Overview of This AI Expert Argues We're Underestimating AI's Productivity Impact — With Ethan Mollick
Alex Kantrowitz speaks with Wharton professor and AI researcher Ethan Mollick about why AI’s economic impact may be showing up later — and differently — than many expect. Mollick argues that AI is already producing real ROI inside companies, but the bigger productivity gains are being held back by organizational inertia, unclear operating models, and the fact that most firms still haven’t adapted their workflows to truly exploit the technology. The conversation also covers AI’s rapid capability gains, the rise of agents and “harnesses,” the risks of using AI in education and personal life, and why Mollick thinks the world is still early in the transformation.
Main Thesis: AI Is Creating Value, But Organizations Haven’t Caught Up
Mollick pushes back on the idea that AI “isn’t working” because broad productivity stats haven’t exploded yet.
- Individual gains are already real, but organizational productivity takes much longer to materialize.
- AI can make a single coder, analyst, or manager dramatically more productive, but companies still have to:
- coordinate teams
- redesign processes
- decide how work gets assigned
- determine who gets access to tools and tokens
- In other words, AI’s impact is being bottlenecked by management and organizational design, not just model quality.
Why productivity stats look underwhelming
- Companies are seeing more code, more content, and more experimentation.
- But the increase in outputs does not automatically translate into immediate economy-wide productivity growth.
- Mollick says the problem is that AI is single-player technology in a multiplayer organizational world.
Where Companies Are Seeing ROI Today
Mollick argues the ROI question is no longer “does AI work?” but “how do we scale it?”
Examples he cites
- Leadership teams using AI as a super-analyst to surface problems in company data.
- Financial services firms launching new product lines with AI support.
- Hedge funds / PE firms using AI to improve management and portfolio performance.
- Walmart and other large enterprises using AI for frontline support, store operations, and management guidance.
- Data migration and legacy IT projects being dramatically accelerated.
Key takeaway
Most large companies are already seeing some return, but they are still in the experiment-and-restructure phase, not the fully transformed phase.
The Real Bottleneck: Organizational Change
Mollick says the biggest unlock is not a better prompt — it’s structural change.
What change looks like
- Breaking up old team structures
- Recombining roles across engineering, marketing, product, and operations
- Giving small, high-talent teams large token budgets and hard problems
- Using AI to rethink what work should exist at all
The “leadership, crowd, lab” framework
He says successful transformation requires:
- Leadership that commits to change
- Crowd members who are using and inventing with AI
- Lab efforts that spread the lessons across the organization
AI Progress: Why It Keeps Getting Better
Mollick is bullish on continued capability gains and argues that many “walls” people worried about have not materialized.
Three things driving recent progress
- The LLM brain
- Reasoning improvements
- Better post-training and reinforcement learning
- The app layer
- New interfaces beyond the chat window
- Tools like Codex, Claude Code, NotebookLM, and other task-specific products
- The harness
- Tools that let AI act on the world
- Web access, code execution, computer control, and connected systems
Why reasoning mattered
- Reasoning models can “think out loud” before answering.
- This improved performance in math, science, and other tasks that require exploring intermediate steps.
- Mollick argues this changed the trajectory of the field.
His bottom line on progress
- He does not think AI progress has stalled.
- He does think every exponential eventually becomes an S-curve, but he doesn’t see a clear wall right now.
- Compute and power are real constraints, but not a sign of imminent stagnation.
The Overhang: Current AI Capabilities Are Underused
One of Mollick’s central ideas is that there’s a huge gap between what current models can do and what organizations actually use them for.
- He calls this the capability overhang.
- His view: we don’t necessarily need better models to get major transformation.
- We need better adoption, better workflows, and better organizational imagination.
AI Agents, Personal Assistants, and the Privacy Tradeoff
Mollick is enthusiastic about personal agents — his “Claw-like” category for AI systems that connect to personal data and act proactively.
What these systems can do
- Catch errors in emails
- Cancel subscriptions
- Negotiate refunds
- Manage routine digital tasks
- Flag mistakes before you send a message
The catch: privacy and security
He’s very clear that this comes with risk:
- prompt injection through email or other channels
- accidental data leakage
- access to sensitive financial or medical data
- unclear long-term trust boundaries
His view is basically: the risks are real, but current security hasn’t been as bad as many feared so far.
Education: AI Is Both a Tutor and a Cheating Machine
Mollick says education is one of the clearest examples of AI’s dual nature.
His view
- Bad use of AI: outsourcing the work and short-circuiting learning
- Good use of AI: using it as a personalized tutor
What works
- AI tutors outside class
- Active learning in class
- In-person exercises, case discussions, and debates
- Traditional assessments like blue books and other proctored formats
His takeaway
The solution is not “ban AI”; it’s to redesign education around:
- tutoring
- active engagement
- better assessment
Faith, Writing, and Cultural Authenticity
The conversation also explores how AI is entering intimate and value-laden spaces.
Examples
- AI-written sermons
- AI-assisted obituaries
- Religious interpretation and spiritual advice
- Personalized writing for families and communities
Mollick’s point is that society has not yet negotiated the right boundaries for these uses.
Cultural concern
As AI-generated writing becomes more recognizable, people may become more skeptical of:
- sermons
- LinkedIn posts
- ad copy
- podcasts
- public-facing writing that feels “too polished” or generic
Jobs and the Labor Market: Disruption Is Coming
Mollick emphasizes that AI will likely reshape jobs unevenly.
What he thinks is happening
- Some jobs will disappear
- Some will be augmented
- Some will be transformed rather than replaced
- Coders are already seeing their roles change dramatically
Broader message
- This looks like an industrial-revolution-style transition
- Big gains are possible, but the transition may be painful
- Society needs safety nets and policy preparation
Token Costs, Compute, and Deflation
Mollick argues that token pricing is falling fast enough that cost worries will likely fade over time.
- AI is highly deflationary on a per-task basis.
- More efficient models can delegate work to cheaper models.
- Even as usage increases, the cost per unit of intelligence keeps dropping.
His view: token anxiety is temporary; capability growth and usage expansion matter more.
Notable Insights
“You don’t get organizational productivity gains instantaneously from individual productivity gains.”
A core theme of the episode: AI can supercharge individuals without instantly transforming large firms.
“We don’t need new models for complete transformation.”
Mollick believes the bigger issue is adoption, not capability.
“The penalty of locking in a good enough solution based on today’s limits grows.”
Companies should avoid overcommitting to workflows built around today’s temporary constraints.
Bottom Line
Mollick’s message is cautiously optimistic: AI is already producing real value, but the biggest gains are still ahead because most companies, schools, and institutions haven’t fully adapted. He sees the current moment as one of coexistence — a period where humans and AI are learning how to work together, while organizations slowly rewrite their structures around the technology.
The big risks are not that AI stops improving, but that:
- institutions misuse it,
- over-automate with it,
- or fail to reshape themselves quickly enough to capture its benefits.