A ChatGPT subscription is the right first step for most small businesses, and nothing in this note argues against it. It stops being enough at one specific place: the last mile, where the work repeats every week, the documents are private, and the output has to fit your process instead of a chat window. That gap is what private AI solves, and this note explains how to tell which side of it you are on.
What a subscription does well
For writing help, research, brainstorming, and one-off questions, a subscription is cheap, fast, and good. The quality per dollar is hard to beat and no small business should feel bad about using one. If that is 80 percent of your AI use, stop here. You do not have a last-mile problem.
Where the last mile starts
The subscription model breaks down in a predictable way, and it is worth naming the exact cracks:
- Re-entry tax. Every session starts from zero. Your context, your voice, your rules get pasted in again, or get lost.
- Private documents. Client files, contracts, and financial material should not be pasted into a consumer chat window, and most owners correctly refuse to do it. So the AI never sees the actual work.
- No repeatability. A great result on Tuesday does not guarantee the same result Thursday. For a draft that is fine. For a weekly intake summary, it is not.
- No approval boundary. The chat cannot tell the difference between preparing work and publishing it. Your process can, and the owner approval rule has to live somewhere.
- Per-seat math. Five people each re-learning the same tricks, each with their own prompts, none of it written down.
What private AI changes
Private AI means the system runs where your data already is: on a machine you own, on your own network. Three things change because of that. Your documents stay yours, inside your boundary. The workflow gets encoded once, so intake or summarization runs the same way every time. And the whole path, what the AI is allowed to touch and what stays human, gets written down instead of living in someone's head.
This is not about a smarter model. The honest position is that top cloud models are still ahead of what runs locally. The question is never which model is smartest in the abstract. It is which setup does the actual repeated work, reliably, inside your rules. On a $1,400 mini-PC we measured that question directly: 495 trials over 29 job types with certified reliability floors, all public in the delegation-bench repo. Local AI is not a belief system. It is a measured yes or no per task.
The three-question decision rule
Ask these in order:
- Does the task repeat every week with a clear definition of done? If no, subscription.
- Does it need private documents or customer data? If yes, private or hybrid.
- Does an error cost real money or trust? If yes, the setup needs a measured reliability number before anyone walks away from it.
Most businesses end up hybrid: subscription for exploration, private setup for the two or three workflows that actually repeat. That is a normal landing point, not a compromise.
What the first step costs
You do not have to guess. The Local AI Setup path starts with a $175 Setup Hour that ends in a written recommendation: local, cloud, hybrid, or stop. The starter setup is $450 beta or $650 standard. If you want the comparison made against your actual workflows instead of a general article, that hour is the cheapest way to get it.
FAQ
Is private AI safer than a subscription?
It is different, not automatically safer. Data staying on your machine removes the sending-it-outside risk. It also means you own updates, backups, and boundaries. Starter setups deliberately exclude sensitive and regulated data until there is a written scope for it.
Can local AI really do useful work?
For defined, repeated tasks, yes, and that claim is measured, not vibes. The delegation-bench results show certified floors for code, document search, summarization, and similar job classes on hardware a small business can buy.
Do I have to choose one?
No. Hybrid is the most common result: subscription for exploration, private setup for repeated work.