Wharton Entrepreneurship Professor: What I Push My Students to Build Now That They Have AI

Summary

Ethan Mollick argues that AI makes it easier to produce work and build products, but relying on it too readily can weaken skills and erase the human differences that make people or businesses distinctive. His central advice is to use AI deliberately: let it handle suitable work while preserving human input in the areas that matter most.

Entrepreneurship and what to build

• Encourage ambition. Mollick asks students to consider what they could do that is more ambitious than their initial idea. AI lowers the amount of technical work required, making bigger goals more feasible.
• Start with problems you know well. His students—including surgeons, finance professionals, humanitarian workers, and soldiers—have used their field experience to build tools for specific problems, such as tracking supply-chain risks or making medical information easier for patients to understand.
• Don’t dismiss familiar startup ideas too quickly, but understand their real-world complexity. Ideas such as replacing real-estate brokers or improving home-cooked meal delivery keep recurring because people want them. Yet they may be harder than they appear due to trust, legal, political, and social factors.
• Aim beyond selling tools. Mollick cautions against thinking only in terms of selling software or “picks and shovels.” If AI helps make an existing business more efficient, entrepreneurs may be better off applying that capability directly to the business or problem.
• Software is becoming more accessible. People who previously needed a technical co-founder can increasingly build tools themselves. Mollick sees this as an opportunity for entrepreneurs to use software to create newsletters, videos, and other ventures—not necessarily to start a software company.
• Ideas matter most when paired with execution. AI can generate many ideas, but ideas often change as a business develops. Customer conversations, experimentation, and implementation remain essential. AI can make early testing cheaper, including by acting as a rough simulated customer, though it cannot replace real customers.

Four capabilities for using AI well

Mollick identifies four capabilities that help people use AI effectively:

1. Deep knowledge: Expertise in the relevant field helps a person recognize whether AI’s advice is useful and apply it realistically.
2. Broad knowledge: Familiarity with other areas—such as design or accounting—helps people use AI across disciplines and make connections they might otherwise miss.
3. Agency: People need to decide what to delegate, what to retain, and where to participate in the work.
4. Taste: The ability to choose promising ideas or outputs from many possibilities helps people guide AI toward work that is distinctive and worthwhile.

He describes the AI “jagged edge”: systems can perform impressively in some areas while failing at tasks that seem simpler or more obvious. Their capabilities are hard to predict, so people need to evaluate results rather than assume reliability.

Preserving human judgment and distinctiveness

• AI can push users toward similar ideas, strategies, and writing styles. Human variation can therefore become a competitive advantage, particularly in areas where a business needs to stand out.
• To develop taste, Mollick recommends studying good and bad examples, learning how experts explain what makes something effective, and seeing a wide range of approaches. Education can help by giving people language and practice for evaluating work.
• AI can make creative and evaluative decisions, but the important question is whether those decisions match the human’s goals, knowledge, and preferences.
• Mollick advises people to avoid “cognitive surrender”—handing over tasks so automatically that they stop practicing important skills. In areas where a person wants to maintain an advantage, it may be worth preserving some effort and difficulty rather than delegating everything.

Learning, junior jobs, and career skills

Mollick says AI is disrupting the traditional apprenticeship model for white-collar work. Junior employees once learned by doing tasks for senior colleagues and receiving feedback. Now, a junior worker may use AI to produce the immediate deliverable, while a manager may see less reason to spend time teaching or reviewing the work.

He sees two possible organizational responses:

• An easier but concerning route: Hire more senior employees and reduce entry-level opportunities.
• A more deliberate route: Make training a real part of work, giving people time to learn and assessing their development rather than expecting them to be productive immediately.

The skills that make someone “good enough” may also change. For example, coding may depend less on writing code and more on defining a vision, communicating clearly, or understanding what should be built. The relevant human bottlenecks can shift as AI improves.

How to manage AI agents

Mollick recommends treating AI as a team that needs clear direction and supervision, not as a magic box that can be trusted after one prompt.

A practical approach he describes:

• Set the goal and standards. Explain what you want, what counts as good or bad, and what constraints matter.
• Break the assignment into stages. Ask the AI to propose approaches before asking it to carry one out.
• Create decision points. Tell it when to report back, present alternatives, or ask for human input.
• Request options rather than a single answer. Compare several approaches and decide which ones to pursue.
• Review early and intervene. Watch for the agent going in an unhelpful direction, and correct it before it has done extensive work.
• Use feedback loops. Explain what worked, what failed, and why, then ask the system to revise or investigate.
• Keep human agency in consequential choices. Delegate tedious execution, but remain involved in decisions that shape the strategy or outcome.
• Build trust gradually. Test an agent on low-stakes tasks first, understand its tendencies, and be especially cautious when models or interfaces change.
• Keep oversight visible. Know what instructions and context the agent is using, and avoid allowing it to take significant actions without appropriate review.

As an example, Mollick describes asking agents to research the structure of stories and business pitches. He first asked for possible research approaches, then reviewed and combined options, requested results for several approaches, and questioned the agent about apparent problems. AI handled repetitive work such as coding and sorting a large collection of stories, while Mollick retained the more interesting decisions.

Thinking, creativity, and idea generation

• Mollick recommends generating ideas individually before using AI or starting a group brainstorm. Group brainstorming can make people anchor on the first ideas raised and reduce variety. Individual thinking first preserves more independent perspectives.
• AI can still be useful afterward for expanding ideas, testing possibilities, and doing labor-intensive work.
• He stresses the value of “desirable difficulty”: people often need to exert mental effort to learn. If AI removes every challenge, users may lose practice without realizing it.
• Creativity is not the only valuable human contribution. People may bring useful strengths in performance, communication, emotional connection, or execution even if idea generation is not their strongest skill.

Fact-checking and delegation

Mollick says AI can be especially effective at fact-checking and tedious administrative work. While checking his book, AI reportedly found not only errors in his own text but also errors in sources he had cited—including a mistake in an academic paper’s table and an incorrect city in an article.

He nevertheless cautions against delegating consequential actions too quickly. AI-written messages can sound unlike the user, and agents may send messages or act more proactively than intended. He recommends understanding the system’s instructions and behavior, using it first in low-stakes situations, and building trust through experience.

Overall perspective

Mollick acknowledges that rapid changes in AI can be frustrating and unsettling. He recommends combining curiosity and experimentation with a clear-eyed recognition of the disruption. His broader goal is an “augmentation” model: AI handles tedious work while people spend more time on meaningful decisions, learning, and creative or strategic work.

Speakers and sources featured

Speakers

• Ethan Mollick — Wharton professor of entrepreneurship and innovation; interviewed guest.
• The host — Interviewer and creator of the Silicon Valley Girl channel; her name is not given in the subtitles.
• MongoDB sponsor narration — A sponsored segment describing MongoDB Atlas Agent Engine; the narrator is not identified.

Sources and studies mentioned in the interview

• Anthropic studies — Cited in discussion of coding ability and the benefits experienced users may get from AI.
• A study of Kenyan entrepreneurs using GPT-4 — Described as finding an increase in profits for already successful entrepreneurs who could implement the advice.
• Alex Eames and colleagues’ negotiation-agent research — Discussed in relation to differences in people’s ability to build agents and the effects of repeated negotiations.
• Research on brainstorming and group idea generation — Invoked to support generating ideas individually before group discussion.
• The host’s closing reference to Brian Chesky — A recommendation to watch a separate interview, not a speaker in this video.