Ethan Mollick on AI Agents: Keep Your Thinking, Build Beyond SaaS

Should you still build a software company when AI makes software easier to produce? And how do you use increasingly capable agents without letting them make every meaningful decision? In Marina Mogilko's interview with Ethan Mollick, the two questions lead to the same answer: bring your knowledge of a real problem, let AI expand what you can attempt, and deliberately keep the choices that make the work yours.

Watch Marina and Ethan's Conversation

Mollick is an associate professor of management and co-director of Wharton's Generative AI Labs, according to his faculty profile. TIME included him in its 2024 TIME100 AI list. His book Co-Existence: The Next Phase of AI was released in October 2026. The interview includes a MongoDB-sponsored segment; the discussion here is editorial, not a product endorsement.

The Startup Question: What Happens When Software Is the Easy Part?

At 07:26, Mollick says the outlook for entrepreneurship is encouraging, but he thinks founders have become too focused on software-as-a-service as the thing to sell. In his Wharton class, surgeons, supply-chain specialists and people in humanitarian work are using AI to build around problems they already know. His point is not that a prototype equals a business. Domain knowledge helps identify which problem is worth solving; customer access, trust, delivery and economics still have to work.

He distinguishes software-heavy companies, which may need exceptional engineers working with AI, from software as a toolkit for making a newsletter, service or physical-world operation better. He also notes that established software vendors still serve real needs. The useful question for a founder is therefore not "Can an agent clone this app?" but "What outcome will a buyer value enough to keep using, and what can we do for them that a generic tool cannot?" That last question is our editorial interpretation of the interview, not a claim that all SaaS is obsolete.

Four Human Advantages Mollick Wants You to Bring

Mollick names the same four assets in the interview and in his own essay, The Overhang. They are not a guarantee of business success; they are a way to see why two people using the same model can get different results.

Mollick's view of taste is more specific than "AI cannot make creative choices." It can. The question is whether those choices express your aims. Seeing many good and bad examples and learning the language to discuss them helps build that ability, he says at 16:10.

Protect the Thinking You Still Want to Own

Mollick calls the risk of accepting AI's answer without doing your own thinking cognitive surrender. In his essay on choosing when to use AI, he points to research where AI improved immediate task performance but could also weaken independent understanding. This is a risk, not proof that every AI-assisted task deskills its user.

His personal boundary is to type the first draft of his writing himself. He also recommends making your own initial list of ideas before asking AI to brainstorm, so the first model suggestion does not anchor the whole conversation. He is not prescribing a universal ban on AI writing or ideation. He uses AI to critique, research and fact-check, while choosing where effort remains valuable to him. The concern about junior jobs is similarly forward-looking: if agents take the routine work through which novices used to develop judgment, employers will need intentional ways to train people. The interview does not establish that every entry-level role has already lost that function.

Manage Agents by Designing Choice Points

At 33:23, Mollick describes managing agents less like reciting a fixed workflow and more like delegating to a capable research team. He asks for check-ins, diverse options to choose among, and examples of good and bad work. Then he checks direction early, before the agent spends a long time optimizing for a misunderstood goal.

His story-shape research example is unusually concrete: ask the agent to establish what prior research is real; request several possible approaches; select some for a first pass; compare the results; combine the best parts; and question a suspicious finding. The agent does the tedious coding and sorting, while the human keeps the research question and the important choices. This is not a claim that the agent's work is automatically correct.

The book fact-checking account shows both the upside and the limit. Mollick says AI flagged mistakes in his draft and in cited material, alongside human fact-checkers. Those are his reported findings; this article has not inspected the manuscript or independently confirmed each correction. For high-stakes claims, use the lead as a reason to revisit the primary source, not as a final verdict.

A Small Agent Brief With Two Stops

This original JQ AI SYSTEMS brief turns the interview's choice-point idea into a bounded first test. Use a low-risk question, write your own initial view before pasting it, and compare the agent's alternatives with what you would have done. A successful run should leave you with a better decision and an auditable trail, not merely a polished answer. It has not been tested as a product workflow.

For a business idea, add one human customer conversation before committing to build. Mollick notes that AI can simulate a customer as an inexpensive first pass, but the simulation is not a customer.

Video Chapters

Ambition and what to build

Skills and independent thought

Agents, work and the future

Sources and Limits

Checked 9 November 2026. The video, supplied transcript and host's episode page support the account of what Marina and Ethan said. His Wharton profile, publisher listing, and essays on the four advantages and cognitive surrender provide primary context. Startup forecasts, the effects on junior training, and the reported book corrections are Mollick's analysis and experience, not independently measured outcomes in this article.