Agentic AI and the CFO’s Role in the Productivity Shift
n the second article for CFOs, Salvatore Cantale explores how finance leaders can help redefine output, workforce design, and performance measurement in an AI-enabled enterprise. ...
Artificial Intelligence
Artificial confidence: How LLMs undermine executive thinking
Professor Amar Bhidé advises executives dabbling in LLMs to stop.
Not unexpectedly, Nvidia CEO Jensen Huang is an AI super fan. “I use it (AI) every day, and it’s improved my learning, my thinking,” he said during a recent visit to the UK. “It’s helped me access information and knowledge a lot more efficiently. It helps me write, helps me think, it helps me formulate ideas.”
He’s not alone. An IBM study released in May 2025 found that 61% of CEOs are adopting AI agents.
Amar Bhidé, Professor of Health Policy and Management at Columbia University Irving Medical Center, is skeptical. “CEOs fib about their use of AI because they want to be perceived as forward-thinking or are trying to sell AI products,” he says. “Anyone who has tried to use LLMs to do anything useful would stop fairly quickly. If I knew of a CEO who was using AI in a meaningful way, I’d be tempted to short their stock.”
Bhidé is no technophobe. He obtained his first degree in chemical engineering from the Indian Institute of Technology, Mumbai, purchased his first PC in 1980, and coded his personal website. “I try every new technology,” he says. “I’m a sucker for it.”
But a series of underwhelming personal and professional experiences undermined Bhidé’s faith in artificial intelligence. For example, one of his students submitted an AI-assisted assignment that falsely claimed that Russian-American Nobel Prize-winning economist Simon Kuznets had helped make the discovery that ulcers could be treated with antibiotics. More generally, he found that AI-generated summaries of his academic papers completely missed the main point of the thesis. He’s well aware that large language model (LLM)-generated code can be riddled with bugs or simply fail to work.
“I see two issues with LLMs,” he says. “One is that many of them are rubbish. But even the ones that work well are not cost-effective. The current AI mania means investors and CEOs are subsidizing investments in AI.”
A stubborn cost curve
Bhidé disagrees with predictions that the cost of LLMs will plunge as the technology matures. Such a reduction would require a breakthrough, for example, Tesla’s discoveries around alternating current, which drastically reduced the cost of electricity transmission. Many assume that AI will inevitably experience something similar. The case of nuclear fusion shows that it is not a given. If there is no dramatic cost breakthrough, many promising AI use cases will be unviable because they cannot be marketed at an affordable price.
Bhidé believes that, ultimately, the cost and adoption trajectory of AI could be similar to that of robots. “We’ve been told for decades that robots will take over everything,” he says. “They didn’t. There are valuable use cases like managing warehouses. Like this, there could be useful applications for LLMs without dramatic cost breakthroughs, in restricted domains.”
LLMs seem to have been designed to flatter the user, thereby encouraging their continued use.
Flattery gets LLMs everywhere
Meanwhile, psychological manipulation can get users dangerously hooked. “LLMs are programmed to project authority while preying on the insecurities of users,” says Bhidé. “Their responses are at once utterly confident and aim to affirm users’ opinions or desires.”
If a CEO asked an LLM a series of questions about business strategy, the replies would more than likely align with the CEO’s existing views, assuring them that their thinking is sound and that they are on the right course.
“They play courtier, not devil’s advocate,” continues Bhidé. “LLMs seem to have been designed to flatter the user, thereby encouraging their continued use. ’Oh, what a brilliant idea!’ they will effectively say, convincing the user that they are really smart.”
“Continuously changing circumstances, from regulation and geopolitics to technological advancements and customer preferences, require executives to make decisions in a state of uncertainty.”
AI loves risk, hates uncertainty
To grasp the genuine potential of both LLMs and a broader set of AI technologies, we must understand the difference between statistical risk and contextual uncertainty. This distinction is central to Bhidé’s most recent book, Uncertainty and Enterprise: Venturing Beyond the Known. In the book, he points to the self-evident truth that statistical risk is quantifiable. Based on the assumption that the future will be similar to the past, risk can then be assessed using numerical, probabilistic measures.
A player sitting at the roulette wheel plays a game of risk, deciding how much money to bet on a certain probability of winning. A doctor explaining the life expectancy of a newly diagnosed lung cancer patient turns to data about similar cancer patients.
Uncertainty arises in events that are to a meaningful degree without precedent. It cannot, therefore, be accurately measured because past events no longer inform future possibilities.
“Risk is predicated on stationary phenomena, like a game of chess with fixed rules or the motion of planets, where what is going to happen tomorrow is exactly or very like what happened in the past,” says Bhidé. “In these cases, AI could, if programmed correctly, help determine the most likely outcome and the best choice to make. But outside of chess and the natural world, we invariably face uncertainty. In human affairs, we should expect the future to diverge from the past and therefore be skeptical about statistical extrapolation.”
This is especially true in business. Continuously changing circumstances, from regulation and geopolitics to technological advancements and customer preferences, require executives to make decisions in a state of uncertainty, rather than based on defined risk. They have to imagine what will happen – and imaginatively persuade others that their imagined future is desirable and attainable.
Faced with uncertainty, neither old-fashioned statistics nor oracular LLM offer much assistance. For executives, the challenge is to distinguish between what AI can reliably compute and what the human imagination can create.
Expert
Related
Will AI be a catalyst for net-zero transition … or a drag on it?
AI is power hungry but, used wisely, its benefits could outweigh the costs. Business leaders will play a key role in determining whether innovation and sustainability can advance together....
Audio available
Why cyber is no longer just a technology concern
From psychological manipulation tactics to social engineering and weaponized wearable technology, 2025 has seen it all. ...
Making agentic AI work: What CFOs need to know
CFOs can turn agentic AI from experiment to enterprise value by mastering execution, autonomy, and goal-driven workflows....
Learn Brain Circuits
Join us for daily exercises focusing on issues from team building to developing an actionable sustainability plan to personal development. Go on - they only take five minutes.
Explore Leadership
What makes a great leader? Do you need charisma? How do you inspire your team? Our experts offer actionable insights through first-person narratives, behind-the-scenes interviews and The Help Desk.
Join Membership
Log in here to join in the conversation with the I by IMD community. Your subscription grants you access to the quarterly magazine plus daily articles, videos, podcasts and learning exercises.