Chapter 33: Workflows, Not Roles

Maya ran the experiment on a client she liked, which is how she found out she'd been an idiot about it.

It was month nine. The team was three people now, plus her, and the bottleneck had moved. Her first lead handled delivery. The VA ran the calendar. And Maya still wrote every word that went out under a client's name, because writing was the one thing she was certain only she could do. She was proud of it. That pride was the problem, and it had two opposite shapes.

The first shape was avoidance dressed as taste. She wouldn't touch AI for client work at all, because the people flooding her feed with bot-written carousels produced exactly the bland mush she'd built a business killing. She'd watched a competitor go all-AI and turn into wallpaper. So she folded her arms and wrote everything by hand, slower than her competitors, and told herself slow was the same as good.

Then, one tired Tuesday, she swung the other way. A client named Priya, who ran a founder-led logistics consultancy, needed a month of email copy. Maya was buried. So she fed the brief into the model, told it the audience and the offer, and let it run. It wrote thirty emails in four minutes. They were grammatical. They were fast. She skimmed them, told herself they were "90% there," and scheduled the whole month.

The emails were garbage. Confident, fluent, average garbage. They sounded like every other logistics consultant on the internet, the precise bell-curve voice Priya had hired Maya to escape. They went out. Priya read the third one in her own inbox, the one promising "tailored solutions that drive results," and called Maya, not angry, which was worse. Disappointed. "This doesn't sound like me," she said. "This sounds like the thing I was running from." Maya pulled the rest of the queue and rewrote it by hand that night, free, and lay awake doing the math on how close she'd come to torching a client she actually liked over four minutes of saved labor.

Both moves were the same move. Ignoring AI out of pride and surrendering to it out of exhaustion were two ways of refusing to do the actual work, which was to decide, task by task, where her judgment had to live and where it didn't. She'd been treating AI as a yes-or-no. The council treated it as a where.

She closed Priya's tab and went back to the operators, the ones who ran teams ten times her size and had clearly already paid for this lesson.

The first thing they did was take the urgency out of her hands. This was not optional, and it was not later. The corpus said it in a chorus too broad to pin on one person:

AI will never be worse than it is right now. And if you assume any rate of improvement over any reasonable time period, learning how to use AI should become your number one priority, your number two priority, number three priority, and your number 10 priority.

Then Hormozi reframed the entire question she'd been asking. She'd been asking "should I use AI for writing," a role question, a yes-or-no. Wrong unit. Don't think in roles:

You have to stop thinking in role-based thinking and start thinking in workflow-based thinking. For every hire you're considering, write down the four to six things or eight things or ten things that that person actually does, and then ask whether each of those activities could live inside of a workflow instead of head count.

That cracked it open. "Writer" was not a task. It was eight tasks wearing a trench coat. Researching the client's market. Pulling their actual customer language. Drafting structure. Drafting sentences. Checking the claim against what the founder really believed. Cutting the lines that lied. Some of those a model could do better than her at 2 a.m. for free. Exactly one of them was the reason Priya paid her.

The mistake with Priya was diagnosable now. She hadn't trained the AI. She'd judged a cold model against a craft she'd refined for a decade, which Hormozi calls a rigged fight:

Business owners will install a half-built function of AI into their business, compare it to something that's been optimized over the last 6 years or 10 years or 20 years, and say, "See, it doesn't work as well." You have to do an apples-to-apples comparison. How long did you work on training and putting all the SOPs together? Well, then you have to apply that same methodology to the AI.

She'd never given the model her SOPs, her samples, her rules. She'd hired a stranger off the street and blamed it for not sounding like a ten-year employee. So she built what Martell calls a master prompt, and what Maya thought of as onboarding:

You need to create your master prompt. A master prompt is essentially having a document of all of your preferences. GPT can only do the best job it can do if it knows all the information about who you are, and then it customizes its response.

She fed it her positioning method, twelve of her best rewrites, her list of banned phrases, the wince test in full. And she stopped letting it flatter her. Sabrina Ramonov's line became a rule taped to the work:

Most people use AI as a faster search engine. They ask for a generic marketing plan and AI gives them generic fluff. That is glazing, not coaching. A real co-founder tells you what you need to hear. They find the flaws in your logic and the leaks in your funnel. This prompt stops AI from being a yes man and turns AI into a strategic sparring partner that demands data and challenges your assumptions.

The model went from a yes-man to a sparring partner that told her which of her own headlines were lazy. That alone was worth the weekend.

Then came the fork she actually had to stand on, because three credible camps wanted three different businesses from her.

Camp one, Dan Martell, drew the exact line: most of the work, then stop.

You need to get your work to a point where AI is doing 92% of it. The 8% is the artist part. The 8% is the you part. But everything else AI can do for you.

Camp two wanted her to go further and turn the whole thing into a machine. Chris Williamson, relaying YMH Studios, described the dream of a business with the human nearly removed:

He's using Claude code and Flux and ElevenLabs, and basically she can chat 24 hours a day with all of her subscribers. She doesn't sleep. There's nothing to film. No one's typing. Claude code writes every message, Flux generates every photo, and ElevenLabs generates her voice.

Seductive at 2 a.m. with a full pipeline. Throughput without bodies. And it was the exact road her wallpaper competitor had driven straight off.

Camp three, Kallaway, pulled the opposite way: don't hide the human, sell it. Make the handmade the whole pitch.

I think we're going to see an anti-AI shift, we're going to see a shift to human-made. It's like the Etsy version of content. You're going to have handcrafted human only.

Three roads. The deciding variable was not which sounded smartest. It was what her specific moat was made of. Maya didn't sell volume. She sold the line a founder secretly didn't believe, found by a human who'd sat across from forty founders and watched them flinch. Her moat was taste and lived experience, the two things the machine camp throws away and the Etsy camp can't scale. So she rejected pure automation, because it would sand her down to the average voice that had nearly cost her Priya. And she rejected pure handmade, because doing the 92% by hand was just the slow pride she'd already proven was a stall.

She took Martell's line and made it load-bearing. Taki Moore had named why the 8% could only be hers:

That's why I just believe that fundamentals are so important now, because when people go down the road and they just go AI content but they got no taste, it's like they're just putting out crap, and then it actually does a negative thing to your brand.

The wince test was the 8%. A model could draft thirty competent emails. It could not read a founder's homepage out loud and feel the small flinch at the claim that overreached. That flinch was lived experience, and it was the entire reason anyone paid her instead of paying nothing.

The toll was specific and it stung. She had to hand the part she was proudest of, the drafting, to the machine, and admit the artistry she'd been protecting lived somewhere narrower than her whole job. For a decade "I write it all myself" had been her identity. She traded it for "I judge it all myself," and that trade felt, for a week, like a small demotion.

She rebuilt Priya's month the new way. The trained model drafted the thirty emails, now in Priya's actual cadence because it had her samples and her rules. Then Maya did the last mile by hand: ran the wince test on every one, killed the four lines that lied, rewrote the two openings that went soft, and added the one lived-experience story the model had no way to know. Ninety minutes, not three days. Priya read the new third email and wrote back one word: "Finally."

The capacity jump was not subtle. Work that had eaten her week now took an afternoon of judgment. She put the freed days where only she could go, the diagnosing calls and the taste, and let the boring 92% run without her. The business that had needed all of her to produce now needed only the irreplaceable part of her.

Which opened the next wall, the one Part V had been driving at the whole time. She now had AI doing the 92% and three humans doing real work, and a finite number of her own hours left. So the question stopped being can someone or something else do this and became sharper and more uncomfortable: of everything left on her plate, what was actually worth her own hands, and what was she keeping out of habit. She had bought back the week. Now she had to decide what an hour of her was even for.

My verdict. You are hiding from AI in one of two coward's crouches, and they look like opposites. Either you won't touch it, calling your slowness craft, or you've handed it everything and called the average output "good enough." Both skip the only real work, which is deciding task by task where your judgment has to live. The machine drafts. You decide. The four-word version: direct AI, don't compete. Don't race the model at the thing it does for free. Stand on the 8% it can never reach, the taste and the scars and the story only you lived, and sell that. The rest is just typing, and typing was never the part they were paying for.