Hand-Onboarding Is Research, Not Inefficiency
Graham's core argument is that recruiting users one by one and delighting them absurdly is how you learn what your product actually needs to do. When agents handle your onboarding before you have done it yourself fifty times, you skip the curriculum. The manual labor of early customer service is market research in disguise — you learn the exact friction, the exact language, the exact moment someone trusts you enough to stay. That knowledge is what you encode into your workflows later.
Trust Cannot Be Automated Before It Is Earned
An agent can execute a perfectly designed onboarding sequence and still fail to earn the customer's belief that you will be there when something breaks. The founder who shows up personally in the early days is building a different asset than a smooth product — they are building the reputation that makes distribution possible. Software has become cheap to create, which means the moat is who trusts you and why. Trust is earned at unscalable moments before it can be systematized at scale.
Judgment Comes Before Autonomy
Graham's insight maps cleanly to a design principle for AI-native companies: human judgment must precede agent autonomy. You cannot delegate to agents what you have not yet understood yourself. Until a founder has sold the product, onboarded the customer, and felt the friction personally, they lack the ground truth to set goals, define boundaries, and write the evals that make agentic work reliable. The unscalable phase is not opposed to automation — it is the prerequisite for automation that actually works.
