Should CEOs Lead More Like Large Language Models?

Previous Chief Executive of the Year winners like Lisa Su, Marc Benioff, and Kenneth Frazier have not yet been joined by Claude or Grok, but perhaps large language models (LLMs) still have something to teach business leaders.

The statistics suggest that CEOs need all the help they can get. A record-breaking 2,221 American chief executives left their posts in 2024, exactly a thousand more than in 2015. Figures for 2025 dipped 9% but remained the second highest ever.¹

Prefer to listen to this article? You can do so here:

Why is there so much more chief executive churn than just a decade ago? The 2026 AlixPartners Disruption Index offers some clues:

Some 70% of CEOs report high organisational disruption, and 72% say it is increasingly difficult to know which threats to prioritise. More than half doubt their own team has the agility to respond, and 45% already fear for their jobs.

In this context, the old paradigm of analyse, decide, communicate, and execute is obsolete, and any CEO applying linear thinking to this new reality is a prime candidate for an early exit.

LLMs are built to operate amid exactly this kind of ambiguity, so how do they do it, and can business leaders usefully emulate them?

How LLMs do what they do

In a roughly chronological sense, LLMs are initially trained on very large datasets, with an effective cutoff in what they have seen. This training establishes their baseline parameters. On topics poorly represented in that training, however, they are more prone to hallucination – producing plausible but incorrect outputs.

Grounding can alleviate this by supplying the model with additional relevant, timely, and authoritative information at the point of use, often dramatically reducing error rates.

Once operational, an LLM uses its attention heads – each rather like an eye that has learned to look for particular relationships in the data – to determine what matters most.

That prioritisation can be radical: some information may receive very high attention, while other data receives almost none, although the many attention heads do this differently and simultaneously.

Even after this prioritisation, an LLM does not simply arrive at one absolute answer. At each stage of generating its response, it calculates probabilities for many possible next tokens and selects amongst them.

These choices can be influenced by a temperature setting. A low temperature concentrates probability on the model's most likely choices, making output more predictable and conservative; a higher temperature gives less likely alternatives a greater chance of being selected, making output more varied and creative – but potentially also more erratic.

Finally, through this repeated process of attention, probability, and selection, an LLM can produce a remarkably coherent narrative – and, when connected to tools, can use that output to take action.

What if a CEO followed the same playbook?

Like LLMs, CEOs are trained on finite, time-bound “data” that rarely prepares them for today’s challenges (e.g. AI, cybersecurity, and geopolitics).

That means they also need grounding in up-to-date information to inform key decisions. Our Disruption Index bears this out: CEOs are roughly twice as likely as the rest of the C-suite to feel they are personally falling behind in knowledge and skills.

The worthwhile challenge, amid the swirling hype cloud, is how to get down to the raw signals through trusted advisors, quality research, and confidential peer forums.

There is also merit in an LLM-like discipline of attention management through prioritisation, delegation, and active disengagement from the unnecessary. Otherwise, CEOs eventually dissolve into the noise.

Now the most radical idea: a strategic move towards probabilistic reasoning. Classical approaches have tended to set specific goals with singular, often immutable, plans to achieve them.

However, amid compounding disruptive forces, CEOs need strategic goals with multiple paths to fulfilment. Where the objective was once the pursuit of the perfectly optimised business, now optionality and adaptability are worth trading some efficiency for.

In practice, this might mean multiple supply chain routes, more than one prototype ready for finalisation and launch, or proficiency with frontier and open-weight AI models.

CEOs can even set a mindset temperature: dialling up creativity for “out-of-the-box” strategising or setting a narrow range to serve optimisation goals.

Finally, CEOs must give boards, shareholders, and staff the same strategic narrative fluency as LLMs. If these stakeholders understand where the business is heading, they are far more likely to align, even in a confusing environment.

Leadership is about the who, as well as the how

I think the case for LLM inspiration is strong.

However, Kouzes and Posner, who spent more than thirty years asking people worldwide what they most want in a leader, identified not only what I would call functional “how” characteristics, like “competence”, but also “who” qualities, such as “honesty” and “fairness”.

AI models can raise the bar for functional competence but not for character qualities, which are deeply rooted in distinctly human traits.

For now – and, I believe, for the years ahead – humans will therefore shoulder the true burden of leading businesses through disruption.

How can business leaders take action?

1. Get grounded. Do not rely solely on your experience, but gain access to the best information on the topics that matter. Seek quality over quantity.
2. Focus your attention. Many of the topics that cross your desk daily could be handled by other team members. Get very clear on what will determine success or failure and focus intensively on only those things.
3. Create options and manage them probabilistically. Singular hyper-optimised strategies should give way to multiple options that can be switched as the odds of their success change.
4. Communicate fluently. Provide a clear, cohesive narrative that inspires confidence, reduces anxiety, and enables everyone to align.
5. Be an outstanding human. People look for distinctive human qualities in their leaders. Dial up honesty and fairness because AI can’t.

So yes, CEOs should lead more like large language models, but with a big qualification: AI offers nothing on vital leadership character, but a lot on the “how” of navigating complexity.

That also means no AI winners of Chief Executive of the Year for the foreseeable future.

If you'd like to listen to an audio version of this article, please click here.

Footnote

^{1. Whilst total US CEO departures dipped in 2025, exits at publicly traded companies specifically set a record – 446, up from 373 in 2024.}

The content of this article is intended to provide a general guide to the subject matter. Specialist advice should be sought about your specific circumstances.