Agent permissions

What is meant by agent permissions?

Agent permissions are the rules that set which systems, data and actions an AI agent may use, and which actions need a person's approval first. They work like user access rights for software that acts on its own. Claude Code, for example, uses rules that allow an action, ask a person first or deny it. OWASP's AI security guidance names excessive permissions as a root cause of damaging agent actions.

Why it matters to a revenue leader

An agent that runs with a seller's full CRM access can change anything that seller can, including amounts and close dates on deals in the forecast. OWASP recommends giving each agent the narrowest access its job needs and requiring human approval for high-impact actions. OpenAI's agent guide suggests rating each tool by risk, weighing factors such as write access and financial impact. Permissions decide whether an agent's mistake ends up as a bad draft or as a change to the forecast.

A PE-backed fire protection company gives its renewal agent read access to agreements and service history and lets it write only draft quotes. In its first quarter the agent drafts 340 renewal quotes, and none reach a customer without the account manager's approval. When it misreads a contract term on 6 drafts, the errors stay in draft and get fixed.