The solo operator's AI stack

Most solo operators do not have an AI problem. They have a sprawl problem. Two assistant subscriptions because a friend recommended one and a newsletter recommended the other, a half-built automation that fires into a spreadsheet nobody opens, three note apps, and a monthly card statement that nobody has read line by line since spring. The work still gets done, but the tools are not doing much of it, and the cost is real.

A stack is not a list of tools you like. It is four decisions, made once, that you can defend and undo. This guide is for a one-person or two-person business that already pays for at least one AI subscription and wants the whole arrangement to make sense. It is not for teams with a procurement process, and it is not for anyone building a product on top of model APIs. Prices below are the list prices shown on each vendor’s own page on 4 September 2026, and vendors change them without much notice.

Four layers, and nothing else counts as a layer

The stack has an interface, a worker, an automation, and a record.

The interface is the one general assistant you open every day without thinking. It answers questions, drafts the awkward email, reads the contract you were sent, and tells you what you are missing. Claude, ChatGPT and Gemini all sit here. So does Perplexity, if what you do all day is closer to looking things up than to drafting.

The worker is the tool that produces the actual artefact when the artefact is specialised. Code is the clear case: the output is a changed repository rather than a block of text you then have to put somewhere, and tools built for that job, Cursor among them, are shaped differently from a chat window. For most non-technical operators, the interface and the worker collapse into one tool, and that is fine. Do not add a worker layer to justify a subscription.

Automation is the thing that runs while you sleep: a form submission becomes an invoice, an inbox label becomes a task, a weekly report assembles itself. The record is where outputs land and stay findable in six months. A Notion database, a Google Sheet, a folder on your own disk. If your outputs live in chat history, you have no record layer, and the first time you need last quarter’s numbers you will discover this the hard way.

One tool per layer, and you say which one out loud

The rule that saves the most money is boring: one tool per layer, named. Not “I mostly use Claude but sometimes ChatGPT”. Pick, then cancel the other one. Two general assistants at $20 each is $480 a year for a difference you will not be able to describe [1][2].

The reason is not only cost. A second tool in the same layer splits your context. Half your prompts, uploads and saved instructions live in one account and half in the other, so neither becomes good at knowing your business. The tool you use for everything gets better because you feed it everything.

There is a second cost, harder to see on a statement. Every tool in a layer is a place you have to look. Two assistants means two histories to search when you half-remember a draft, two sets of uploaded files, two settings pages to check when a vendor changes a default. The tax is a few minutes at a time, and it lands on the days you are busiest.

The exception worth allowing is a free second opinion. Keeping a free account on a rival assistant costs nothing and is useful when the paid one is confidently wrong. That is a habit, not a layer.

What $0, about $40 and about $100 a month actually buys

At $0, every layer has a real free tier. A free assistant account, files kept in folders you already own or a free Notion workspace, which Notion lists at $0 per member per month [3], and automation on a free plan: Zapier’s Free plan is $0 with 100 tasks per month, and Make’s Free plan is $0 with 1,000 credits per month, where each module action in a scenario counts as one credit [4][5]. That gap matters more than it looks. 100 tasks is one low-volume workflow. 1,000 credits is a few of them, which is enough to find out whether you will use the layer at all before you pay for it.

At about $40, you are buying the paid assistant and a record you trust. Claude Pro is $20 per month, or $17 per month with the annual discount, billed at $200 up front; ChatGPT Plus is $20 per month [1][2]. Notion lists Plus at $10 per member per month and Business at $20 [3]. Make Core is $12 per month for 10,000 credits [5]. An assistant, Notion Plus and Make Core lands at about $42 a month, and it is the shape most solo operators should stop at.

Check one thing before you take the cheaper Notion tier. Notion AI is a trial on the Free and Plus plans, and the agent that chats, generates, autofills and translates is included on Business [3]. If you want the record layer to do work rather than only hold it, that is the $10 per member the decision turns on.

At about $100 there are two honest shapes, and they are not equivalent. The first spends almost everything on the interface: Claude Max starts at $100 per month, with a choice of 5x or 20x more usage than Pro, and free tiers underneath it for record and automation [1]. The second spreads the money: a $20 assistant, Notion Business at $20 per member, and Zapier Professional from $19.99 per month comes to about $60, leaving headroom for a worker tool or a busier automation plan [1][3][4]. Take the first if you hit usage limits weekly. Take the second if you keep hitting limits on everything else.

One note on automation pricing, because it is the layer that surprises people. n8n Cloud Starter is €20 per month billed annually for 2,500 workflow executions, and Pro is €50 for 10,000 [6]. There is also a standard self-hosted Community edition on GitHub, governed by n8n’s fair-code licence [6]. Self-hosting moves the cost off the invoice rather than out of existence: you still own the server, the upgrades, and the evening a workflow stops firing without telling you. Choose n8n self-hosted if you genuinely enjoy running things. Choose Zapier if you want the connector to exist already. Choose Make if you want visual branching and per-action pricing you can predict.

What you keep out of the stack

The default question is not “is this tool secure”. It is “what does the policy on this specific plan say happens to what I paste”. On consumer Claude plans, Anthropic uses your chats and coding sessions to improve Claude if you choose to allow it, and may also use conversations flagged for safety review to enforce its Usage Policy; Incognito chats are not used to improve Claude even when model improvement is switched on in Privacy Settings [7]. For ChatGPT, the setting is named “Improve the model for everyone”, you turn it off under Settings then Data Controls, and your conversations stay in your chat history either way; Team, Enterprise and Edu plans carry additional data controls [8]. Read the current version of both before you decide, because these terms have moved more than once.

That gives you a working line. Your own drafts, your own numbers, your own strategy: paste freely. Material that belongs to a client, a candidate, or a patient goes in only if your contract with them and the vendor’s policy on your plan both allow it. If you have to reason for more than a minute about whether a document qualifies, that is your answer.

Two things stay out regardless of policy. Credentials of any kind, because an assistant conversation is not a password manager. And anything you would not want to see quoted back to you in a dispute, since a record of the exchange exists on the vendor’s side whatever the training setting says.

Swapping a layer without rebuilding the stack

Layers stay swappable when the connective tissue does not live inside any one of them. Three habits do most of the work.

Keep your prompts in the record layer, not in the assistant. A page of saved instructions in Notion or a text file on your disk moves to a new assistant in a copy-paste. The same instructions saved as custom instructions inside a vendor’s account do not move at all.

Let automation write to the record, not from assistant to assistant. If a Zapier or Make scenario ends by writing a row, changing the assistant that produced the text means editing one step. If the scenario chains two vendors together, you rebuild it.

Name the layer, not the vendor, in anything you write down. A note that says “assistant drafts the reply, automation files it, record holds the thread” survives a switch. A note that says “Claude drafts it and Zapier files it” has to be rewritten every time you change your mind, and after two rewrites nobody keeps it current.

Test a replacement on live work for two weeks before cancelling anything. Run the same real tasks through both, and keep the new one only if you can say which specific job it did better. Monthly billing exists for this. The annual discount on Claude Pro is real money, $200 up front against $240 paid monthly, but only take it once you have stopped shopping [1].

What still goes wrong

Prices and limits move. Every figure here was read from the vendor’s own page on 4 September 2026, and any of them can change next quarter, along with what a plan includes at that price. Plan pages also default to a billing mode, so the number you see first is not always the number you pay. Treat all of it as a snapshot to re-check, not a standing fact.

The bigger failure is quieter. A stack this small assumes you are the only operator and the only reviewer. Nothing in it catches an assistant output that is fluent, plausible and wrong before it reaches a client, because there is no second person in the loop. Automation makes that worse by removing the moment where you would have read the thing. Automate the delivery of work you have already checked; do not automate the checking.

And a tidy stack can quietly become an excuse. Four layers, one tool each, a spreadsheet of prices, and no more revenue than before. The stack is overhead until something in it is doing a job you used to do by hand. If you cannot name that job for a given layer, cancel the layer.