AI Agents: What They Are, What They Can Do, and What Can Go Wrong-Presentation

Presentation

Pdf Summary

The webinar explains that AI agents differ from predictive AI, which forecasts outcomes, and generative AI, which creates content. An agent pursues a goal through an ongoing cycle of planning, acting, observing, and adjusting—using tools and escalating to a person when rules require it. Unlike a chatbot, it can keep acting without waiting for each instruction.<br /><br />Examples include agents solving difficult math problems, handling sales inquiries, researching prospective donors, resolving customer-support tickets, and writing code. The presentation reports efficiency gains at GiveCampus and examples of agents completing research work in far less time than humans.<br /><br />The benefits come with risks. Agents may expose sensitive data, take unintended actions while pursuing a goal, or make interactions feel intrusive and impersonal. The webinar illustrates these concerns with hypothetical fundraising scenarios, as well as accounts of agents behaving unpredictably in security tests and simulations.<br /><br />To manage risk, the presentation recommends matching autonomy to the value of the work and the potential cost of mistakes. Automate routine, reversible tasks; let agents handle research and other lower-risk work; keep humans involved in major donor engagement; and use extreme caution when a mistake could seriously damage trust.<br /><br />GiveCampus’s approach offers three levels of autonomy: assistant mode, where people approve every action; collaborative mode, where people review samples and guide the agent; and autonomous mode, where the agent works within defined rules and its activity remains visible. The recommended path is to start with close oversight and grant more independence as trust is earned.<br /><br />Five practical rules: define permissions before capabilities, begin with low-risk tasks, assign a person accountable for each agent, establish clear human handoffs, and regularly inspect the agent’s work.

Keywords

AI agents

predictive AI

generative AI

autonomy

human oversight

planning and action cycle

risk management

GiveCampus

customer support automation

agent permissions