Ethan Mollick on Entrepreneurship, AI Agents, and Preserving Human Judgment

Marina Mogilko interviews Wharton professor Ethan Mollick about entrepreneurship, AI and his book Co-Existence. She begins by asking how he encourages students to be ambitious when they start companies in class.

Mollick says there is no simple playbook for agency, but people often impose barriers on themselves. Asking what they could do beyond their first plan and placing them among ambitious peers can expand their view of what is possible.

In a recent class, surgeons, finance professionals, soldiers and humanitarian workers built tools around problems they knew. Examples included understanding supply-chain vulnerabilities and explaining medical information to patients more clearly.

He says entrepreneurship teachers repeatedly see ideas that underestimate social, political or legal barriers, such as replacing real-estate brokers. Students also propose solutions to problems specific to campus life. These ideas are not automatically bad, but can be harder than they appear.

Mogilko points to businesses that succeeded with ideas that once sounded unlikely. Mollick agrees that he cannot know in advance which will work, which is why he does not simply forbid students from trying.

On ambition, Mollick says a large goal can attract good people and motivate a team, but starting and scaling a company are different challenges. Running a small business and raising successive venture rounds require different skills.

He thinks many founders are too focused on selling software as a service. His historical analogy is to use a new technology to operate a business more effectively, rather than assume the best opportunity is always selling tools to others.

Mollick expects AI to lower barriers to entrepreneurship while changing the market for software. Some companies will still need strong engineers; for others, software becomes a toolkit for delivering a different product or service.

He says large companies are considering which tools they can build themselves, though existing vendors still provide value and the transition will take time. Products that merely repackage capabilities a customer already has may face particular pressure.

Mogilko describes considering a custom newsletter tool when subscription costs rise. Mollick sees a broader benefit: a nontechnical founder can use software and advice to pursue the actual business rather than first find a programmer for every need.

They discuss a study of entrepreneurs in Kenya using an older model for advice. Mollick reports a profit improvement among the stronger entrepreneurs in that study and says the advice did not similarly rescue those already struggling. The result is specific to that study and its setting.

Mollick names four capacities for effective AI use: deep knowledge in a domain, broad knowledge of adjacent fields, agency and taste. Expertise helps a person judge results; breadth makes it easier to apply the tools across tasks.

He argues that if everyone accepts the same model’s first strategy or prose, their work can converge. In experiments he describes, AI-generated ideas could become diverse but still did not cover the entire range produced by a group of people.

The implication, in his view, is to add a distinctive human perspective where it matters competitively. Using a generic tool for an unremarkable support task may be fine; the area where a business wants to stand out deserves more deliberate direction.

Taste can be developed by seeing good and bad examples, learning the vocabulary to discuss them and examining many alternatives. Mollick brings a design professor into class to help students articulate why a font or image works.

Mogilko asks how taste differs from judgment. Mollick says AI already makes judgments and creative choices; the question is whether those choices match the person’s goals and standards. He thinks people are still developing language to distinguish these forms of judgment.

He uses mathematics as an example: a correct proof may matter differently to researchers if it also reveals an interesting idea. AI forces people to articulate qualities that previously went unnamed because human judgment was the only reference point.

Their attention turns to education and junior work. Mollick thinks AI can be a useful tutor, but simply submitting its answers can bypass learning. He expects more active learning and assessment alongside tutoring.

He worries more about apprenticeships. Junior workers traditionally learned by doing tasks for experienced colleagues and receiving feedback. If both junior workers and managers rely on AI for those tasks, the old path to professional judgment may weaken.

Mollick calls the loss of practice cognitive surrender. He suggests deliberately preserving some effort in areas where a person needs to learn or maintain an advantage, even when delegating would be easier in the moment.

Organizations face a choice, he says: hire only experienced people, or invest in structured training that gives newcomers time to learn. He notes early signs of pressure on junior roles but says the evidence is not yet conclusive.

Mogilko asks what happens to competent but not exceptional professionals. Mollick says it is too early to know and returns to the jagged frontier: AI can be unexpectedly strong at one task and weak at another.

Jobs bundle many activities, including coordination with people, institutions and legal systems. As the bundle changes, different combinations of technical vision, communication and other skills may become valuable; he does not assume any particular task is permanently protected.

Mollick says expertise increasingly helps people use agents well, and that some people also develop an intuitive feel for how a model responds. He describes a negotiation-agent study in which the people best at designing agents were not necessarily the strongest human negotiators.

Mogilko asks whether the key is explaining one’s own mental workflow to AI. Mollick is unsure that this is universally true. His preferred analogy is managing a team: delegate work, watch where it goes wrong and provide feedback, even if the system does not imitate his own process.

He writes the first draft of his own work because model-generated prose can steer his thinking. For him the aim is to preserve a human process for the choices that matter, while delegating other work.

On brainstorming, Mollick recommends generating ideas individually before asking an AI or group for alternatives. Early suggestions can anchor later thinking; starting independently can preserve a wider range of ideas.

He calls the useful effort of thinking “desirable difficulty.” If a process is entirely effortless, a person may not be learning. The point is not to avoid assistance, but to decide which mental work is worth retaining.

People differ in creativity, and Mollick does not think every person must protect idea generation in the same way. Someone can use AI more heavily where ideas are not their distinctive strength and focus on execution, connection or another contribution.

Entrepreneurship methods such as customer interviews and minimum viable products developed around learning cheaply. AI can lower the cost of an early test further, although Mollick cautions that a simulated customer is not an actual customer.

He says ideas gain value through execution and exploration. In his classes, a startup’s initial idea often changes substantially as the team tests it; the person who first proposed it is not the only source of the eventual outcome.

Managing agents requires many familiar management skills. Mollick recommends specifying what good and bad outcomes look like, identifying checkpoints, requesting alternatives and inserting the human’s own choices into the process.

He watches early runs to catch communication failures before an agent pursues the wrong direction. His concern is not only token expense: monitoring a large amount of delegated work also costs human time and attention.

The biggest mistake, in his view, is trusting agents too much. He argues that interfaces emphasizing magical automation can make it harder to see what is happening, so managers need explicit points at which the system reports back.

He illustrates the approach with research into the shape of stories. He asks an agent to propose methods, provides feedback, has it try several approaches and then selects and combines promising elements.

The agent handles tedious work such as gathering thousands of stories and building tools; Mollick keeps the interesting decisions. He contrasts this augmentation model with handing over every choice and receiving a finished result.

When Mogilko asks how often a workflow must be rewritten as models change, Mollick says the example is not a fixed recipe. It is a management process that adapts to each interaction and preserves opportunities to choose.

He is happy to delegate fact-checking and routine forms. In checking his book, he says AI identified not only his own errors but possible errors in sources he cited. These are his reported experiences and still require verification.

Mogilko describes an assistant sending a form before she intended it to. Mollick recounts an unwanted email sent by an assistant to his dean. Both experiences illustrate why convenience does not replace clear boundaries and a process for building trust.

Mollick says he starts scheduled tasks at low stakes and adjusts instructions over time. He worries that some assistants expose too little about their configuration or proactivity, making it harder for users to predict behavior.

Rapid updates create organizational friction: a company can train hundreds of people on an interface only to find it changed. Mollick nevertheless thinks people and organizations often adapt more effectively than advance predictions suggest.

Mogilko describes a model finding a forgotten conversation in her records and using it to improve an interview workflow. Mollick says researchers face a comparable challenge when AI identifies overlooked connections or tensions in their work.

His advice is to acknowledge anxiety and frustration while keeping curiosity and a sense of play. New tools can expand what people do, even though changing interfaces and canceled products make adaptation tiring.

On local models, Mollick rejects a precise forecast about their future enterprise share as unsupported. He says organizations differ in their preferences and cautions against assuming today’s constraints will remain fixed.

He recommends keeping a few personal benchmarks that models cannot yet satisfy. For him these include an interesting original academic paper, a startup executed from an idea alone, a compelling game and a coherent story he wants to read.

Asked what happens if AI learns taste, Mollick says the question is partly about a person’s individual taste. He expects unevenness and similarity across systems to remain meaningful for some time, but admits he has no answer for a future in which those gaps disappear.