There is a question I get from business owners now that I did not get a year ago: is there a way to use AI without sending our information to somebody else’s computer?

Yes. You can run a capable AI model on hardware sitting in your own office. I do it. Whether you should is a different question, and it depends on things most articles about this skip past.

This is the honest version.

What “local” actually means

A local model is a language model that runs on your own machine instead of in a provider’s data center. You install software such as LM Studio or Ollama, download a model file, and it answers from your hardware. No account, no per-token bill, no network request leaving the building.

I run one on a desktop with a strong graphics card. It is not a frontier model and it does not pretend to be, but it drafts, summarizes, rewrites, answers questions about documents I give it, and helps with scripting. For a meaningful share of everyday work, the gap between it and a paid cloud model is smaller than people expect.

The three reasons a small business would want this

1. The data never leaves

This is the real one. In my guide to AI-built tools, the third failure pattern is business data getting pasted into AI services with nobody checking where it is retained. Customer records, invoices, employee files, patient names.

A local model removes the question entirely. There is nothing to check, because nothing was transmitted. If you are subject to a regulation that cares where data lives, that distinction is worth more than any feature comparison.

2. The cost is a purchase, not a meter

Frontier AI is currently cheap because investors are paying for a lot of it. The providers are spending enormous amounts to keep these systems running while the market sorts itself out.

I do not know when that changes or by how much. Nobody does. But planning your operations around a price that is being subsidized during a land grab is a risk worth naming out loud. Hardware you own has a known cost and does not reprice.

3. It works when the internet does not

Smaller point, but a real one for anyone who has watched a business stop because a service was down.

What it actually takes

The honest constraint is graphics memory, not money.

Rough guide, using the compressed model files most people run:

Treat those as approximate. Model efficiency keeps improving, and what needed 24 GB last year often fits in less now.

Practically: a decent gaming graphics card is the entry point. Plenty of businesses already have a machine that qualifies, or can get there for the price of about a year of team AI subscriptions. This is one of the few cases where a well-specified gaming PC is a legitimate business purchase.

What local models are genuinely bad at

I am not going to sell you something that does not work.

A local model is a competent generalist, not a replacement for a specialist. The businesses that get value from this use it for the eighty percent that is routine and keep a paid account for the hard twenty percent, with a rule about what may go into the paid one.

When not to do this

Skip it if any of these are true:

That last one matters most. Running your own model is a technical control layered on top of a decision about handling data. Without the decision, the control is decoration.

A short decision checklist

Three or more clear answers means it is worth pricing out. Fewer means fix the policy first.

Where this goes next

The direction that interests me is models that do more than answer. Agent software can let a model run commands and complete tasks rather than just producing text, and it runs on a local model perfectly well. I have written about that at length in When AI Gets Hands, along with why I think rising costs push more organizations toward hosting their own.

That is genuinely powerful and genuinely worth being careful with. Same rule as everything else here: understand it before you point it at anything that matters.

Where to go next

The free Small Business Security Checkup covers whether you have clear rules for AI tools and unapproved cloud services, which is the decision that should come before any of this.

If you want to talk through whether local AI makes sense for your situation, or you want help specifying and building the machine, that is what IT Wes is here for.