AI isn’t a productivity tool. It’s an instrument. Five jobs it does well. Four it shouldn’t touch. Solo founders who can name the difference get more out of AI than the ones running on the “use AI for everything” advice still circulating on every platform.
The “everything” advice doesn’t fail because AI is bad. It fails because the line isn’t drawn. Without a line, you let AI into the four jobs where it produces output that looks competent and isn’t. The work feels productive. The result is thinner. Six months later your archive sounds like everyone else’s archive. The line below is short on purpose. Five in. Four out. Reasoning attached.
Five things AI is actually good at
First drafts. Not finished writing. Raw text you’ll cut, restructure, and rewrite, but text that exists rather than not existing. The blank page is the most expensive part of any writing task, and AI is good at filling it. The trick is treating the first draft as scaffolding, not content. Cut 50 to 70 percent. Rewrite the rest. The piece that ships is yours.
Structured extraction. Pulling specific data out of messy text. The relevant numbers from a long email thread. The action items from a 30-minute transcript. The list of company names mentioned in a research note. AI is reliably good at this kind of work because the source contains the answer and the job is just to surface it. The error rate is low and the verification is fast.
Reformatting. Converting between formats while preserving substance. Bullet points to prose. A paragraph to a tweet. An outline to a brief. AI handles these transformations cleanly because nothing new is being created. The substance is the same. The shape changes.
Volume and variation. Generating ten versions of a subject line. Five different intro paragraphs. Twenty possible headlines for the same article. Solo founders are usually capacity-constrained on these tasks – the time cost of generating ten variations by hand is higher than the time cost of picking among ten variations AI generated. The picking is the work. The generating is the scaffolding.
Iteration partner. A cheap rubber duck. Talk through a half-formed argument. Ask AI to argue against your position. Ask for three weaknesses in a plan. Ask what a smart person would push back on. The output won’t be the answer. It will be enough friction to surface the answer you already had access to but hadn’t found. Ethan Mollick’s Co-Intelligence is the cleanest articulation of this stance: AI works best as a collaborator in the thinking, not a substitute for it.
These five share a property. The work is real. The judgment about whether the work is good still belongs to you.
Four jobs AI shouldn’t do
Source claims. AI fabricates citations. Confidently. With plausible-looking URLs and made-up author names. This isn’t a bug that’s getting fixed in the next model. It’s a property of how language models work. Citing something AI produced as a source is how you publish wrong information under your own name. The fix is to use AI for the first draft and verify every citation against the actual primary source before anything ships. A 2025 MIT Media Lab paper, Your Brain on ChatGPT, found measurable reductions in critical engagement when participants relied on AI for the verification step. The lesson is operational: keep the verification in your hands.
Decide what matters. Editorial judgment isn’t a skill AI has. It’s a synthesis of taste, context, and stakes that requires you to know your reader, your archive, and what you’re trying to be known for. AI doesn’t know any of these. The output it produces about what matters is the average of what mattered for someone else. Asking AI “what should this article be about” is asking the average of the internet to set your strategy. The output will be average accordingly.
Edit your voice. This is the one most solo founders miss because it looks like polish. You write a paragraph. You ask AI to “improve” it or “make it more readable.” AI smooths the rough edges – which are usually the parts that made the prose recognisable as yours. After three months of this, your archive sounds like everyone else’s archive. The distinctive phrasing is gone. The competent generic prose has replaced it. The audience can’t tell what was you and what wasn’t. Neither can you, if it’s been long enough.
Choose the angle. The decision about which idea becomes the next piece is upstream of writing the piece. It contains the strategic question of what you’re for and what your archive needs next. AI doesn’t have a strategy. It has access to your prompts and the average of what’s been written about your topic. Asking it to pick the angle outsources the most distinctive decision you make as a publisher. The output is a piece that could have been written by anyone, on a topic everyone has covered, from an angle that doesn’t position you against anything.
These four share a property too. The work looks done. The thinking it skipped will surface as drift in your archive across the next six months.
Where the line actually runs
The line isn’t between “AI tasks” and “human tasks.” It runs between work where the substance is fixed and work where the substance is judgment.
When the substance is fixed – the meaning of a sentence, the data in an email, the structure of an outline – AI is a fast, reliable instrument. The work was always going to come out the same shape. Letting AI handle the mechanical part frees the human part for what only the human does.
When the substance is judgment – what matters, what’s distinctive, what positions you against alternatives, what your voice sounds like – AI doesn’t have access to the variables that produce a real answer. It produces the average. The average is the failure mode. Solo founders who treat AI as a judgment engine produce average work without noticing.
The Solo Operating System frame from the cluster A opener maps here directly. AI does well in the Work part of the four-part system. It’s bad at Decisions. It’s bad at Review. It’s neutral at Inputs – it can help capture, but the sorting is still yours. The line is structural, not stylistic.
Why the line moves you can’t see
The reason this matters more for solo founders than for teams: a team has friction. Other people disagree. Editors push back. The judgment work happens inside the team’s interactions and the AI output gets reviewed by someone whose taste is different from the AI’s. A solo founder doesn’t have that friction. The AI’s judgment, accepted silently, becomes the only judgment in the room.
The drift is invisible because each individual piece looks fine. The aggregate is the failure. You publish for six months. Each piece sounds like a piece. The archive starts sounding average. Search starts ranking you for terms you didn’t intend. The audience that arrives doesn’t feel addressed by anything specific because nothing specific is being said. By the time the pattern is visible, six months of work has shaped your archive in a direction you wouldn’t have chosen.
The fix is the line. Five jobs in. Four jobs out. Both numbered small enough to remember on a Tuesday afternoon when the tool is open and the deadline is close.
The piece on the five questions to ask before you put AI in your workflow goes deeper into the operational version of this line. The cluster closer on when to ignore AI covers the editorial-judgment side at the level of the individual sentence.
The rest of the AI & Workflows archive lives at the AI & Workflows category.
FAQ
What about the new models that promise better reasoning?
The line above isn’t about model capability. It’s about what kind of work has a verifiable right answer. A better model produces better first drafts and faster structured extraction. It doesn’t change whether you should outsource editorial judgment. The four out-of-bounds jobs are out of bounds because the substance is judgment, not because the current models are bad at it.
Does AI cite sources I can use directly?
Sometimes. Often not. AI-generated citations need to be checked against the actual primary source every time. A reliable workflow is: AI helps locate the rough area of the source (“there’s a 2024 study about…”), you find the actual source, you cite that source. The AI provides the lead. You confirm the lead. Citing AI’s citation directly is how factual errors get published under your name.
Is using AI for first drafts hiding something I should be doing myself?
No, if you treat the first draft as scaffolding and rewrite the rest. Yes, if you ship what AI produced with light edits. The difference is whether you keep the work of selecting, structuring, and rephrasing. The first draft is the cheapest part of writing. Keeping the rest is what makes the published piece yours.
How do I know if my voice is drifting?
Read three pieces from your archive at six-month intervals. Ask honestly whether you would still describe them as written in your voice. If the older pieces sound sharper than the newer ones, AI editing is the most likely cause. The next move is to stop running AI passes on your prose for 90 days and see what comes back.
Where does this line move as AI gets better?
The five in-bounds jobs get faster and slightly more reliable. The four out-of-bounds jobs don’t move. Editorial judgment, source claims, voice, and strategic angle don’t become AI tasks because the model improved. They’re not a model problem. They’re a category problem.
What stays yours
The work that stays yours is the work the audience came for. The angle, the judgment, the voice, the source. None of it is the part AI does well. All of it is the part that decides whether what you publish is yours or whether it’s a competent average of what someone else might have written.
Use AI for the scaffolding. Keep the work.
The Solo Founder’s AI Stack. A printable matrix of the in-bounds and out-of-bounds jobs, twelve prompt patterns that respect the line, and the retire-this-prompt checklist that keeps the stack honest. Download the kit →
