AI Agents are the new SaaS

Summary

Key Points

• —Agents fundamentally differ from traditional SaaS by selling the completion of a job or work, rather than just providing a tool for a team to use. 1:55
• —To find a viable agent idea, focus on workflows where people are already paying for the work, such as an employee or agency, and where the buyer feels a significant loss from missed opportunities. 4:26
• —A good agent workflow is frequent, has a clear finish line, integrates with existing software, involves learnable but annoying edge cases, and addresses a pain point where the buyer feels a tangible loss. 5:36
• —Before building, shadow a human performing the target job 10-20 times to gain deep insights into the real workflow, edge cases, and decision-making processes. 7:33
• —Design the agent with seven key parts: what triggers it, the context it needs, tools it can use, what it's allowed to do, when it needs approval, when to escalate to a human, and what defines success. 9:07
• —Leverage 'workflow teardowns' and consistent content creation to demonstrate the agent's value by contrasting the inefficiencies of the old way with the efficiency and benefits of the new agent-driven process. 10:06
• —Start by building the smallest useful agent (MUA), such as a draft-and-approve, triage, coordinator, or bounded action agent, earning autonomy gradually by starting with predictable workflows. 11:25
• —The 'product wrapper,' including logs, approvals, controls, and analytics, is crucial for building customer trust and transforming a cool automation into a true agent-first SaaS product. 12:52
• —Create an 'eval set' of 50 real examples to test the agent's performance, identify mistakes, and build trust with potential customers by demonstrating its accuracy and how errors are addressed. 13:08
• —Sell initial pilots to three customers in a single niche, focusing on the outcome (e.g., answering missed calls) with a setup fee and simple monthly pricing, then productize the repeatable patterns. 16:36