Harvard Professor Karim Lakhani on Who Is Most at Risk From AI and How to Build Real Expertise

The video opens with Karim Lakhani's warning about the workers most at risk from AI. He says he has fired people who told him that the AI gave them their work, and that the danger lies with people who hand their thinking to the machine.

Lakhani is a Harvard Business School professor. He describes research with BCG that identified three ways people work with AI. Two of the groups developed their skills, and one did not. The host asks how a person can tell which group they belong to.

Lakhani answers that self-automators do the worst because they have stopped thinking and stopped engaging in the activities. People who stay engaged can do far more complex things and push their expertise much further than before.

The host then asks for practical tips on how to train AI assistants and what the prompts should be. The conversation turns to Lakhani's view that the near future is humans with AI rather than AI acting alone. He argues that AI cannot yet set its own hypotheses, and that it needs direction, curation and taste from people. He also says that execution tasks are being absorbed by machines while organizations and occupations are being reconfigured at the same time.

Lakhani refers to a study by Fabrizio Dell'Acqua at Michigan showing that people with good AI can fall asleep at the wheel, which he says is where self-automation can go wrong quickly. He describes cyborgs as people who intimately work out what AI is good at and what they are good at. In his observations of R&D, sales and marketing tasks, he says cyborgs still have a lot to learn and figure out.

Centaurs divide the work between themselves and the AI. Lakhani's worry is that they stay static in terms of what they expect the AI can do. Because AI capabilities are improving exponentially, he argues, people must keep updating the separation between themselves and the machine.

The host notes that the 'jagged frontier' makes this complicated: a task that looks as hard as one the AI handles well may suddenly fall outside its abilities. Lakhani says BCG hypotheses that AI would be good at everything did not hold. For consultants, some tasks are handled well and others are not. The challenge is that the AI's intelligence is not perfect but talks very well and can be persuasive, so users must apply skeptical thinking and high standards instead of falling asleep at the wheel.

Lakhani says that workflows and routines are changing. As AI does more of the execution, he says, the question becomes what humans are really adding value for. He sees this as a great place for humanity to be, because people should not be comfortable with the current equilibrium and could go solve bigger problems.

The host raises a finding from a persuasion study: when BCG consultants pushed back on GPT-4, it often argued even harder, even when it was wrong. Lakhani says there are now ways around this. He describes a prompt approach on which Kai-Fu Lee posted on LinkedIn, in which the AI starts with the counterargument, suggesting the user's perception of the problem may be wrong. For every claim it makes, it indicates how much is based on its own training and its own confidence, and where its arguments are weak.

Lakhani says users must train the model and system prompt to push back from day one, and also learn to push back against it. He calls this a back-and-forth that becomes critical, and warns that people who are naive about it easily fall into the trap of just believing what the AI says.

Lakhani says that when people ask AI for a design, it picks models that are the average answers rather than ones for the task at hand. He suggests using AI to educate and learn, while pushing it toward several different targets and being adversarial, asking it to prove itself wrong and list pros and cons.

Asked how a one-person team can tell whether AI's design is world-class or merely mediocre, Lakhani says the responsibility does not go away, because in the end the person puts their name on the work. He repeats that he has fired people who said 'the AI gave me that,' and asks whether the person is just a meat interface to AI, because then he does not need them.

He says people must stay engaged and that it is easy to fall into cutting and pasting. A one-person product team needs to build knowledge about design, learn it, and set up the AI so it can be prompted in different directions rather than returning a first design. In his words, the burden of learning and taste goes back to the human; otherwise machines will replace humans when humans stop exercising their judgment.

When the host asks how students can learn taste if basic tasks are already done by AI, Lakhani says he feels he is in the middle of a tsunami.

Lakhani describes a conversation with Stephen Wolfram in which Wolfram said that although he designed mathematical languages to do integrals decades ago, he saw his kids in school still learning integrals by hand. Wolfram's point, as Lakhani relates it, is that you need to learn to do integrals by hand initially so you understand how things work, but the next 50 or 100 integrals should be left to the machine.

Lakhani says the reason is that observing a thousand integrals done by a machine teaches you far more about integrals than doing the next 40, 60 or 80 by hand. He presents this as a thesis for handling automation: learn the foundational piece yourself first, then give the rest to the machine and become an observer of the underlying patterns.

He ties this to the role-expansion story. Anyone responsible for these tasks, he argues, needs to understand the fundamentals well enough to use the machines properly. He admits he is partly self-centered as an educator, but says the importance of learning keeps increasing because what one learned at school or university is no longer enough.

Lakhani says the main thing he teaches his students is learning how to learn. Students who have a routine for learning how to learn are not afraid of challenges and actually thrive on them. He says that with AI, learning how to learn is now quite easy, because there are different explanatory ways of teaching and quizzing that are available to anyone who knows how to use them.

On the question of whether students are learning more than they did two years ago, Lakhani describes his required MBA AI course, which is three years old and has had to pivot every year because of AI. When Claude-style tools appeared suddenly last year, the course had to teach them to keep its legitimacy. The team's key realization was that business school cannot only talk about AI, so half the class now builds with AI, creating RAG systems, running evaluations and building multi-agent systems to solve business problems.

Lakhani describes executive activation sessions with C-suites. Nearly every executive or chief of staff says they are already doing a lot of AI, which he says he does not believe. He wants the entire C-suite to vibe-code applications that matter for their own businesses rather than trivial ones, such as a sports-betting app. Asked about operations inside a major global firm, executives tend to resist and point to the IT department.

He says that once another CEO backs the trial, most of the leaders he has encountered, who had been cushioned from what the models can do, build something in about 90 minutes. Their reactions are twofold: they yell at the CIO for not delivering this for five years, and they realize that if they are not doing AI themselves, they cannot guide their teams to do it.

Lakhani says that if programming is the leading indicator of what will happen in AI, every other occupation is six months to two years behind.

The host notes that in Silicon Valley everyone is coding and building, while friends building AI companies say they hold conferences on very basic things that people find exciting. Lakhani responds, citing William Gibson's line that the future is already here but not evenly distributed. He says Silicon Valley and a few other pockets live in a very different future than most companies, and that most companies are probably at GPT-4.5 level, which he calls basically advanced search.

Lakhani adds that his research team, through close relationships with companies, has observed prompt-level data and AI usage across hundreds of thousands of people in large organizations. He says there is a steep curve inside organizations: the difference between the 95th or 99th percentile of users and the average is between 10x and 20x. As models improve, those who have figured out how to use them skyrocket, while the rest stay roughly where they are, like someone on a moving walkway while others ride an escalator. He says studies on this will be released soon.

Lakhani says targets should be acts of imagination, not AI targets, and should be business results such as revenue increases. People should then work backwards: if this were going to be true, what else would need to be true?

The host recalls a study Lakhani did for H&M in which, when AI created the solution, it was hard for people to sell internally, which suggests AI cannot set your ambition, because if it sets a strategy you cannot sell it to your team. Lakhani agrees completely, saying the leader must push first and AI can then help bolster it, not the other way around.

On teaching ambition, Lakhani says there is a normal distribution of ambition in the world, and the aim is to shift it toward more ambition. He says fear of failure and worry about what others think make this deeply personal. He argues that when leaders set outlandish targets that seem unachievable, most of the time they come true. Setting a 10x target and achieving 7x instead of 2x changes the world.

Lakhani says that to get into Harvard Business School, applicants need an impeccable CV and grades. If students have not encountered enough failure, they may judge that a 10x goal might end in failure and so set a 2x goal instead, which may be good enough.

He says the school must keep bolstering students, telling them they will always be fine because they have been to top schools, and asking them to have faith in themselves while setting incredible goals for themselves and the world. Pushing students in that direction is, he says, a constant task for the faculty.

The host says the idea of AI as a team of experts is empowering but worries about the claim that the marginal cost of expertise is going to zero. Lakhani says he is obsessed with this question and that he came to academia to understand open source software in the late 1990s, when it made no economic sense. Open source relies on multiple shots at goal, with many people trying many things, so that the best-performing solution can be made available and the right tail is reached.

Lakhani says AI is in many ways an absorption of human expertise, which amounts to an open-source model at scale. What AI is doing is shifting the human-plus-AI expertise curve to the right. His empirical studies at P&G, H&M and BCG, he says, support the view that people draw from a distribution of expertise that AI provides. Everyone now needs to understand what they have developed individually, what the AI has, and how they differ from it. 'It cannot just be you without AI. It has to be always you with AI.'

He compares this to CAD/CAM. He learned drafting by hand, and could only do so much until CAD/CAM allowed designers to do far more complex work. AI likewise lets experts push further, or the machine replaces them.

Lakhani says he finished his undergraduate degree in 1993, when the internet was appearing, and has since lived through waves of mobile and machine learning. He keeps asking himself whether this is a difference of degree or a difference of kind, and says he keeps landing on the view that it is a difference of kind.

If that is right, everyone must keep asking what their expertise is, how they are developing it, and how it changes with AI. He says this is experiential: it cannot simply be taught, and people have to keep working with the tools. He describes his wife, an artist, and his daughter, a designer, both struggling to work out how AI can take their work further.

He points to hip-hop as an example of technology changing an art form, from two turntables and a mixer to a global form of expression. He argues that technology is changing every profession and occupation in a similar way.

The host notes that a model may be reviewing code and another model may be stepping in to do the review. Lakhani asks what the human now does. In his open-source studies he saw that many people love open source because they love the grind of coding. He says he was a terrible coder and never got into it, but others do it. If that is taken away, he asks, where does meaning come from in work?

He suggests that the answer may be managing agents, which may or may not be meaningful. He says everyone across occupations is trying to figure out something new and does not yet have answers, which is why it feels so uncertain and why people feel unmoored.

On consumption, Lakhani says models will increasingly consume outputs, so creators may need to publish machine-readable files as well as content. He argues that human views still matter, but will increasingly be mediated through these tools, so creators need to make both available.

Lakhani responds to colleagues worried about junior workers by saying that, if they encountered real organizations and real workers with real data, they would be skeptical of the claim. The host says it depends on the junior worker. Lakhani says that if all a person is, is a copy-paste meat interface to the model, that job is gone, and completely gone.

He says he asks law firms and other CEOs the same question: if a law firm is a thin wrapper around an AI company's model, or a manufacturer is a thin wrapper around another model, why should it exist? He says the question to ask is how jobs should be redefined and what the nature of work now looks like. He suggests people could do ten times more if AI were deployed more widely at scale.

On 2030, he expects a tale of two cities: a set of firms that understand the change and redefine work, including startups and large enterprises, and others that do not. He cites a colleague's Frontier Firm Radar, which found that startups using AI better achieved faster ARR growth and faster product launches.

Lakhani relays a test from Greg Brockman, president of OpenAI: ask yourself whether your business is improving with a new frontier model or whether it is disappearing. Lakhani says this is how you figure out your moat and how it is affected by the frontier. He stresses that this thinking cannot be outsourced to someone outside the organization or to someone junior, and that you must do it yourself.

He says leaders must play with these models now. Recalling the arrival of email, he describes bosses who had assistants read their email and dictate replies, some of whom never used email themselves. He says the same thing is happening with AI leaders, who see it as somebody else's job.

He also reports that CEOs who understand AI are spending between 20 and 30 percent of their time discussing AI issues. At one global firm, he says, half of senior meetings were about the impact of AI on the business.