When AI Starts Acting, Who Is Still Doing the Thinking?

AI is moving from answering questions to taking action. At the same time, organisations may need to pay much closer attention to what happens to human judgement when more of the thinking is delegated to machines.

Several developments this week point towards the same emerging tension.

AI systems are acquiring greater agency. Research into cognitive offloading is raising questions about what happens when people increasingly delegate parts of their reasoning to AI. And, perhaps tellingly, senior executive development appears to be moving in the opposite direction, towards philosophy, ethics, intellectual challenge and deeper thinking.

Taken separately, these look like different stories.

Together, they raise a much bigger leadership question.

As artificial intelligence becomes more capable of acting, how do organisations ensure that human judgement does not quietly become the thing they automate away?

"Intelligence is not Judgement"

What’s Happening

Reuters reports that OpenAI has introduced autonomous agents designed to pursue objectives across applications rather than simply respond to individual prompts.

The significance is not simply improved automation. It represents another movement from AI as a source of information towards AI as an actor within organisational systems.

That creates a different governance problem.

An AI that drafts an answer can be checked before anything happens. An AI capable of initiating actions, changing information or progressing work requires organisations to think much more carefully about permissions, visibility, intervention and accountability.

A second development provides an interesting comparison.

OpenAI and Synopsys have announced work on specialised AI for semiconductor design. The AI may produce increasingly sophisticated engineering outputs, but those designs will still proceed through established sign-off verification before being accepted as physically valid.

There is a valuable distinction here:

Generation is not validation.

Validation is not authorisation.

Authorisation is not accountability.

The system may generate the work. Another mechanism establishes whether that work is valid. Someone then determines whether it should be accepted or acted upon. Accountability ultimately remains somewhere within the human organisation.

As AI becomes capable of doing more than making recommendations, knowing exactly where those boundaries sit becomes increasingly important.

"Recent workplace reporting has also examined whether growing reliance on AI may affect managerial thinking and cognitive independence."

What’s Being Said

Much of the current AI governance conversation understandably concentrates on safety, data, permissions and control.

But another strand of research is examining what happens to the human side of the relationship.

Cognitive offloading is not new. People have always used tools to reduce mental effort. Diaries remember appointments. Calculators perform arithmetic. Search engines retrieve information.

Generative AI changes the nature of that relationship because it can increasingly perform elements of analysis, interpretation and reasoning itself.

A recent study published in Frontiers in Psychology examined different forms of AI-assisted cognitive offloading among 589 participants. It found different relationships depending upon whether people retained cognitive agency while using AI or became more dependent upon it.

The authors rightly caution against claiming straightforward causation, but the distinction is important.

Using AI and delegating judgement to AI are not necessarily the same behaviour.

That matters because workplace use is becoming routine. Recent workplace reporting has also examined whether growing reliance on AI may affect managerial thinking and cognitive independence.

Managers can now ask AI to interpret information, generate alternatives, construct an argument, challenge a proposal, draft the communication and recommend what to do next.

Each individual step can look entirely reasonable.

Taken together, however, they raise a difficult question about where the human contribution begins and ends.

"A human can remain formally within a decision process while contributing remarkably little independent judgement."

What I’ve Noticed

We frequently talk about keeping a human in the loop.

That may no longer be sufficient.

A human can remain formally within a decision process while contributing remarkably little independent judgement.

Imagine an executive receives an AI-generated analysis.

The system identifies the problem, interprets the available evidence, generates the alternatives, assesses the risks and recommends a course of action.

The executive reads it.

It appears persuasive.

They approve it.

There was a human in the loop.

But was there independent human judgement in the loop?

The distinction becomes more important as AI becomes more capable.

If people repeatedly delegate the uncomfortable cognitive work involved in forming a judgement, something may gradually change. The immediate outputs may actually improve. Decisions may become faster. Analysis may become broader. Communications may become clearer.

Nothing necessarily fails.

That is precisely why the issue deserves attention.

Capabilities generally develop through use. Judgement is unlikely to be an exception.

An executive who repeatedly encounters ambiguity, weighs competing evidence, tests assumptions and makes consequential decisions develops something different from an executive who primarily evaluates recommendations that have already been assembled for them.

The important question therefore becomes:

Does AI challenge human judgement, or gradually replace the need to exercise it?

Those are profoundly different uses of the same technology.

"Capabilities are develop through use. Judgement is no exception."

What This Means

Another development this week makes the timing particularly interesting.

The Financial Times has reported growing demand among senior executives for development experiences drawing on philosophy, ethics, history, neuroscience and wider intellectual challenge rather than relying solely on conventional management competencies.

There is even experimentation with AI used as a Socratic coach, where its purpose is not to supply the answer but to make the human think more deeply about the question.

That shift makes sense.

For much of management history, information and expertise were relatively scarce. Leadership development therefore concentrated heavily on transferring knowledge, frameworks and techniques.

AI changes that equation.

Executives already have access to extraordinary amounts of information. Increasingly they also have access to systems capable of synthesising it, analysing it and recommending what they should do.

The scarce commodity may therefore move elsewhere.

• Towards the ability to examine those recommendations independently.
• To recognise assumptions.
• To tolerate contradiction.
• To seek disconfirming evidence.
• To distinguish confidence from evidence.
• To hear dissent without immediately defending an existing position.
• And ultimately to decide while retaining responsibility for having made the decision.

That creates an interesting paradox.

As artificial intelligence becomes more capable, human judgement may become more valuable precisely because intelligence itself becomes abundant.

Executive development may consequently move away from simply providing leaders with more answers.

Its greater value may lie in creating structured intellectual friction.

• Someone introduces another interpretation.
• Someone questions the assumption that has quietly become fact.
• Someone asks what evidence would cause the executive to change their mind.
• Someone prevents an uncomfortable contradiction from being rushed past simply because a plausible answer is already available.

That is also where independent executive challenge becomes increasingly important.

AI can supply intelligence at a scale no executive coach could realistically attempt to match so there is little value in competing with it there.

The more valuable role may be creating the conditions in which executives continue to exercise their own judgement around that intelligence.

"Value lies in creating the conditions in which executives continue to exercise their own judgement around available intelligence."

Pressure Test

Take one consequential decision in your organisation where AI now contributes to the analysis or recommendation.

Before seeing what the AI suggests, could the accountable executive state what they currently believe, why they believe it and how confident they are?

Then introduce the AI analysis.

If their position changes, could they identify the evidence or argument that caused the change?

Could they challenge the recommendation?

Could they explain what would make them stop the system acting?

And if the AI were wrong, is it completely clear who would remain accountable?

Hesitation around those questions does not necessarily indicate poor AI governance. It may reveal something more useful: the organisation has not yet separated intelligence, judgement, authorisation and accountability clearly enough.

The next phase of AI will make that distinction increasingly difficult to avoid.

The question is no longer simply whether AI can see more, know more or do more.

It is whether the people accountable for its actions can still independently judge them.

Can you see it?

Can you judge it?

Can you stop it?

And increasingly:

Who remains accountable when you don't?

AI can supply intelligence at a scale no executive coach ever could.

That makes intelligence a poor place for an executive coach to compete.

The more valuable role may be creating the independent friction that prevents intelligence becoming a substitute for judgement.

That is how I increasingly think about The Boardroom Coach: independent space for consequential judgement.