AI May Prepare the Boardroom. It Must Not Become It.

In a previous essay, I argued that the real AI employment risk may not be the one everyone is watching. The obvious panic is that AI will take the jobs. The quieter danger is that it may remove the work through which people learn judgement.

Now the same problem is moving up the organisation.

Boston Consulting Group has published an argument for “decision agents” in the boardroom: AI systems that help executive committees navigate complex, high-stakes choices by gathering inputs, testing scenarios, synthesising evidence, formulating recommendations and translating them into follow-up actions.

The pitch is seductive. The modern executive committee is overloaded. Capital allocation, market entry, supply chains, pricing, capacity, risk, regulation and technology choices increasingly cut across functions. The information is fragmented. The meetings are too slow. The politics are real. The data is often inconsistent. The consequences of getting the decision wrong are large.

So why not bring in a decision agent?

Why not have an AI system arrive with the briefing already prepared, the scenarios already tested, the trade-offs already mapped, the assumptions already challenged, and the recommendation already formed?

BCG describes the possibility of a custom-built decision agent as “another team member in the boardroom” — an omniscient chief of staff with supercomputing power.

That phrase should make every serious board pause.

Not because BCG has identified a foolish use case. It has not. Used properly, this kind of technology could be valuable. Most boards and executive committees would benefit from better evidence, faster scenario testing, clearer assumptions and more disciplined follow-up. There is nothing noble about badly prepared meetings, unread board packs, siloed inputs or decisions made on partial facts.

The problem is not that AI enters the room.

The problem is the role it is being invited to play.

There is a hard boundary between decision support and judgement substitution. On one side of that boundary, AI helps leaders see the decision more clearly. On the other, it begins to frame the decision, narrow the choices, rank the options, recommend the outcome and create the appearance that judgement has been exercised.

That is the line boards must not cross casually.

An AI decision agent can gather evidence. It cannot carry responsibility.

It can test a scenario. It cannot know what kind of organisation you are trying to become.

It can produce a recommendation. It cannot stand behind that recommendation when employees, investors, customers or regulators ask why the decision was made.

It can simulate a point of view. It cannot possess duty, courage, memory, reputation or accountability.

Those are not technical details. They are the substance of governance.

The most dangerous metaphor in the whole discussion is that of the AI agent as a “team member.” A team member owes duties. A team member can dissent. A team member has experience, incentives, fears, loyalties and consequences. A team member can be questioned not only about the answer but about their conviction, their hesitation and their responsibility for what follows.

A decision agent has none of that.

It may be useful. It may be powerful. It may be more comprehensive than any human analyst. But it is not a colleague. It is not a director. It is not an executive. It is an instrument.

And instruments must remain under the control of accountable people.

The risk is not science fiction. It is organisational psychology.

When an AI system produces a confident, polished recommendation, people respond differently. The recommendation looks neutral. It appears to have absorbed more evidence than any one person could process. It speaks in the language of reason. It turns ambiguity into structure. It gives the room something to converge around.

That is useful when the recommendation is treated as a challenge.

It is dangerous when it becomes the centre of gravity.

Anyone who has spent time around senior decision-making knows that the official decision process is only part of the story. Boards and executive committees are human systems. They contain hierarchy, fatigue, deference, politics, fear, impatience, overconfidence and career risk. People do not merely evaluate evidence. They read the room. They sense what the chief executive wants. They avoid being the difficult voice too often. They look for defensible positions.

Now put a highly capable decision agent into that room and give it the role of synthesising the evidence and formulating the recommendation.

What happens?

The first danger is judgement laundering.

Management can say, explicitly or implicitly, that the recommendation emerged from the agent’s analysis. The machine becomes a buffer between the human decision-maker and the human responsibility. No one has fully abdicated. Everyone remains “in the loop.” Yet accountability has become slightly more diffuse.

The second danger is synthetic objectivity.

Because the agent integrates data, the output feels neutral. But the data is not neutral. The assumptions are not neutral. The weighting of risk is not neutral. The framing of the question is not neutral. The choice of scenarios is not neutral. The definition of success is not neutral.

All of those things are human choices, even when they disappear inside a machine-generated recommendation.

The third danger is automation bias at the top.

Automation bias is not only a junior-user problem. Senior people are not immune to the persuasive force of a confident system, especially when the subject is complex, the time is short and the recommendation is convenient. A board does not have to be naïve to defer too much. It only has to be tired, overloaded and reassured by apparent analytical completeness.

The fourth danger is the loss of dissent.

The best boardroom contributions are often not synthetic. They are awkward. Someone notices that the numbers are right but the conclusion is wrong. Someone challenges the premise. Someone asks why the decision is being rushed. Someone has seen the pattern before. Someone says: “This may be technically correct, but commercially stupid.” Someone spots the political, human or reputational consequence that is not in the model.

That kind of dissent is not noise. It is governance.

A decision agent may be able to generate objections. It may even be instructed to play devil’s advocate. But a generated objection is not the same as a human being putting their judgement, status and responsibility behind a challenge in the room.

The fifth danger is that the board confuses a better process with a better decision.

A well-designed AI agent can make the decision process faster, richer and more consistent. That is not the same as making it wiser. Many catastrophic decisions were internally coherent. Many failed strategies had excellent models. Many acquisitions that destroyed value were supported by sophisticated analysis. Many boards that went wrong did not lack information. They lacked the judgement to interpret it.

That is the point.

The boardroom is not merely a place where information is processed. It is where accountability is supposed to concentrate.

That is why the language matters. If AI is described as an analyst, a simulator, a challenge engine or an evidence instrument, the hierarchy remains clear. The machine supports the accountable humans. But when AI is described as another team member in the boardroom, the hierarchy begins to blur.

And once that blur is accepted, the next steps become easier.

Today the agent prepares the memo.

Tomorrow it updates the recommendation during the discussion.

Then it proposes the follow-up actions.

Then it monitors execution.

Then it flags underperformance and recommends corrective moves.

At every stage, the human committee remains formally in charge. But in practice the decision architecture has shifted. The machine has moved from informing judgement to shaping it.

This is not an argument for keeping AI out of the boardroom. That would be both unrealistic and unwise. The question is not whether senior leaders should use AI. They will. The question is whether they will govern it before it governs the meeting.

There is a simple rule:

AI may prepare the boardroom. It must not become the boardroom.

That means boards and executive committees need to define the boundary before adoption, not after.

First, the agent should never own the question.

The framing of a strategic decision is itself an act of judgement. Are we asking whether to enter a market, or whether the organisation has the right to win there? Are we asking how to reduce cost, or what capability must be preserved? Are we asking which investment has the highest return, or which investment best protects the company’s future options?

The question determines the answer space. The board must own the question.

Second, the agent should expose assumptions, not hide them.

Every recommendation should show the assumptions on which it rests: market growth, execution capacity, capital cost, regulatory risk, customer behaviour, competitive response, timing, downside case and organisational strain. The board should be able to change those assumptions and see how the recommendation moves.

If the recommendation cannot be interrogated, it should not be trusted.

Third, the agent should produce options, not a single preferred answer too early.

One of the most subtle dangers in executive decision-making is premature convergence. Once a recommended answer appears, the room often starts debating that answer rather than reopening the field. A useful decision agent should widen the aperture before it narrows it. It should show credible alternatives, minority cases, downside paths and reasons not to proceed.

The machine should make disagreement easier, not harder.

Fourth, every AI-supported recommendation should have a named human owner.

Not a system owner. Not a data owner. Not a vendor. A human executive who says: “I have reviewed this. I understand the assumptions. I accept responsibility for the recommendation I am putting before this committee.”

Without that, AI becomes a convenient place for accountability to dissolve.

Fifth, the board should require a human first view.

Before the agent’s recommendation is shown, the accountable executives should state their own view. What do they think? What are they worried about? What trade-off do they believe matters most? Where is their uncertainty?

Only then should the AI recommendation enter the discussion.

This is the same discipline required lower down the organisation. If junior people use AI before they have formed their own view, they do not develop judgement. If senior people do the same, they risk outsourcing judgement at the very point where it should be most mature.

Human first. Machine second. Human reconciliation third. Human accountability always.

Sixth, the board should audit not only the output, but the influence.

The most important question is not merely whether the agent’s recommendation was accurate. It is how the agent changed the decision process. Did it narrow the debate? Did it suppress dissent? Did it privilege what was measurable over what mattered? Did it make directors more confident than the evidence justified? Did it shift responsibility away from named executives?

AI governance cannot stop at technical validation. It must include behavioural governance.

That is the missing layer in much of the current enthusiasm for agentic AI. We talk about model performance, data quality, security, explainability and workflow integration. Those matter. But boardroom AI creates a deeper question: how does the presence of synthetic judgement alter the behaviour of accountable humans?

That is not an IT question.

It is a governance question.

It is also a leadership question. Because the uncomfortable truth is that some executives will welcome decision agents for the wrong reason. Not because they clarify accountability, but because they soften it. Not because they improve judgement, but because they provide a defensible recommendation around which people can gather. Not because they force harder thinking, but because they make the meeting feel more rational than it is.

A board should be alert to that temptation.

The right use of a decision agent is to make the human decision harder before it becomes easier. It should surface the awkward facts, the hidden dependencies, the weak assumptions, the unfunded capabilities, the second-order risks, the uncomfortable alternatives and the reasons the preferred answer may be wrong.

If it merely makes the preferred answer faster, smoother and more persuasive, it has become part of the problem.

The irony is that the organisations most eager to put decision agents into the boardroom may be the same organisations already weakening their own judgement pipelines. At the bottom, they automate the junior work through which future leaders once learned. At the top, they introduce systems that simulate the judgement those future leaders may no longer possess.

That is the dangerous loop.

First, remove the apprenticeship.

Then, compensate for the missing judgement with machines.

Then, mistake the machine’s fluency for the judgement you failed to grow.

No board should accept that bargain.

The future organisation will need AI. It will need faster evidence, better simulation, richer scenarios and more disciplined execution. But the more powerful these systems become, the more important human judgement becomes — not less.

A weak board with a powerful decision agent is not a strong board. It is a weak board with a more persuasive briefing pack.

A timid executive committee with synthetic recommendations is not a courageous executive committee. It is a timid committee with better cover.

An organisation that cannot form judgement in its people will not become wise because an agent joins the meeting.

That is why the boardroom test is simple.

Does the AI make accountable humans think harder, or does it make it easier for them not to think?

Does it expose judgement, or does it simulate it?

Does it sharpen responsibility, or does it blur it?

Does it prepare the room, or does it become the room?

BCG is right that decision agents are coming to senior decision-making. The use case is real. The productivity case is obvious. The temptation will be enormous.

But the governance principle must be equally clear.

A decision agent can inform judgement. It cannot possess it.

It can support accountability. It cannot carry it.

It can prepare the boardroom. It must not become it.

In the first AI apocalypse, we worried that machines would take the jobs. In the second, quieter one, we may invite them to occupy the place where human responsibility used to sit.

That would not be the machine’s failure.

It would be ours.

*