The rule I use to decide what I automate
AI without control is worthless. You can automate everything, but you have to know what to automate and in what order.
Here’s the rule I follow:
I automate where the output is the execution of a criterion I already have. I don’t automate where the output is the criterion.
It sounds obvious until you try to apply it, because from the outside the two tasks look identical. Writing a post is writing a post. But deciding which angle nobody has defended yet, and writing the 1,400 words that develop it, are two different operations. One creates information that didn’t exist. The other formats it.
AI is extraordinary at formatting. It’s mediocre at creating information that didn’t exist, because its raw material is whatever is already written somewhere. When I hand it the part that creates information, it hands me back the average of the internet. Which is exactly what I don’t need.
How I decide, in three questions
For any task:
Is there a written rule that covers 80% of cases? If the rule only lives in my head, or changes every time I apply it, it isn’t a rule. It’s criterion, and it isn’t ready to leave me.
Is the value in the output, or in what I learn producing it? The value of a follow-up is in the output. The value of a customer conversation is in what I hear. I automate the first. If I touch the second, I lose my nervous system.
Would I catch the mistake before publishing it? Badly written copy, I see. Wrong segmentation, I see six months later, when the half-year is already spent. The longer an error takes to surface, the more expensive it is to delegate.
Three greens: I automate without guilt. One red: not yet.
The order
Order matters as much as selection, and it’s the part I got wrong at the start.
First, execution that already has a rule. It’s boring and it pays back immediately: prep, transcripts, follow-ups, CRM hygiene, reporting, tier-1 support, content formats and distribution. None of it teaches me anything, and all of it was eating my hours.
Second, distilling criterion into rules. Turning what I do on instinct into something written and verifiable. This phase doesn’t automate anything by itself. It manufactures the material that makes automation possible later.
Third, automating what’s been distilled. Then going back to the frontier where I still know nothing.
I started at step three without doing step two. I produced a lot and nothing happened.
Where my line sits today
AreaAutomatedMineContentDrafting, formats, calendar, distributionWhich category I’m defending, which phrases I want repeated, which proprietary data I useGEOQuery monitoring across LLMs, HTML, schema, internal linkingWhich queries, and what answer I want citedCustomer successTier-1, health checks, onboarding, reportsRenewals, large accounts, any customer who sounds annoyedSalesPrep, transcription, follow-up, CRMThe conversationProductTests, docs, triageRoadmap and architecture
The right-hand column isn’t there out of ego. It’s there because that output doesn’t have a rule yet.
What I almost automated
Demos were the largest block of hours in my week, and I’m the bottleneck of the company. Automating them looked like the obvious call. I got as far as designing how.
I stopped when I ran them through my own three questions.
I don’t have the playbook written. I’m selling a category I’m still defining, to people who until that call believed their GA4 was fine. I still don’t know which objection kills the deal, or which piece of data opens someone’s eyes and which one sounds like marketing.
Those conversations aren’t a sales channel. They’re my market research.
If I put them through an agent, the learning ends up in a log I don’t read. I’d have execution solved and criterion unbuilt — the worst place to be.
What I did automate was everything around the demo. That’s where 80% of the hours were.
The line moves, but you have to push it
Today’s criterion is tomorrow’s execution — but only if I distil it into an explicit rule.
My content agent runs on 24 verifiable rules. Every one of them used to be a decision I made on instinct, on every piece. Writing them down took an afternoon. That afternoon turned a criterion task into an execution task permanently.
This is the work that shows up in no AI keynote. It isn’t “I handle the important conversations.” It’s: I handle them while distilling what makes them work, and the moment the rule exists, I automate the distilled part and move to the next frontier.
If I run 200 demos without extracting patterns, at demo 200 I’m exactly where I was at demo 20. Just more tired.
The three mistakes
Automating criterion too early. The symptom is steady output with no effect: content that ships and produces no lead, no citation in an LLM, no reply. It’s the most dangerous failure because volume masks the absence of signal for months.
Not automating execution, out of attachment. “I do this myself because the personal touch matters.” Sometimes that’s true. In my case it was usually identity dressed up as criterion.
Never distilling. The most expensive and the most invisible, because nothing breaks. I just keep doing the criterion work forever, and the ceiling of the company quietly becomes my calendar.
What I’ve taken from this
Using AI isn’t an advantage. We all use it, with the same models and similar prompts. Generic AI content about cookieless analytics already exists by the million. It doesn’t rank, doesn’t get cited and doesn’t sell.
The advantage is in the two inputs AI can’t manufacture: proprietary data nobody else has, and criterion accumulated by having been there. Without those two, a stack of agents produces interchangeable material at high speed.
So for me, AI-native doesn’t mean less human in the loop. It means being in fewer places, and the right ones.
A few weeks ago I opened a file called criterio.md. Every time I make a decision I used to make on instinct, I write it down as a rule.
It’s the only thing that lets me move the line. Without it, the line stays where it is — and so does the ceiling.