Local vs cloud AI: which setup fits your business?
Most owners treat this as a single verdict — am I a local AI business or a cloud AI business? That framing is wrong, and it's what keeps people stuck.
The honest answer: it's a split, not a verdict. Your work falls into a few tiers by sensitivity. Everyday work belongs in cloud tools, where quality is highest and setup is zero. Anything you couldn't put on a billboard belongs on machines you own. Most businesses end up running both.
Sort your work by sensitivity about 20 minutes
The setup question can't be answered in the abstract — it depends entirely on what your work contains. Every task you'd hand to AI falls into one of three tiers. Sorting them takes one sitting and a honest look at what flows through your business.
The three tiers
- Fine anywhere: public facts, marketing drafts, general questions, social posts. Nothing here would hurt you if a stranger read it — because strangers effectively do.
- Handle with care: customer lists, contact details, anything tied to a named person. Not radioactive, but it needs a tool that contractually keeps its hands off.
- Never leaves the building: financial records, contracts, health information, client matters, proprietary process — anything a leak would actually cost you.
If a task straddles two tiers, put it in the stricter one. A proposal with a client's name in it is a tier-three task with a name in it, not a tier-one task with a name in it.
Done when: you've tagged your regular AI-suitable tasks with a tier number.
Give each tier its setup about 30 minutes
Each tier has a setup that fits it, and the fit is decided by the data — not by what's trendy. This is the whole decision:
The tier-to-setup map
| Tier | Setup | Why |
|---|---|---|
| Fine anywhere | Cloud tools, free tiers — ChatGPT, Claude, Canva | Best quality, zero setup, $0. These tools are strongest exactly where the stakes are lowest — take the free win. |
| Handle with care | Cloud, paid tiers or AI inside software you already run — Copilot, Gemini, your accounting package | Paying moves you onto contract terms where your data isn't training anyone's model. Same convenience, real terms. |
| Never leaves the building | AI on your own machines — Ollama | The only setup where pasting sensitive material doesn't break a line — because there's no third party to breach. |
Two things this map gets rid of:
- The guilt of using cloud tools. If your cloud work is genuinely tier-one, you're doing nothing wrong. The people in breach stories weren't using cloud — they were using it for tier-three material.
- The fantasy of going fully local. Local models are good and improving monthly, but on ordinary hardware they still trail the best cloud models for hard writing and reasoning. Forcing everything local trades real quality for control you only need on some tasks.
Done when: every tier on your list from step 1 has a setup and a named tool.
Know what local actually requires read once
Local AI has a reputation for being complicated. The software isn't — tools like Ollama install in minutes and run free forever. What you're actually signing up for is owning the whole stack: the hardware sets the quality, and the upkeep is yours.
The honest requirements list
- Hardware. A modern machine with decent RAM runs capable open models. Old laptops won't — your hardware sets the ceiling on how good the answers get.
- Patch discipline. The data stays private, but the machine becomes the thing that needs securing. Updates and basic lockdown are your job — nobody emails you about them.
- Lower ceiling on hard tasks. For everyday drafting, summarizing, and Q&A, good open models hold up. For the hardest reasoning and long-form writing, cloud still leads. Keep those tasks in the cloud tier where they belong.
- One seat at a time. Local AI serves one user comfortably. A team passing one machine around is a bottleneck, not a setup.
None of these are disqualifiers — they're just the real shape of the trade. You give up some ceiling and some convenience, and in exchange the data never leaves. For tier-three work, that trade is worth it every time.
Done when: you know whether your hardware and habits can carry the local tier — or you've decided to keep that tier small for now.
Price it honestly about 30 minutes
The money works out less dramatic than the debate suggests. Most split setups cost less than the all-cloud plan they replace — because the sensitive work stops buying subscriptions it shouldn't be using anyway.
What each setup really costs
| Setup | Real cost |
|---|---|
| Cloud free tiers | $0 — plus the privacy trade: free plans often train on what you type. Fine for tier-one work only. |
| Cloud paid tiers | $20–30 per person monthly for the strong assistants; more per-seat inside office suites. The price of contract terms and better models. |
| Local | $0 in software, forever. The costs are hardware you likely own and your time keeping it current. |
The honest comparison isn't dollar-for-dollar. It's whether the sensitive work is currently being done in a way you'd defend to the client whose data it is. If the answer is no, the local tier's $0 price is beside the point — it's the only option on the list that passes that test.
Done when: you know your monthly number and which tier each dollar is buying.
Two things to keep in mind
Local is private, not automatically secure. Data on your machine isn't in anyone's cloud, but it's on a machine that connects to the internet. Patch it, lock it down, and treat it like the asset it now is. The privacy win is real; the security work is still yours.
If you're in a regulated industry, start with your brief. The industry hub routes to decision briefs for law firms, therapists, dentists, and others — with the lines that matter for your field already drawn.
Sort your work into three sensitivity tiers → cloud free tiers for public-safe work → paid or in-app AI for customer data → your own machines for anything a leak would hurt. Run both. Most businesses should.
