Not to decide in the directors' place, but to prepare them better, sharpen their capacity to reflect, and ask the right questions.

The boardroom is arguably one of the most complex levels in any organisation. That complexity is not incidental. It is the very nature of the exercise.

An analysis by PwC, featured on the Harvard Law School Forum on Corporate Governance, notes that high-performing boards tend to rest on three pillars: varied experience that informs judgement, a long-term view at the scale of the whole enterprise, and a one-step remove from day-to-day management that makes the difficult questions possible.

Those three pillars translate into three demands.

A board must first reconcile divergent points of view. Each director brings the wealth of their own experience, their own background, and a reading of the situation that is theirs alone. The value of a board arises precisely from the friction between these perspectives, provided that friction produces a decision rather than paralysis.

It must then understand the company's DNA. A board decision is never taken on figures alone. It engages a history, a culture, tacit commitments, and a trajectory that the documents never fully capture.

It must, finally, resolve complex situations. Incomplete information, high stakes, long consequences. A board often decides where there is no obvious right answer, only trade-offs.

the temptation

In this context, inviting artificial intelligence can be tempting.

The promise is real. A language model reads in seconds what takes a director hours to work through, and board packs often run to several hundred pages. Writing in the Harvard Business Review, Stanislav Shekshnia (INSEAD) describes pioneering boards already using LLMs to prepare for meetings and to test assumptions in session.

The same research identifies three contributions: preparing directors better, improving the information that reaches the board, and, in time, taking part in the discussion itself.

Beyond synthesis, the most interesting contribution lies elsewhere. A well-configured model helps surface the questions that would otherwise stay in the blind spots, challenge an assumption no one has dared to voice, and test management's position against alternative scenarios. PwC points the same way: AI can reduce the long-standing information asymmetry between board and management, and put strategy under pressure by combining public and internal data.

The state of practice, however, deserves a sober measure. A study published in the Harvard Business Review and cited by the Harvard Law School Forum on Corporate Governance reports that its two authors convened more than fifty directors in focus groups: most say they rarely, if ever, use AI in discharging their mandate. The promise runs well ahead of the practice.

the risks

That promise comes with real risks, which must be named plainly.

The first is confidentiality. A board handles what a company holds most sensitive.

The second is inaccuracy. The MIT Sloan Management Review puts it bluntly: an AI must be watched because it can drift or hallucinate. A model may perform well in testing, then degrade quietly over time.

To this is added the absence of critical judgement. Left to itself, an LLM is the perfect yes-man: fluent, assured, quick to confirm what is expected of it. The risk, for a board, is that it becomes the single voice behind which everyone falls into line, when good governance is measured precisely by each director's ability to challenge without deferring to it.

Other risks deserve to be named:

• Dependence. By leaning on a synthesis, a director may stop reading the source and lose the very material of their discernment.
• Premature anchoring. A well-written summary steers the debate before it has taken place, and manufactures a false consensus.
• Erosion of friction. Debate is a board's function, not its flaw. A tool built to smooth things over can impoverish what gives the board its value.
• Opacity. What models internalise, and how far they truly align with human intent, remain difficult to audit. Yet governance demands traceability, challenge, accountability, and the ability to explain why a decision was made. A fast, fluent, plausible answer is not enough.
• Accountability. A director carries fiduciary and legal responsibility. The MIT Sloan Management Review situates this within the Caremark and Marchand line of case law and draws a direct consequence: overseeing AI risk presupposes greater technological fluency on the board. That responsibility cannot be delegated to a machine.

addressing the risks

None of these risks is disqualifying. Each one can be managed.

On confidentiality, the news is a regular reminder that nothing is entirely secure. That holds for a web tool, an email system, a cloud, as it does for any digital infrastructure. Zero risk does not exist.

But let us ask the real question. How many directors and executives already use a public LLM in their work, and drop internal information into it without the slightest framework? The figures are telling.

According to the LayerX report of 2025, 45% of employees use generative AI tools, and among them 77% paste data into their queries, more than a fifth of which contains personal or payment information.

More striking still, roughly 82% of these copy-and-paste actions run through unmanaged personal accounts, outside any corporate security perimeter. A 2026 survey confirms the trend: 43% of employees have entered work correspondence into public tools, 34% customer data, and 31% financial information or confidential documents and strategies. The Samsung case, in which engineers disclosed confidential code through ChatGPT while seeking help with debugging, is only the visible illustration of a phenomenon that is both massive and silent.

The debate, then, is not "an LLM or no LLM in the board". It is "a framed LLM or the unmanaged use already under way". A tool hosted in a controlled environment, contractually bounded, and governed by clear rules, is more defensible than today's actual situation. The MIT Sloan Management Review describes exactly this logic under the heading of minimum viable governance: oversight must be built into the platform itself, which logs every prompt and every output, masks sensitive data, screens for hallucinations, and leaves an auditable trail. Control stops being a checkpoint upstream and becomes continuous monitoring. And this framing serves more than security: without serious configuration, an LLM also produces poor answers, when they are not simply wrong.

On inaccuracy, everything turns on the build. A board tool is not a raw chatbot, it is an architecture.

At Nobilys, it takes the form of board patterns, a structure of specialised competencies, and a virtual board designed not to reassure but to challenge. Grounding in the board's real documents rests on a technique known as RAG, for retrieval-augmented generation. Rather than answering from its general knowledge alone, the model first retrieves the relevant passages from the company's own documents, then builds its answer upon them. It stays held to the real substance of the file, and not to its approximations. The design intent is the reverse of the usual reflex: the aim is not a model that pleases, but a model that contradicts, that flags the blind spot and surfaces the missing question. One requirement remains, which the MIT Sloan Management Review captures in a single line: when a model does what it should not, someone must have the authority to stop it. That authority stays human.

a possible model

What, then, does such a model look like?

It prepares the board, digesting voluminous files so that meeting time can be given back to judgement. A director who receives two hundred pages the night before can have, the next morning, a structured synthesis, the points that call for vigilance, and the areas of risk, and walk into the session a step ahead.

It augments the board's capacity, through specialised profiles that extend the committees: audit, remuneration and nomination, risk, sustainability. A sector-intelligence module, specific to the client's business, fills a frequent blind spot. That blind spot is documented: close to three-quarters of boards rate themselves as moderately or barely AI-literate, and many carry a generational deficit on digital matters.

It challenges decisions and questions management, holding a line most experienced directors will recognise: a board decides, but above all it questions. The true value of such a tool is not to answer, it is to ask the right questions, whether of substance, of process, or of strategy. In practice, one can load a preparation pack and ask not for a summary, but for the five questions a rigorous director ought to raise in session on a given investment or acquisition. The value shifts from the answer to the question.

It accelerates, finally, the onboarding of new directors, giving rapid access to past decisions, their rationale, and the historical record, which shortens a learning curve that is usually long.

the fundamental objection

A demanding board will raise the objection itself, so it is best to name it. A consumer LLM already gives good surface answers, and an experienced board will often reach the same conclusions. That is true. But it confuses the answer with the value.

The value of such a tool is not to produce an unexpected conclusion. It lies elsewhere: in the speed of preparation, in aligning the board around a shared base of facts, in the accelerated onboarding of a new member, and in a grounding in the company's real documents that a generic tool will never provide.

A public chatbot answers an isolated question. A well-built virtual board holds the memory of past decisions, knows the file, and is designed to contradict rather than to please. The difference is not in the fluency of the answer, it is in the depth and traceability of what underpins it.

Not to replace, but to enhance

There remains the most intimate fear. A director may dread losing their added value. The opposite is what happens.

AI does not decide, and has no business deciding. It expands the capacity to reflect, sets a decision in perspective, and opens angles the room had not seen. Judgement, the weighing of trade-offs, and responsibility remain entirely human.

This contribution is not confined to mid-sized organisations without a structured board, where the gain in structure is obvious. It holds equally for mature, professional boards, which gain in effectiveness, in impact, and in alignment. A well-built tool tightens a board's alignment as much as it accelerates the onboarding of its new members.

The right way to invite AI into the boardroom, then, is not to entrust it with the decision. It is to make it an instrument of preparation and a partner in contradiction, in the service of directors who remain, and remain alone, accountable for their choices.

References



• Stanislav Shekshnia, « How Pioneering Boards Are Using AI », Harvard Business Review, juillet-août 2025.
• « How Forward-Thinking Boards Are Using AI », INSEAD Knowledge, janvier 2026.
• Paul DeNicola, Barbara Berlin, Ariel Smilowitz (PwC), « Using AI in the Boardroom : New Opportunities and Challenges », Harvard Law School Forum on Corporate Governance, novembre 2025.
• « How Boards Can Lead in a World Remade by AI », Harvard Law School Forum on Corporate Governance, février 2026.
• « Why Your Board Needs a Plan for AI Oversight », MIT Sloan Management Review.
• « A Framework for Assessing AI Risk », MIT Sloan, janvier 2026.
• Nick van der Meulen, Jennifer Jewer, « Minimum Viable Governance for Generative AI », MIT Sloan Management Review, juin 2026.
• « The Real Question to Ask About AI Governance », MIT Sloan Management Review, juin 2026.
• LayerX, Enterprise AI and SaaS Data Security Report 2025.
• « Shadow AI Is Happening Within Your Organization », enquête 2026.
• Lutz Finger, « AI in the Boardroom : What Directors Need to Know », Forbes, avril 2026.