For several years, the leadership conversation about artificial intelligence has turned on one question: what can the machine do? AI systems can screen thousands of applications, flag underperforming units, adjust prices, approve transactions and reroute supply chains, at a speed and scale no executive team can match.

A second question is now becoming more important than the first. Not what the machine can do, but what we should allow it to decide. The distinction sounds academic. It is not: it is where a conversation about technology becomes one about leadership and, ultimately, about responsibility.

When AI can decide, who should decide?

We are not only automating work. We are redistributing authority.

I have spent much of my professional life running companies and watching leaders decide, across countries, cultures and organisations. Contexts change enormously; one thing does not. To lead is, in the end, to decide. Some decisions move millions. Others move something harder to measure: a career, a reputation, a relationship of trust, the dignity of the person on the other side of the table.

This is why I think the current debate contains a quiet misunderstanding. We speak of delegating tasks to AI. Increasingly, we are delegating authority, and the two are not the same. There is a real difference between a system that says “this is what I found”, one that says “this is what I recommend” and one that says “I have already acted”. At each step, the technology simply seems a little more useful. At each step, something else also happens: power moves.

The leadership question is whether we are choosing, deliberately, where it moves.

The silent delegation

Very few organisations will announce that, from Monday, AI holds more decision-making authority. It will happen one recommendation at a time, one automatic approval at a time, one exception that quietly becomes procedure.

The reasons are always sensible. The system is faster. The numbers look convincing. The executive has two hundred other things to resolve, and yesterday’s recommendation worked. Clicking “approve” is far easier than rebuilding the reasoning behind it. Over months, the exceptional automated decision becomes the normal one. The executive is still there, but authority has changed place. Nobody decided to hand it over. Nobody noticed it was moving.

The greatest leadership risk of this period is not that machines will suddenly seize power. It is that people will gradually stop exercising it.

Who is setting the price?

Hospitality offers a clear illustration, and it is a sector I know from the inside. For a long time, a hotel’s rate was a judgement made by its general manager and commercial team, informed by occupancy, the calendar, the competitor across the street and a good deal of instinct. Revenue management systems changed that, largely for the better. They read demand, booking pace, competitor rates, local events and historical behaviour far more thoroughly than any team could.

First, the system informs the manager. Then it recommends a rate. Then, because its recommendations are usually right and the team is stretched in high season, it is allowed to change prices on its own within a range. The range widens. Eventually a question appears that nobody asked at the beginning: who is actually setting the price?

A more important one follows. When a rate is wrong, because it damages a long relationship with a tour operator, alienates loyal guests, contradicts the positioning the owners have built or treats some customers unfairly, who answers for it? The revenue manager who no longer sets rates? The general manager who approved the configuration? The software vendor? Revenue management is not the problem, and this kind of automation is often an improvement. The risk appears when authority shifts without anyone having decided that it should.

The regulatory calendar has moved. The leadership question has not.

According to the OECD, 20.2% of firms in the countries with available data used AI in 2025, up from 8.7% in 2023; among large firms, the figure reached 52%. The conversation is moving from experimentation to integration, and integration is where authority starts to shift.

Europe, meanwhile, is applying its AI Act progressively. The Digital Omnibus on AI (Regulation (EU) 2026/1744), in force since July 2026, moved the rules for high-risk systems listed in Annex III to December 2027, and those for AI embedded in regulated products under Annex I to August 2028. The principle, however, is already written into the text: once those rules apply, organisations using high-risk systems will have to assign human oversight to people with the necessary competence, training and authority. The word authority is there for a reason.

Law can tell an organisation what it may not do. It cannot tell its leaders what they should do. Compliance is the floor; judgement begins above it.

From Human in the Loop to Human in Authority

For years, one phrase has reassured us: human in the loop, a person kept inside the process. I am no longer sure it is enough.

Picture an executive who receives two hundred AI-generated recommendations every morning, each built on more information than any person could process alone. Approve. Approve. Approve. For six months. Technically, a human is still in the loop. But is that executive deciding, or have they become the final click in a process the system effectively runs?

This is why I propose a more demanding principle. Human in Authority is the principle that human beings must retain meaningful authority over consequential AI-assisted decisions. The difference from human in the loop is not presence but power: a person can remain technically inside the process and still have lost, in practice, the real capacity to decide. Consequential decisions are those that materially affect people, strategy, capital, reputation or safety, or whose effects are hard to reverse. Meaningful authority is more than the right to press “approve”: it is the capacity to understand, question, refuse, intervene and, when necessary, stop the system.

The case for it is not that people are smarter than machines. Often we are not, and we are not free of bias either. It is that certain decisions carry something intelligence alone cannot carry: responsibility.

“The algorithm decided” is not an answer

An algorithm rejects a candidate. Who answers? A system declines an important client. Who explains why? A model concludes that the organisation would be more productive with three hundred fewer people. Who looks those people in the eye: the chief executive, the board, the vendor?

There is one sentence I hope never becomes normal in the vocabulary of leadership: “the algorithm decided”. The algorithm may have calculated, predicted, recommended, even executed. Responsibility cannot disappear inside it. A simple rule follows: never delegate a consequential decision until you can name the human being who will still answer for its consequences.

The Human Authority Test

Most AI strategies start with tools. I would start with a decision. Put one consequential decision your organisation makes today on the table (a hire, a dismissal, an investment, a price, a client, a supplier) and ask four questions before delegating it.

Capability — Can it?

What can the system genuinely and reliably do? Not what the vendor promises or the demonstration suggests, but what it does with our data, in our context, given these consequences. Capability has to be demonstrated, not assumed.

Authority — Should we allow it?

What are we prepared to let the system decide? Analysing, recommending, deciding and acting autonomously are four different levels of power, and drawing the lines between them is a leadership task.

Accountability — Who answers?

Who remains responsible for the consequences? Delegating execution cannot create a vacuum of responsibility. A person, a role or a body must answer for the outcome; ideally, a name.

Reversibility — Who can stop it?

Who has the information, authority and operational ability to intervene or reverse the action?

If nobody around the table can answer all four clearly, the organisation is probably not ready to delegate that decision. If it can, the answers belong in writing: what AI may analyse, recommend, decide and execute; who remains accountable; who can halt the system. A short human–AI decision charter of this kind is a leadership document, not a technical one, and it is best written before an incident reveals where the boundaries were.

Power without accountability is dangerous. Authority without reversibility becomes dependence.

When intelligence is abundant, judgement becomes scarce

For centuries, access to information was power; later, knowledge became the advantage. AI is making both abundant. Analysis, prediction and expertise are becoming cheaper. What becomes scarce is something much older: judgement. Knowing not only what can be done but what should be done. Knowing when the numbers are right and the decision is still wrong, where efficiency should stop, and when to have the composure to say no.

That is where competitive advantage is moving. More than half of large firms were already using AI. Saying that a company “uses AI” will soon say very little about it. The difference will lie in who knows how to govern its power, and the strongest organisations may be those that know precisely where automation should stop.

The question that remains

This is the conversation about AI I increasingly want to have with leaders. It does not start with a tool. It starts with a decision on the table and three questions: what would you delegate to a machine, what would you never delegate, and why? The answers reveal something about the technology, and far more about the leader.

AI may become increasingly capable of making consequential decisions. The leadership question is whether humans remain willing to carry responsibility for them.

Which consequential decision in your organisation would you never fully delegate to AI, even if the system could make it faster and more accurately?

Abderrahim Ouadrassi
Executive Coach · Strategic Thinking Partner for Leaders
Founder, AbdeCoach

Sources

  • OECD, AI use by individuals surges across the OECD as adoption by firms continues to expand, January 2026.
  • European Commission, AI Act Service Desk, Timeline for the implementation of the EU AI Act.
  • Regulation (EU) 2026/1744 (Digital Omnibus on AI), Official Journal of the European Union, 24 July 2026.
  • Regulation (EU) 2024/1689 (AI Act), Article 26(2).