The Agent That Can Sign a Quote: What to Require Before AI Touches Revenue
About this session
Enterprises are racing to give AI agents real power: scoring leads, generating quotes, applying discounts, triggering fulfillment. Every one of those actions has a dollar value attached, yet most teams plan to govern them like chatbots: a system prompt and hope. Security researchers call this the non-human identity problem. Machine identities already outnumber humans roughly 40 to 1, and agents are the fastest growing, least governed slice. I lead the Salesforce architecture behind lead to cash automation at Black Duck, a software supply chain security company, and I'm designing the governance gate our future agents must pass before any of them touches revenue. This talk shares that blueprint: classifying agent autonomy by financial blast radius, designing a scoped and auditable identity per agent, evaluating agents whose failures cost money, and defining where human approval stays mandatory.
Speaker
Key takeaways
- Why agents that act on revenue are a different risk class than agents that draft or recommend, and why prompt level guardrails can't govern them.
- A practical model for classifying agent autonomy by financial and contractual blast radius, not task complexity.
- What a scoped, auditable non-human identity plus a runtime authorization gateway looks like as a pre deployment requirement, and where human approval stays mandatory.