Generative AI tools are freely available and demonstrably productive, yet adoption among microentrepreneurs remains low. A natural explanation is that entrepreneurs hold pessimistic beliefs about AI’s profit potential, and correcting those beliefs should, in principle, be sufficient to trigger adoption. We test this directly in a pre-registered field experiment with 736 Chilean micro-entrepreneurs, randomly assigning participants to informational vignettes depicting a fictional entrepreneur applying ChatGPT to her own business, or to an uninformative control. The vignettes shift profit expectations upward in a pattern consistent with Bayesian updating, with revision concentrated among the most skeptical participants. Yet willingness to pay for personalized AI coaching and interest in AI training are each unchanged in the same survey session as the belief update, ruling out belief decay. A two-month follow-up confirms the disconnect — AI adoption, business practices, sales, and profits are statistically indistinguishable across arms. Belief correction is complete, Bayesian, and precisely estimated, yet adoption is unmoved. The result establishes a boundary condition for information provision policy: when the technology is general-purpose and firm-specific capability gaps persist, correcting profit beliefs is necessary but insufficient for adoption.

This study examines how mandatory disclosure policies affect entrepreneurial resource acquisition when traditional credibility signals vary. We propose a conditional substitution framework where platform certification replaces missing reputational signals. Exploiting Kickstarter’s mandatory AI use disclosure policy as a quasi-natural experiment across 27,845 projects, we find that mandatory disclosure significantly enhances funding outcomes for AI-using entrepreneurs overall, though effects vary substantially by entrepreneur type. First-time entrepreneurs lacking track records experience positive effects, while serial entrepreneurs with established reputations experience negative effects. While prior work demonstrates that signals typically complement one another, we provide empirical evidence that platform verification’s value depends critically on entrepreneurs’ existing signal portfolios. These findings identify boundary conditions to signal complementarity and reveal how governance design democratizes market access.

Technology licensing and inventor mobility have largely been treated as parallel but separate channels: codified knowledge flows through contracts, tacit knowledge through careers. We develop a demand-side perspective linking these channels: licensing creates targeted demand for inventors at precisely one firm—the counterparty—and relational governance provisions substitute for such hiring by providing structured access to tacit knowledge. Using 3,558 pharmaceutical licensing agreements linked to inventor career histories, a staggered difference-in-differences design finds that outlicensing nearly doubles inventor mobility toward the licensee relative to other firms (18.6 vs. 10.4 percentage points). Relational governance provisions attenuate this targeted outflow by half while leaving other-firm mobility unaffected. These findings reveal a previously unrecognized human-capital consequence of licensing: contract design shapes post-agreement human capital distributions.

We develop a model of mentorship that centers on how mentors reduce uncertainty rather than how they transfer knowledge. Mentors do not merely add information; they filter and reorganize the mentee's decision environment, pruning dominated paths and transmitting heuristics for future evaluation. We formalize two distinct filtering mechanisms and show that they map onto the exploration-exploitation tradeoff at the level of the mentoring interaction. Under internal filtering, the mentor exposes the mentee to varied framings and prompts self-directed reasoning, building the mentee's own capacity to evaluate strategies. Under external filtering, the mentor delivers deterministic, high-confidence guidance, reducing ambiguity directly. We derive conditions under which each mechanism dominates for skill persistence after mentor removal and for mentees differing in initial self-regulatory capacity. We test these predictions in a randomized controlled trial with approximately 90,000 enrolled unemployed job seekers through France Travail, the French national employment agency. Participants in online soft-skills courses — Personal Initiative and Negotiation Skills — are randomly assigned to course content alone, content with an exploratory AI mentor, or content with a directive AI mentor. Both AI mentors share identical self-regulated learning infrastructure — prompting learners through cycles of planning, monitoring, and reflection — and identical course material; they differ only in the filtering mechanism. We cross-randomize across courses to test generalizability. Outcomes span course completion, knowledge retention, entrepreneurial skills, job search, venture creation, employment, income, and wellbeing, measured up to 18 months post-intervention.

Organizational theory predicts that when productive capability disperses to frontline workers, firms should decentralize governance authority to match it. General-purpose AI has dispersed productive capability to every knowledge worker, yet firms restrict access and workers conceal productivity gains even where access is permitted. We develop a formal model showing that both responses are individually rational equilibria, not failures of attention or management intent. IT restricts access because it bears the full cost of any security incident but captures none of the productivity benefit. Workers conceal AI-enabled gains because disclosure invites surplus extraction, a hold-up problem that arises at the moment of knowledge revelation rather than at the investment stage of relationship-specific assets. We prove that a polycentric governance design resolves both failures jointly and outperforms any hierarchical alternative under the conditions the model identifies. Three mechanisms work in combination. Contribution-based screening identifies productive adopters without self-disclosure. Collective observability within a community of practice makes the commitment not to extract the surplus self-enforcing in a strictly larger set of organizational environments than any hierarchical arrangement can achieve. A community-governed repository aggregates the tacit, task-specific knowledge that centralized collection cannot reach. Enterprise AI programs that transition from shadow adoption to large-scale knowledge commons of community-built tools are consistent with the model’s predictions.

Quality mentorship improves economic outcomes, but its reliance on scarce human capital limits scalability — particularly in developing economies. A fundamental question is whether the filtering and guidance functions of mentorship require a human relationship or whether they can be delivered algorithmically. We investigate this in a randomized controlled trial with 120,000 secondary students across Ecuador, Peru, and Chile. Students enrolled in online Personal Initiative courses are randomly assigned to course content alone, content augmented with an AI mentor, or content with access to trained professional mentors through a digital platform. Within the human mentoring arm, we randomize whether students observe mentor characteristics before selecting a mentor, generating exogenous variation in the information available for matching. This design separates the effect of mentorship itself from the effect of informed self-selection into mentors. We develop a framework distinguishing the informational function of mentorship — filtering uncertainty, transmitting heuristics — from its relational function, and derive conditions under which algorithmic delivery can substitute for human mentors. A secondary experiment randomizes whether follow-up surveys are conducted by AI voice agents or human interviewers, testing whether AI measurement reduces social desirability bias in noncognitive outcomes and estimating the cost-effectiveness of algorithmic data collection at scale. Outcomes include entrepreneurial skills, metacognitive capacity, and post-graduation employment and entrepreneurship.