How the work gets done, chapter fifteen

The machine's share and the human's share

Strategy
Machine: research summaries, lists of competitors.
Person: the angle and the bet.
By hand
Copy
Machine: first drafts and variants.
Person: the words that ship.
By hand
Design
Machine: layouts, resizes, theme colours.
Person: what the eye meets first.
By hand
Email
Machine: formatting, merge tags, scheduling.
Person: who gets it, and why.
By hand
Social
Machine: cut-downs, captions to edit, posting times.
Person: the voice.
By hand
Paid
Machine: ad variants, bid rules, reports.
Person: the offer and the budget call.
By hand
Pages
Machine: builds from a pattern, alt text to check.
Person: the headline and the proof.
By hand
Analytics
Machine: pulling and cleaning the data.
Person: what the numbers mean.
By hand

Hours saved a week 14.9 hours of 21

Risk to voice and accuracy 19 of 50 (8 by hand)

My split saves 14.9 hours a week and keeps the risk at 19 of 50. Strategy stays by hand; every station a customer or a decision sees keeps a human check.

Assumptions you can change. Hours and risk are my judgement for a small team, illustrations, not data.

Strategy

Copy

Design

Social

Paid

Pages

Analytics

The eight stations: who does each, the assumed hours and risk, and what each adds to the two meters
Station	Done by	Hours the machine could take off	Risk if unchecked	Hours saved	Risk
Strategy	by hand	1	9	0	1
Copy	machine, checked	4	8	2.8	2.75
Design	machine	3	4	3	4
Email	machine, checked	2	6	1.4	2.25
Social	machine, checked	3	6	2.1	2.25
Paid	machine, checked	2	5	1.4	2
Pages	machine, checked	2	5	1.4	2
Analytics	machine, checked	4	7	2.8	2.5
Total		21	50	14.9	18.75

The pipeline: eight stations on one line of light, from strategy to analytics. Work moves along it from left to right. Hand a station to the machine and the work there moves faster; leave it unchecked and some of it slips off the line. My split, shown here, saves 14.9 hours a week of a possible 21 and keeps the risk at 19 of 50, against 8 with everything done by hand. Switch any station, add or drop a human check, and watch both meters. The hours and risk weights are my illustrations, not data.

think like making 80% of a presentation like the background design and theme colors and stuff and then human goes in and adds the text

Something I said on 8 June 2026, about splitting the work.

The idea

I think every marketing task has two parts. There's the mechanical part: pulling the data, formatting, first drafts, variants, scheduling. And there's the judgment part: the angle, the offer, the words that actually ship, and the call on what the numbers mean. A machine is good at the first part. I wouldn't trust it alone with the second.

So I automate one process at a time, time it before and after, and keep a human check anywhere a mistake would cost trust. I brief the machine the way I'd brief a person: real context, a clear format, the phrases I never want to see, and a request to mark anything it guessed. Most of this is borrowed, and that's fine. As I said the same day, "A lot of the time you really don't need to invent the wheel."

How the meters work

Each station has two numbers: the hours a week a machine could take off its mechanical part, and how much damage its work could do to the voice or the facts if it shipped with nobody looking, from 0 to 10.

By hand, a station saves nothing and carries a risk of 1, because people make mistakes too.
By machine with no check, it saves all its hours and carries its full risk.
By machine with a human check, the check costs back 30% of the hours and leaves a quarter of the extra risk.

The 30% and the quarter are my guesses, rounded on purpose. What I'd trust is the shape: the hours come from the mechanical stations, and the risk comes from the ones a customer reads.

The AI-tell check

The same linter that X-rays the cold email in chapter seven reads anything a machine drafts before a person signs it. Below is a first draft for Kiln's newsletter, the kind a machine writes when the brief is thin. Edit it and the tells drop off the list.

A first draft from the machine, for Kiln, an invented coffee roaster

What reads as machine-written

With JavaScript on, this panel fills with a machine-written draft and marks each tell as you edit it.

Where I used it

Ivanna, the assistant system I run on my own Mac: voice requests, one shared memory and scheduled jobs. It's built on this split. Machines pull, sort, draft and run the jobs; anything risky waits for my yes, and I merge every change myself. Solo
The working notes I kept from my chats while I studied digital marketing. Three of them hold up. A research brief gets everything a person would: a role, the real background, a fixed format, the phrases to avoid, and an instruction to mark guesses as guesses. The prompt is the work: once the brief carries the context, the voice and the checks, generating is the easy part, so the brief is what I keep and reuse. And checking is its own pass: I flag shaky figures while drafting and check them before anything goes out with a name on it.

What I'd do now

On 8 June I wanted to automate the course work one day at a time, using each day's lesson as the rules. I'd finish that as this Playbook's own production line. Every chapter's download, the checklists and worksheets, would come out of one documented pipeline: the machine does the formatting and the first pass, I do the angle and the final words, and I log the hours before and after. Then the hours meter above would be measured instead of guessed.