This paper extends the standard cheap talk and optimal delegation models by assuming that the agent's private information is endogenously learned instead of given by nature. When learning is overt, we show that the classic "delegation principle" on the project selection still holds, but a similar "delegation principle" on the learning decision does not hold; the learning task may not be assigned to the party who has a superior learning technology. The principal either delegates the learning task and the decision authority or retains both for herself. It is never optimal for her to delegate one and retain the other. Compared to centralization where the principal invests in learning directly and makes decisions without consulting the agent, cheap talk generates under-investment in learning, while optimal delegation generates over-investment in learning. In the cheap talk model with covert learning, the agent invests less in learning and both parties have lower welfare than when learning is overt. In the optimal delegation model with covert learning, the principal has to eliminate a set of "intermediate" projects in order to induce an ex ante uninformed agent to learn.
Available at: http://works.bepress.com/jia_xie/20/