ludo Van der Heyden, founder of the INSEAD Corporate Governance Centre, envisages a fundamentally different way of working: information work less tied to individual meetings, and more continuous, forward-looking engagement with the company.

Among other possibilities, the authors see AI reducing the board's dependence on management for information. Their position is optimistic. It is not, however, empirical evidence that AI is already producing this effect on a broad scale.

For me, it nevertheless raises an important question: What changes when a governing body receives information not only at scheduled intervals, but can examine particular assumptions more deliberately between those dates?

In the hypothetical expansion case, this might mean that the company does not simply track ever more metrics. Instead, it monitors the specific conditions on which the decision depends, such as a partner's ability to supply or a particular change in customer demand. An agent could connect several steps: checking approved sources for changes, comparing them with documented assumptions and preparing a question supported by evidence. This is a possible way of working, not a capability demonstrated here in a particular system.

That need not mean a permanent meeting or a supervisory board interfering in day-to-day operations. The new possibility would be to use relevant changes more deliberately as a trigger for advice or oversight. Whether this works also depends on data quality, attention and available expertise. Another stream of alerts could just as easily create more overload.

Could this lead to smaller boards — or none at all?

The question of replacement cannot be answered simply by saying that humans must remain responsible.

Corporate governance scholar Martin Petrin explicitly develops a more far-reaching scenario. He assumes that AI will also take over demanding forms of human judgement. Under that assumption, boards could become smaller; eventually, management and oversight could merge into a single AI management entity.

This is not a description of corporate management that works today. Nor is it a prediction that German supervisory boards will be abolished in two years. Petrin explores the institutional consequences of a far-reaching assumption about technological capabilities. The international board models discussed in the sources should not be equated with Germany's separation of management and supervisory boards; the scenario makes no statement about the powers of corporate bodies under current German law.

We should therefore neither dramatise the question nor close it prematurely. Four developments need to be distinguished:

• Automating tasks: Finding, comparing and preparing information.
• Changing the work: Different questions become possible, reviews more frequent and information flows different.
• Displacing roles: Some contributions may require fewer people or different capabilities.
• Replacing a corporate body: Powers, legitimacy and responsibility are reconfigured institutionally and legally.

The first development does not automatically lead to the fourth. Equally, the continued existence of a corporate body does not guarantee that its current work will still be needed in the same form.

Who oversees the AI that does the overseeing?

Legal scholars Luca Enriques and Dirk A. Zetzsche offer an important counterargument here. They warn against comparing supposedly perfect machines with real, fallible humans. Their central argument is that conflicts of interest do not automatically disappear when technology takes over information processing.

Whoever selects the system, determines its objectives and controls its data can still exert influence. The institutional question therefore remains, even if models become technically much better. This is not a blanket argument for human superiority: Enriques and Zetzsche explicitly acknowledge that better monitoring outcomes than those achieved by human boards could justify algorithmic boards. Their objection is to the assumption that better technology will eliminate conflicts of interest by itself.

Applied to our expansion example, an AI system given only the same prepared proposal does not thereby gain an independent perspective. It may simply express the existing rationale more persuasively. Even additional data helps only if its provenance, selection and limitations are considered.

An AI system is therefore not an independent challenger merely because we call it one. It can formulate an opposing position. Whether that position is relevant, well founded and actually new to the company has to be established.

Enriques and Zetzsche's work comes from an earlier generation of technology. It tells us little about which tasks a current agent can perform. But it describes a mechanism that greater computing power alone does not resolve: the distribution of influence over the system.

Why today's capabilities are not a fixed boundary

Limited evidence of impact so far does not mean that the possibilities will remain at their present level.

The research organisation METR measures progress on increasingly long digital tasks, mainly in software development, machine learning and cybersecurity. It measures the human time required to perform such tasks at a specified success level, not how long an agent can reliably run an organisation. The frequently cited 50 percent task-completion time horizon describes the task length at which an agent has a modelled 50 percent probability of success in the study. It is not a reliability standard for business decisions.

In 2025, METR expressed the conditional expectation that, if the trend continued, agents might be able to perform a broad range of week-long tasks around 2027 to 2029. This is neither a guarantee nor a forecast for replacing a supervisory board. Its results page has also not been actively updated since September 2026.

Market forecasts do not point in only one direction either. Gartner expects at least 15 percent of day-to-day work decisions to be made autonomously by 2028, but also more than 40 percent of agentic AI projects to be cancelled by the end of 2027 because of costs, unclear value or inadequate risk controls. These are analyst forecasts, not outcomes that have already occurred. The reference is to day-to-day work decisions overall, not executive board decisions.

For management and oversight, none of this provides a reliable date for replacement. It does provide a reason not to build one's way of working around an allegedly permanent limit to today's systems.

What this means for management teams and founders

A supervisory board monitors and advises from a different role than an executive board or management team. Founders, moreover, do not hold a particular corporate office merely by virtue of having founded a company. These differences cannot be dissolved by the word boardroom.

The functional connection nevertheless remains: Anyone leading a company also has to examine their own perspective, interpret developments and follow through on decisions. Management's self-scrutiny does not replace independent oversight. But it is important work in its own right.

This is where a question from board research can be carried across: Could AI make information and counter-perspectives accessible that receive too little attention in day-to-day management?

That is my editorial inference, not an effect demonstrated by the board sources for all management teams. Nor does it require another layer of control for every minor issue. The starting point is one specific decision for which a better information base could actually make a difference.

A limited trial rather than a virtual supervisory board

I would not begin by trying to recreate an entire role. I would choose a real decision and assess the additional AI support in that context. The following trial is my own working suggestion, not a scientifically validated standard.

1. Record the open question and the starting point. In the expansion example: Which assumption could make entry into this market unattractive even if the calculation presented is correct? Before using AI, the responsible team records its open questions and known risks. Otherwise, it is difficult afterwards to distinguish what was new from what was merely phrased differently.

2. Define a traceable information base. Which approved data, earlier decisions and external sources may be included? What is missing? Confidential documents belong only in environments that have been checked and approved for that purpose.

3. Make the additional contribution visible. The AI should provide source references, contradictions and open questions. The team compares these with its starting point: What is new, verifiable and relevant to the decision? Evidence of a previously overlooked supply dependency would be a point to investigate. A long list of generic expansion risks would initially be nothing more than additional material.

4. Do not reduce human judgement to clicking an approval button. Who has the expertise to check the results? Which conversations are needed? Whether a partner will remain reliable under pressure may not be answerable from documents. Experience, relationships and direct clarification are needed here.

5. Look back after the decision. Which new question was asked? Which assumption was corrected? How much additional checking was required? Better preparation may become visible before business success can be assessed. But it does not yet demonstrate a better decision, let alone a causal effect on company performance.

The strongest objection is that, if we have to redo every AI analysis in full, we lose the benefit. That is precisely why the trial should be limited: one decision, an agreed amount of checking and a person with the relevant expertise who is responsible. If traceable evidence is missing, or no additional insight emerges within that scope, the trial is stopped or redesigned. A system does not have to be used everywhere. If the data is incomplete, or checking quality costs more than the additional insight is worth, the right decision may also be to handle that task differently for now.

The role will not automatically stay the same

Perhaps fewer boards will disappear over the next few years than some people expect. Perhaps their work will change more than others currently assume. There are arguments for both, but no established roadmap.

The interesting shift starts earlier: Which contributions to management and oversight become more readily available? What new review becomes possible? And which human contribution consequently becomes more important — or genuinely less necessary?

For me, the discussion therefore begins not with the number of seats in the boardroom, but with the quality of the work that is meant to be done there.

Taking that perspective seriously requires neither declaring today's role irreplaceable nor predicting its abolition.