Ethan Mollick on Agents, Apprenticeship, and Rebuilding Organizations for AI

Ethan Mollick opens with a warning about certainty: despite conversations with AI labs and chief executives, he says nobody has a complete playbook. Lauren Crichton asks what has changed since his previous conversation with Sana roughly a year earlier.

Mollick describes three broad eras: predictive machine learning before 2022, a period of back-and-forth chatbot use that he calls co-intelligence, and an agentic era in which people assign longer-running tasks. He still sees a need for people in the loop, but says models can now act more independently than earlier chatbots.

What surprised him was the speed of the shift. Companies moved from asking whether AI was worth using to asking how to use it, while capabilities continued to improve. He says the uneven shape of those capabilities determines where AI is useful and where integration becomes difficult.

Crichton asks about the ‘jagged frontier’: systems can excel at difficult tasks and fail at surprisingly simple ones. Mollick credits his co-authors and says the boundary moves. Reasoning models closed some earlier gaps, while other capabilities have improved more steadily.

Long-form fiction remains a weakness in Mollick's experience: he finds repeated clichés, similar voices and weak plotting. At the same time, current models can turn a directory of his research material into a reasonably good doctoral-level thesis, although he thinks they are less imaginative about choosing an interesting topic.

He urges people to map the smaller frontier inside their own jobs. A model provider cannot tell every supply-chain manager or podcast designer exactly which tasks work well; someone with expertise in the job has to test those boundaries as they change.

The speakers consider whether readers will care if fiction is human-made. Mollick expects tension between a desire for authenticity and the convenience of personalized entertainment. He is willing to pay for human art, but does not assume every use calls for an artisanal product.

Mollick compares working with agents to working with wizards: one gives a task and sees an impressive result without necessarily understanding how it was produced. He says experiments at Wharton found that changes such as the ordering or attributed source of information can alter AI recommendations unpredictably.

He expects some opacity to persist. Even when a system exposes a reproducible chain, people may not inspect it; future internal reasoning also need not remain in a human-readable language. Crichton asks how he feels about developing systems that may improve other systems in ways people cannot fully understand.

Mollick separates the large question of recursive self-improvement from the everyday experience of using AI. He says no one knows what rapid self-improvement would look like, and that uneven capabilities may limit some outcomes. For daily work, he finds it useful to treat systems as if they had quirks, while remembering they are not people.

Crichton asks how agents can work in organizations when so much context exists only in people's heads. Mollick says organizations already coordinate imperfect people, and could combine people and AI to check each other. The harder problem is conveying unwritten incentives, relationships and history.

A project history can give an AI useful context that bare instructions omit. Some organizations will need to document more, Mollick says, while others may choose to redesign a process from scratch instead of translating an existing human workflow directly into agent tasks.

As an extreme experiment, he describes StrongDM's software dark factory. According to Mollick, people supply product direction but neither write nor inspect code; AI systems write it and adversarial AI tests it, with humans deciding whether to ship. He says the approach also requires substantial spending on tokens.

Mollick does not think removing people entirely is ultimately the right answer. Human participation can introduce variation and preserve meaningful decisions rather than letting one model's preferences shape everything. He expects many organizations to find a point between conventional coding and a fully autonomous factory.

Crichton raises a paradox: when people delegate tasks, they may lose the practice needed to judge the output. Mollick says established experts benefit because they can assess results and delegate execution. The harder question is how to create the next generation of experts.

Expertise develops through trial and error, mentorship and increasingly difficult work. Traditional apprenticeship benefited both managers, who received inexpensive help, and junior employees, who learned explicit and unwritten rules. Mollick argues that AI has disrupted those incentives because both sides can get faster results by delegating to a model.

Rebuilding apprenticeship will require organizations to separate learning from immediate output and tolerate some struggle. Crichton asks how much choice junior employees actually have. Mollick says individual agency depends partly on organizational incentives: someone who refuses AI while colleagues use it may be disadvantaged.

Schools can create incentives for learning by assessing students without AI through exams, essays or role-play. Organizations could also protect time for work without AI and assess the resulting skills. He says leaders cannot assume people will acquire every desired skill if models perform all of the work.

AI assistants are designed to be helpful and provide answers, but receiving an answer is not the same as learning. Mollick argues that removing every bit of friction is a mistake: deliberately adding friction at the right point can support understanding and agency.

Children are intrinsically motivated by different subjects, so education also uses external incentives to ensure broad learning. Mollick points to research in which unrestricted AI use made students feel they had learned when they had not, while a designed AI tutor improved outcomes.

He imagines AI tutors supporting work outside class, with classroom time used for discussion, active learning and assessment. Education will still change slowly and messily. At work, he suggests organizations may need more formal assessments because a manager impressed by AI-generated output cannot necessarily tell what an employee learned.

Possible workplace assessments range from defending a presentation under questioning to professional-style tests or protected non-AI work. The wider design question is what organizations should reward alongside productivity.

If productivity is the only goal, Mollick warns, organizations may generate more low-value output—such as vastly more presentations—without improving the result. Even software development processes must change around faster AI-assisted coding; a daily meeting cannot add much if every report is simply that the model did the work.

Mollick repeats that there is no universal AI playbook. His framework has three parts: leadership sets direction and incentives; the crowd of employees discovers uses; and a lab builds applications, drawing ideas from both leadership and employees. People must be rewarded for sharing uses they might otherwise hide.

He says the framework can apply to both large companies and small startups, with different-sized components. In a 50-person organization, the lab might be a part-time commitment by the CEO. Consultants can help, but cannot replace leaders' own experience, imagination and responsibility.

Organizations routinely experiment with products and marketing, but are less accustomed to experimenting with their own structure. Mollick proposes trying small cross-functional teams and ambitious, time-limited tasks, accepting that some will fail. New technology can initially lower productivity before benefits arrive, so leaders must be willing to work through that period.

Crichton asks about an article titled ‘IT departments are where AI goes to die.’ Mollick says the title was the publication's choice and broadens the argument: IT, legal and other functions are rewarded for reducing risk, while leaders may seek a reassuring standard playbook for a technology that does not fit one.

He worries that organizations normalize AI into a modest component such as natural-language document search and thereby miss more radical uses. He is not arguing for maximizing strangeness for its own sake; he wants organizations to recognize unusual capabilities and leave room for experiments that may change products or structures.

Conventional KPIs can constrain discovery when they ask only for a percentage improvement over past behavior. Mollick does not say every KPI should disappear; he wants room to test new organizational forms, products or markets, with permission for some experiments to fail.

Token costs create real pressure. Mollick says organizations need leadership support and must choose models and spending according to tasks, rather than treating the number of tokens burned as an achievement. Reinventing work may cost money initially while creating efficiencies later, including through better token allocation.

He suggests evaluating innovation through changes in products or organizational practice rather than volume alone. As examples he mentions an AI lab at Colgate-Palmolive that turns ideas into prototypes faster and organizations attempting to shorten very long legacy-software modernization plans.

Mollick argues that interfaces for many jobs lag model capabilities. AI labs understand coding and have built inventive tools for it, but managers and other workers lack equally mature ways to assign tasks, see what an agent did, guide its approach and audit results.

He imagines systems working alongside people through voice, video, documents and shared control, preparing information and offering feedback as work proceeds. He describes using AI to review emails and prepare drafts, while noting that he had not yet been willing to send those drafts.

Asked about human advantages, Mollick names four areas to cultivate: deep knowledge, broad knowledge, taste and agency. Taste can differentiate a person's work without proving that only humans have taste. He cautions against firm claims that judgment is uniquely human when agents already make decisions within delegated tasks.

Mollick expects white-collar jobs to exist in two years because capabilities remain uneven and organizations contain physical activities and unwritten context. Even an extremely intelligent newcomer cannot instantly understand every relationship and workflow. The longer-term outlook remains uncertain.

He explains the O-ring model of jobs: a job contains many tasks, but failure on a critical one can undermine the whole workstream. As AI takes over one bottleneck, another becomes central. In software, he sees a shift from writing every line toward assigning work, designing tests and judging whether an outcome is good enough.

Those bottlenecks will keep moving, and humans can still be cheaper or better for some tasks when AI tokens are costly. Mollick expects a chaotic transition. He notes that earlier industrial revolutions eventually produced more jobs, while stressing that living through the transitions was often painful.

Crichton asks whether management remains valuable. Mollick distinguishes the skill of managing work from the hierarchy of management. Defining tasks, specifying requirements, spotting risks and evaluating output quickly can all make someone more effective with AI.

The organizational hierarchy is a separate question: AI might reduce some layers or expand managers' span of control. Mollick has seen companies combine product and engineering work under a builder role. He cites a Procter & Gamble experiment with 776 employees in which individuals using GPT-4 performed as well as two-person teams without AI.

In that experiment, he says, AI also blurred the distinction between technical and business contributions: participants produced ideas combining both perspectives. He expects managerial roles to change as the boundaries between functions shift.

Asked whether departments might reorganize around problems, Mollick says agents still forget and lose information when working across functions, though better tools and human checks may help. He suggests small cross-functional project teams as a near-term approach and asks when such teams should call in specialists such as lawyers.

For parenting, Mollick favors supervised use when children are young. A parent might ask AI how to explain a problem, create images or play games together; older children can use dedicated tutoring modes with monitoring. He warns that an explanation can create a false feeling of learning if the child never practices.

He is skeptical of trying to predict a single safe future occupation for children. Instead, he recommends allowing them to pursue interests while developing flexibility, broad and deep knowledge, and agency. As with other technology, he emphasizes parental guidance.

The largest open questions, Mollick says, are how capable AI will become and how quickly. Today's organizational advice depends on current capabilities and assumptions about their trajectory; leaders should watch the changing frontier rather than simply extrapolate from a fixed snapshot.

Asked what he is proudest of, Mollick points to making AI more human-centered through educational tools and public writing. He argues that managers, citizens and policymakers should decide how to use the technology rather than treating any particular outcome as inevitable.