Every AI pitch this year promises the same thing: describe what you want, get back finished work. That's half true. What's actually happening, across a newsletter writer, a nonprofit consultant, and a solo software founder, is narrower than the pitch, and more useful.
The newsletter that averaged itself into nothing
Carrie Loranger writes a newsletter to 8,600 subscribers. Her first attempts at using AI to speed it up were the obvious kind: one prompt, "write me a newsletter about X." The output opened with "in today's evolving creator economy landscape." She deleted it and started over, most Sundays, for months.
The problem wasn't the model. A newsletter isn't one job, it's roughly eight: pick a topic, research it, outline it, write a headline, write the body, write a CTA, brief an image, repurpose it into other formats. Collapse all eight into a single prompt and the model does the only thing it can do with no real instructions for any of them. It averages. It returns the statistical middle of everything it was trained on, which is exactly what generic AI writing sounds like.
Her fix was mechanical, not magical. She broke the newsletter into separate jobs, gave each one its own rules and a banned-words list, and ran them in sequence, research to one tool, drafting to another with her voice already trained in. Four hours dropped to one. She still writes the opening, the personal story, and the opinion herself, because the model won't offer one unless asked directly.
The methodology that had to get written down first
The same pattern shows up wherever someone tries to turn static expertise, a book, a course, a consulting framework, into something interactive. Dana Snyder runs a one-person nonprofit consultancy with no technical background. Over six months she turned her own consulting methodology into a platform that walks nonprofit organizations through building a monthly giving program, built specifically to reach the roughly 93% of U.S. nonprofits too small to ever hire a human consultant.
Vishal Sachdev, a business school educator, did something structurally similar in a single afternoon: fourteen interactive lessons, a fictional startup with 26 files of deliberately messy business data, and a "conductor" file that tracks each student's progress. Six AI agents produced 53 files and roughly 50,000 words in about two hours. His own account of it: "I didn't write 50,000 words. I directed them. Every design decision came from years of teaching. The AI handled the implementation."
Neither was blocked by money or by not knowing how to code. Both were blocked, until recently, by the same thing everyone sitting on a framework is blocked by: whether they could describe what they know precisely enough for something else to execute it. That's a specification problem, not a technical one.
The job he gave back to himself
Maor Shlomo built Base44, a platform that lets nontechnical users build software by describing it to a chatbot, largely by himself. It generated close to $1.5 million in revenue within a month of launching, and Wix acquired it for $80 million four months later.
He automated almost everything a normal team would split across a product manager, a QA engineer, and a developer, including an agent that turned his shipped code into marketing posts once it was tuned to sound like him. Then he built a customer support bot, and shut it down after two weeks. Not because it failed. Answering support tickets himself was how he stayed close to what was actually happening in his product, and he needed that friction more than he needed the hours back.
Snyder made a version of the same call from the other direction. She's now building agents for tasks she says she'd never have hired a human for, but she still handles client-facing consulting work herself rather than routing it through her platform. In both cases, the decision wasn't about what could be automated. Nearly everything can be now. It was about which tasks were quietly generating information neither of them could get any other way.
"I didn't write 50,000 words. I directed them. Every design decision came from years of teaching."
Line these three up and the pattern repeats. AI has gotten very good at producing structure: a newsletter's scaffolding, a course's lesson architecture, a support bot's canned replies. It hasn't removed the need for judgment, knowing which eight jobs a newsletter actually is, knowing your own methodology well enough to specify it, knowing which task to keep doing badly by hand because the friction is the point. The gap opening up isn't between people who use AI and people who don't. It's between people who hand AI one vague job and get back something that averages to nothing, and people who've done the harder work of breaking their own expertise into pieces clear enough to direct.