The Unit Economics of AI Agents

An AI agent is only good when the problem it solves is worth more than the cost of solving it. In this talk, I’ll explain how to calculate the Cost per Verified Outcome, and I’ll show how to measure the actual costs for a real agent built on Amazon Bedrock AgentCore. We’ll trace one workflow from invocation to accepted result and account for model inference, AgentCore runtime, gateway and tool calls, memory usage, human review, retries, and failure remediation. You’ll leave this talk with a practical method for deciding whether an agentic product is economically viable before you put it in production, and before you lose money.