Concierge first, code second: why I spent weeks automating nothing

Someone once asked me how long it took to code my AI prospecting agent. The real answer fits in one sentence: I first spent weeks coding nothing at all.

What I did before the first line of automation

Before writing anything, I did the work by hand. For a long time. I sourced my own prospects, with no tooling. I sorted them by hand. I wrote every message myself. And at every step, I noted what was purely mechanical, repeatable without thought, and what required real judgment.

Only once that work was done did I automate. And only the mechanical part.

Why this order, and not the reverse

Most AI projects I see fail do the opposite: they automate a process that nobody in the company has ever actually run end to end by hand. Code comes first, understanding the process comes later. The outcome is almost always the same: the existing disorder gets industrialized, simply faster and at greater scale.

A process you cannot execute yourself is a process you cannot properly automate either. You can only hide it behind an AI that looks confident, without that confidence resting on anything solid.

Doing it by hand is not wasted time

It is the project itself, not a preliminary step you sit through before the “real” work begins. Automation is what remains once you have understood exactly where the value is, and where the human judgment sits that no script can replace.

The first question I ask on an engagement

This is exactly my first step with a client, before AI even enters the conversation: not “which AI should we plug in,” but “show me how you actually do this today, step by step.” Everything else, tool choices, the scope of automation, the level of human validation required, follows directly from that. This mapping of the real process, done by hand, is the core of any serious diagnostic before an AI project begins.