Wharton’s Top AI Professor: Stop Starting Software Companies — The Real AI Money Is Here
Use AI to remove busywork, but keep your own ideas, expertise, and decision points in the process so your work stays distinctive.
- Start with your own ideas before asking AI to brainstorm; compare its options afterward to preserve variety.
- Manage agents like a team: specify what good looks like, request check-ins and alternatives, and intervene early.
- Test business ideas cheaply with real customers; AI can help explore and build, but execution and customer feedback matter.
The agent-management walkthrough shows a practical, iterative feedback process. · Best for: Entrepreneurs and managers using AI to develop ideas, build products, or delegate work.
Wharton professor Ethan Mollick argues that AI makes it easier to build, but entrepreneurs should focus on real domain problems rather than selling software alone. Preserve human expertise and agency, test ideas with customers, and guide AI through frequent choices instead of handing it a task and trusting the result.
Ambition and choosing problems0:57
- AI expands what students can build, especially when they apply expertise from fields they know.
- Look beyond software as the product: use it to make or do something valuable.
- Some problems, such as replacing real-estate brokers, involve social, legal, and political barriers that are easy to underestimate.
Software businesses and entrepreneurship7:02
- Mollick sees software as increasingly accessible and often a toolkit for other work, not necessarily the business itself.
- Companies may build software themselves when vendors mainly package capabilities they already have.
- Experienced entrepreneurs were better able to act on AI strategy advice than failing entrepreneurs in the study he describes.
Expertise, taste, and distinctiveness13:14
- He identifies deep knowledge, wide knowledge, agency, and taste as useful for working with AI.
- AI tends toward consensus and does not cover the full range of human ideas; add distinctive expertise where differentiation matters.
- Build taste by examining varied good and bad examples and learning to explain what works.
Protect learning and generate ideas yourself18:30
- AI can tutor, but using it to get answers without doing the work can prevent learning.
- Add deliberate friction in important areas so you do not surrender skills you may need later.
- Before group brainstorming—or AI brainstorming—write ideas individually to reduce anchoring and preserve variety.
Test ideas and manage agents31:22
- Entrepreneurship is a process of learning cheaply through customer interviews, prototypes, and repeated exploration.
- For agent work, ask for several approaches, review drafts, give feedback, and choose which direction to develop.
- Specify what good and bad look like, set check-in points, request alternatives, and intervene early when the agent goes off course.
Trust, adaptation, and what to aim for39:18
- Build trust with AI gradually: start with low-stakes tasks and understand its behavior before delegating higher-stakes work.
- Changing models and interfaces can disrupt workflows; focus on experimentation and adapt as capabilities shift.
- Keep ambitious benchmarks for work AI cannot yet do well, such as producing genuinely original research or a compelling story.
Step by step
- Write down your own ideas before asking AI to brainstorm.
- Ask the AI for multiple approaches, not one recommended answer.
- Review the options and give specific feedback; ask it to draft several chosen approaches.
- Compare the drafts, select what works, and request revisions or combinations.
- Delegate tedious research or implementation while keeping meaningful decisions and checkpoints for yourself.
- Test the resulting idea with real customers or other evidence; do not treat AI-generated feedback as a substitute for customers.
Tools & resources mentioned
- Wharton Generative AI Lab — Mollick says he leads the lab and describes experiments on AI and work.
- Codex scheduled tasks — described as running recurring summaries and reports.
- Granola — mentioned as a source of notes from a past conversation.
- MongoDB Atlas agent engine — sponsor-described service for building, deploying, governing, and giving agents memory.
- Salesforce, Snowflake, Amazon S3 — data systems the sponsor says Atlas agent engine can access without moving data.
- Slack and Workday — examples Mollick mentions of software companies still using established software vendors.
Notes are AI-generated from the video's transcript and may miss nuance.