Running an Agency Solo: What I Automated First, and What I Still Do by Hand
Six months into running the agency mostly with AI tooling instead of a team, the order I automated things in turned out to matter more than what I automated.
When people ask how I run an AI/automation agency without a team, the honest answer is: badly, for the first month, while I figured out which things were actually worth automating. The order mattered more than I expected. Automate the wrong thing first and you burn a week building plumbing nobody needed.
First: anything with zero judgment in it
The first things I automated had no decision inside them at all — pulling a client's site analytics into one weekly digest, generating an invoice template from a completed project, converting a call transcript into a structured brief. None of these need me. They needed doing, reliably, at the same time, without me remembering to do them. That's the lowest-risk automation there is: nothing gets worse if it's slightly wrong, because there's no judgment call to get wrong.
Second: the first draft of anything, never the final version
Proposal drafts, first-pass code, initial project scoping documents — I let AI tools produce a first version of almost everything now. But I never ship the first version untouched. The rule I settled on: automation is allowed to produce a draft, a human (me) is required to produce the thing that goes out the door. This isn't caution for its own sake — it's that clients can tell the difference between something written for them and something written at them, and the gap is exactly the part that doesn't automate well yet.
What I still do entirely by hand
Three things, deliberately:
- Scoping the first call with a new client. This is where I learn what they actually need versus what they asked for, and that gap is the whole value of the call. Automating it would remove the only part that matters.
- Anything involving a number a client will hold me to. Pricing, timelines, commitments. I want to have personally reasoned through every number that becomes a promise.
- Deciding what not to build. Every agency conversation eventually has a moment where the right answer is "don't build this, here's why." No tool I've used says that on its own — they're all biased toward producing something.
The mistake I made early
I automated a client-communication step before I automated the invoicing. It felt backwards in hindsight — invoicing was pure friction with zero judgment, client communication was almost entirely judgment. I'd optimized for what was technically easy to automate, not for what was actually costing me the most time. Now the first question I ask about anything I'm about to automate is: how much judgment is actually in this step? If the honest answer is "none," it goes first. If the honest answer is "most of it," it probably shouldn't be automated at all yet — it should be made faster to do by hand instead.
