This AI Expert Argues We're Underestimating AI's Productivity Impact — With Ethan Mollick

Apple

Across the shows: 10 speakers in other episodes, 2 bullish and 11 bearish calls, past 30 days

• Ethan Mollick · Wharton School, University of PennsylvaniaBearish

  Apple may run into trouble because it lacks a frontier model needed for personal agents.

  Mollick argues that personal agents, which connect to a user's data and act proactively, need a recent frontier model. Older systems needed elaborate scaffolding such as RAG pipelines and vector databases, but a frontier model now handles that work itself. Companies that did not build LLMs at scale, which he names as Apple, may therefore be unable to offer this kind of work and could run into trouble.

  “It's kind of one of the reasons why I think companies that did not engage in trying to build LLMs at scale like Apple might run into trouble because if you don't have a frontier model, you can't do this kind of work.”at 42:40

AI & jobs

• Ethan Mollick · Wharton School, University of Pennsylvania

  Organizational productivity lags individual AI gains because companies haven't adapted their structures to AI yet.

  Individual workers can produce far more code with AI (a 100x gain is cited), but organizations run on sprints, coordination with marketing, and standups, so the gain does not translate. Mollick cites a study showing huge jumps in coding output but only about 30% more products shipped. He argues AI is a single-player experience while organizations are multiplayer, so the gains stall until processes are redesigned.

  “Produce a 100 times more code than you did before, and that's great if you're a solo entrepreneur producing a lot of code. If you're in an organization that's doing two week long sprints... that 100 times productivity gain is actually... Can be a barrier.”at 3:25

• Ethan Mollick · Wharton School, University of Pennsylvania

  AI is forcing companies to break up teams and redo projects, cutting costs dramatically.

  Mollick says real transformation requires structural change: pulling apart existing teams, dissolving role boundaries (coders doing design, designers coding), and reassessing projects from scratch. His example is a company that renegotiated a $200 million mainframe data migration plan down to $20 million, with the CTO thinking it could be done for about $2 million. He argues most organizations have lost the muscle memory for this kind of organizational experimentation.

  “Thinking about structural change, which takes years for a lot of organizations to do... pulling apart organization teams... went from a $200,000,000 data migration plan... to a $20,000,000 plan.”at 15:03

• Ethan Mollick · Wharton School, University of Pennsylvania

  Expects AI won't deliver the 10% GDP growth some predicted, though broad transformation will still come.

  Mollick is skeptical of the aggressive macro forecasts, including 10% GDP growth, and calls them outside any reasonable expectation. He still expects organizations to transform over the next five or ten years as the capability overhang gets filled, with large socioeconomic effects, so the disagreement is about the size of the productivity effect rather than whether it arrives.

  “I don't think we're gonna see some of the insane numbers predicted of 10% growth, but, you know, who knows? ... But we will see transformation happen one way or another.”at 28:52

• Ethan Mollick · Wharton School, University of Pennsylvania

  Coding jobs have changed radically in two years; the best coders no longer write code themselves.

  Mollick argues that for anyone with access to modern AI tools, coding is no longer the same job. The best coders he knows are not writing code anymore, which is his evidence that disruption is already here. He says the debate among economists about whether there will be more or fewer coders is unresolved, and that he cannot predict whose job is replaced, augmented, or where value goes.

  “Any coders who are listening to this, their job is radically different than it was two years ago... The best coders I know are not writing code anymore.”at 29:26

• Ethan Mollick · Wharton School, University of Pennsylvania

  Senior developer demand is rising as AI expands output, while programmers' work shifts toward managing code.

  Mollick says the debate over coders is a good example of how AI changes jobs. Right now demand for senior developers is rising, because AI lets a small organization with one coder produce work that once needed a team of 20. Programming jobs are also transforming from writing code to managing AI-written code. He hedges: maybe there are more programming jobs, but 'we don't really know,' and he says jobs will be invalidated, transformed or grown in ways we cannot predict.

  “So right now, we are seeing more demand, at least for senior developers, because there's now more coding you could do.”at 1:01:42

Enterprise AI adoption & AI software

• Ethan Mollick · Wharton School, University of Pennsylvania

  Nearly all large companies report AI ROI; the question is now transformation, not whether gains exist.

  Mollick cites his Wharton colleagues' survey of about 1,000 large companies, where 85% of firms with over $1 billion in revenue reported positive ROI from AI as of last year. He says CEOs he talks to are no longer asking whether AI pays off. Individual examples he cites include a Fortune 100 company using AI as a super-analyst for its senior team and PE funds seeing portfolio companies grow faster. He notes the returns are at the transformation stage and many firms are still in experimentation.

  “There is almost nobody at this point in the data from my colleagues at Wharton back this up who isn't seeing some ROI from AI. Like, it's not... This is not an issue anymore that people aren't seeing ROI.”at 7:19

• Ethan Mollick · Wharton School, University of Pennsylvania

  The MIT study showing 95% of AI pilots fail is a bad study; high pilot failure is healthy.

  Mollick calls the MIT NANDA 95% pilot failure study a famously bad study, saying it was based on interviews at a conference that were coded by the author himself. He notes more reasonable reports put the failure rate closer to one in eight. He says he hopes most pilots fail because companies need to run experiments that fail to learn how to use AI, given there is no playbook. He discloses that he takes no money from the AI labs.

  “That is a famously bad study. ... Somebody went out and just interviewed people at a conference and coded that himself.”at 13:33

• Ethan Mollick · Wharton School, University of Pennsylvania

  AI agents became practical only around late 2025, when models reached the Opus 4.5/4.6 level.

  Mollick argues the shift from the chatbot era to the agentic era depends on model capability crossing a threshold. Practical agents were not possible until models reached roughly the Opus 4.5 or 4.6 level around November or December of last year. This explains why product formats keep changing, since each capability step makes new product categories possible.

  “Practical agents were not possible until we reached Opus four six level or four five level about November, December of last year. Like, that was when agents became practical.”at 27:15

• Ethan Mollick · Wharton School, University of Pennsylvania

  Companies locking in AI vendors and solutions based on today's models will be left behind as models improve.

  Mollick gives an example: someone who built an elaborate RAG system around a Llama 3 or Llama 4 model in 2024 saw it become obsolete quickly. He worries companies are choosing vendors and solutions based on current technology and will be blown away as capabilities keep rising. He argues firms need to be ambidextrous, exploiting current technology while exploring for the future.

  “I worry a lot that companies are picking vendors or solutions based on today's technology, and then they are going to be blown away as the technology continues to increase.”at 1:06:09

AI pricing & unit economics

• Ethan Mollick · Wharton School, University of Pennsylvania

  Rising token spend is a management problem, not evidence that AI is failing to pay off.

  Mollick says token costs are a growing expense companies want to control. He argues this reflects a lack of decisions about who uses tokens, for which processes, and under what limits, so people use as many as they can when they are incentivized to. He dismisses the public spending pullbacks as coming mostly from firms with competing AI products (Microsoft, Meta, Palantir), whose revenues are still growing. He reads cutbacks as a sign of unmanaged use rather than uselessness.

  “I think people are worried about token cost, but that's often a management problem. Like, if you don't make any decisions about how tokens should be used... they'll use as many tokens as possible.”at 10:43

• Ethan Mollick · Wharton School, University of Pennsylvania

  Falling token prices are a deflationary push that will make token-cost worries fade, despite rising usage.

  Mollick says he is not worried about long-run token costs. Usage is rising because agentic work consumes many more tokens, and Jevons' paradox applies. Organizations are also getting more token-efficient: a frontier model can orchestrate cheaper small models to do subtasks. The result is that AI does more work for less money, which he calls a really deflationary push. He expects the concern about token costs to fade the way concerns about vibe coding and RAG have faded, because costs keep falling even as more compute comes online.

  “So that lower cost means AI does more stuff for less money in a very short period of time. Really, really deflationary push.”at 1:03:34

AI: overall outlook

• Ethan Mollick · Wharton School, University of Pennsylvania

  Current AI capabilities are underused; transformation needs adoption, not better models, over the next decade.

  Mollick calls the gap between what current models can do and what people actually use the overhang. He argues that the next decade of transformation in how work is done needs no new models, only adoption, reorganization, and imagination. The lack of playbooks is the bottleneck, which is why he says transformation is inevitable and the job is guiding it.

  “I call this the overhang. Right? And there's a huge gap between current capabilities and future capabilities. We don't need better models for complete transformation of a lot of how we work in the next ten years.”at 24:39

• Ethan Mollick · Wharton School, University of Pennsylvania

  AI capability progress is tracking lab expectations; no slowdown or hard technical wall is visible.

  Mollick says he has seen little evidence that AI labs are surprised by their own progress: they 'seem to have been pretty spot on, actually shockingly so, about the continued development and doubling of the technological ability.' He cites reasoning models and the math breakthroughs as evidence. Demand has grown faster than the labs expected, but the technology has overcome barriers he once thought could cause a slowdown. He says forecasters keep underestimating progress because exponentials are hard to predict. He has not identified a technical barrier that would stop steady gains, and he expects compute and power to constrain the pace rather than stop it. He concedes every exponential eventually becomes an S-curve, so a limit exists, but he cannot point to one.

  “I haven't actually seen a lot of evidence of that. They seem to have been pretty spot on, actually shockingly so, about the continued development and doubling of the technological ability.”at 34:03

Robotics & autonomous vehicles

• Ethan Mollick · Wharton School, University of Pennsylvania

  General-purpose AI models are good at robotics, so practical robots may emerge within five years.

  Mollick points to the fact that general-purpose models are quite good at robotics as a sign that physical AI could arrive soon. He says that in the next five years robots may finally emerge and reshape the physical world. He hedges that this depends on how one feels about robots and that it is speculative. Without robots doing work, he expects the built environment and city streets to change little.

  “Do practical robots exist in this world? ... general purpose models actually are quite good at robotics. So maybe in the next five years, we finally see robots emerge.”at 30:19

• Ethan Mollick · Wharton School, University of Pennsylvania

  LLMs may finally unlock general-purpose robotics, though he is not an expert and cannot confirm it.

  Mollick says his view on robotics has shifted over the last few months. Six or seven months ago, robotics experts he spoke to were skeptical, and humanoid robots still do not indicate progress. He now thinks LLMs may unlock robotics. He cites early demos from Astra and especially Argonne doing generalized robotics work with a camera and a robot. He says this might be the moment things start to take off, but he is not expert enough to know whether it is real.

  “I'm starting to think they might be right. ... we've seen some really interesting early demos of Astra and especially Argonne being able to do generalized robotics work with a camera and robot.”at 49:35

Consumer AI products

• Ethan Mollick · Wharton School, University of Pennsylvania

  AI tutors improve learning, while unstructured AI use that does the work for students harms it.

  Mollick says the evidence is clear: AI used as a tutor improves learning, while naked AI use that does the work hurts it. He cites meta-analyses and individual RCTs, including studies from Stanford and Kenya, that show positive effects from AI tutors. He reads the MIT brain-activity study as showing reduced activity when people recall essays they did not write. He also says learning requires effort, so offloading work to AI causes deskilling. Schools should use one-on-one AI tutoring outside class and active learning, with assessment through tests and blue books.

  “But if you use AI as a tutor, you learn. Like, we've got enough evidence... Positive impact on education, naked AI use, negative impact on education.”at 53:58

• Ethan Mollick · Wharton School, University of Pennsylvania

  Personal AI agents now handle tasks proactively, including cancelling subscriptions users couldn't exit.

  Mollick describes personal agents, which connect to a user's data and act proactively. He says his own agent, Dot, has spotted errors in his emails and cancelled several long-term subscriptions he could not exit. The agent handled them without him needing to know how. He argues this kind of friction-removing automation is powerful and that companies are not ready for agents that work at scale with little effort, though the privacy and security risks remain.

  “I've already had my dot cancel several long term subscriptions I couldn't figure out how to get out of. It's dealt with them for me.”at 1:04:28

Frontier AI labs & model race

• Ethan Mollick · Wharton School, University of Pennsylvania

  AI labs' market has grown faster than they expected; demand exceeds supply and all are compute constrained.

  Mollick reframes the surprise around AI progress. The labs were not surprised by capability gains so much as by market demand, which has exploded faster than they expected. Demand has been higher than they anticipated, and they are all now compute constrained. He adds that Ilya Sutskever's data-wall warning was not a claim that forward progress would stop.

  “Their market has exploded [better] than they thought, which is different. Demand has been higher than I think they anticipated. Right? And they're all compute constrained at this point.”at 33:26

AI infrastructure & data centers

• Ethan Mollick · Wharton School, University of Pennsylvania

  Physical compute, chips and power will bottleneck AI inference and will force trade-offs against training.

  Mollick argues that the limit on AI development is physical. The number of chips is finite, and the power available for inference is limited. Labs must therefore choose between spending compute on training new models and selling inference. He says this creates a clear physical-world bottleneck in inference and compute. He adds that the constraint limits how fast gains can be deployed, but he does not think it stops steady gains in capability. He also says labs are compute-constrained because demand has exceeded what they expected.

  “So, I mean, there are limits to how much compute we can deploy. Right? And limits the number of chips out there.”at 37:09

Open-source AI models

• Ethan Mollick · Wharton School, University of Pennsylvania

  Many open-weight models are distilled from Claude or ChatGPT, so they sound like them.

  Mollick says that when he uses open-weight models, he quickly realizes many are distilled from Claude or ChatGPT, so they talk like those models. He finds it refreshing to use a model that does not sound like either of them, which he says is the case for Gemini 4. The argument rests on his observed style similarity, and he does not give a formal source.

  “a lot of the open weights models are actually distilled from either Claude or ChatGPT, so they talk like Claude or ChatGPT.”at 47:04

Cybersecurity

• Ethan Mollick · Wharton School, University of Pennsylvania

  AI-driven cyber threats are definitely coming, even though feared negative impacts have been fewer so far.

  Mollick says he has predicted more negative impacts from AI than have happened so far. Data leakage and deepfakes have been less severe than expected, and people with AI tools have not yet caused a major incident inside companies. He still expects cyber threats to arrive, and he says so directly. He does not give a specific mechanism or timeline in this passage.

  “I think I predicted more negative impacts than we've had, which doesn't mean they're not coming, by way. I think cyber definitely is coming.”at 48:36

Banks

• Ethan Mollick · Wharton School, University of Pennsylvania

  AI agents could undermine banks, which depend on customers' inertia in low-yield deposit accounts.

  Mollick agrees with an ApolloResearch point that AI agents threaten the banking system. Many people keep small sums in low-interest accounts because moving money is too hard. Personal agents could move the money into high-yield savings automatically, which would cut the deposit inertia the economy relies on. He says companies are not ready for this. Kantrowitz adds that large companies move cash between accounts for yield, but consumers don't.

  “one of the most interesting was the threat to the banking system. ApolloResearch pointed this out. Yes. This is true. ... the economy basically depends on us being lazy with small amounts of money that we keep in low interest bearing accounts”at 1:04:57