Can a small office run its own private AI?

Yes. One PC with a good NVIDIA graphics card runs free tools like Ollama and Open WebUI. Everyone in the office opens a web page on the local network and chats with the AI. Nothing goes to the cloud.

How much GPU memory does an office AI server need?

12GB of VRAM handles a few light users on 8B to 14B models. 16GB is the sweet spot for a 5 to 10 person office. 24GB or more runs bigger, smarter 30B class models and handles more people at once.

Is local AI cheaper than ChatGPT Team or Copilot?

Cloud business plans charge per person, every month, so the bill grows with every new hire. A local server is a one-time hardware cost plus electricity, no matter how many people use it. For most offices of 5 or more the server pays for itself.

Private AI for Small Business: One Server, Every Desk

Affiliate note: as an Amazon Associate I earn from qualifying purchases, at no extra cost to you. Prices change constantly, so we do not list them. Click through for the current price.

Tier	VRAM	Office size	Runs well	Price
Starter	12GB	1 to 4 users	8B to 14B	Check price
Sweet spot	16GB	5 to 10 users	14B to 20B	Check price
Power	24GB	10+ users	30B to 32B	Check price
Server RAM	32GB to 64GB	Any	Room for models and apps	Check price
Battery backup	UPS	Any	Rides out power flickers	Check price

Short answer: put one good PC in the back office, give it a graphics card with 16GB of memory, and install two free programs: Ollama (the AI engine) and Open WebUI (a ChatGPT style web page). Every desk in the office opens that web page and uses the same private AI. No per-person monthly fee. Your files never leave the building.

Why one AI box for the whole office

Your data stays home. Customer names, invoices, contracts and patient or client notes never get sent to someone else's servers. That matters for HIPAA, legal, accounting and anyone who signed an NDA.
No per-seat fees. ChatGPT Team, Microsoft 365 Copilot, Gemini for Workspace and similar plans charge per person, per month, forever. Hire two people, the bill goes up. A local server is paid for once. Ten people or one, the cost is the same. Check each vendor's pricing page for today's numbers.
No limits. No "you hit your limit, come back later." Run it all day.
Works when the internet is down. The AI lives on your network, not in the cloud.
One thing to manage. One machine to update and back up, not ten logins on ten accounts.

Be honest about the tradeoff: the biggest cloud models are still smarter for hard coding, deep research and live web search. Local AI is great for the everyday office stuff: writing emails, summarizing documents, rewording letters, answering questions about your own files, and drafting quotes. Many offices keep one cloud seat for the hard jobs and run everything else locally. See local AI vs cloud AI cost for the full break-even math.

What hardware to get (caveman simple)

The one number that matters most is VRAM, the memory on the graphics card. More VRAM means smarter models and more people at once.

Starter (1 to 4 light users): 12GB VRAM. An RTX 3060 12GB runs 8B to 14B models fast. Good for email drafts and summaries.
Sweet spot (5 to 10 users): 16GB VRAM. An RTX 5060 Ti 16GB or RTX 4060 Ti 16GB runs 14B to 20B models with room for a couple of people asking at once.
Power office (10+ users or smarter answers): 24GB VRAM or more. A used RTX 3090 24GB or a newer 24GB to 32GB card runs 30B class models. Needs a 750W or bigger power supply.

The rest of the box:

RAM: 32GB minimum, 64GB if you can.
Storage: a 1TB NVMe SSD. Models are big files (5GB to 20GB each).
Network: plug it into the router or switch with a cable. Do not run the server on Wi-Fi.
Power: a small battery backup (UPS) so a flicker does not crash it.
Already have a decent desktop? Often you just add a graphics card and RAM. Run Check My PC on it first.

Want the details? Read best budget GPU for Ollama and how much RAM and VRAM you need.

How the setup works

Pick the server PC. One desktop that stays on. It can live in a closet as long as it gets air.
Install Ollama. Free download for Windows, Mac and Linux. It runs the AI model. Pull a model like Llama, Qwen, Gemma or Mistral sized for your VRAM.
Install Open WebUI. Free. It gives you a ChatGPT style chat page with user accounts, chat history and document upload. Step by step: install Ollama and Open WebUI on Windows.
Give the server a fixed address on your network (set a DHCP reservation in the router), like 192.168.1.50.
Everyone opens a browser on their own PC and goes to that address, for example http://192.168.1.50:3000. They log in with their own account. Done. Nothing to install on the desks.
Optional: tools and AI agents that speak the Ollama or OpenAI style API can point at the same server, so apps on office PCs use the shared AI too.

Similar tools work the same way: LM Studio, llama.cpp and vLLM can stand in for Ollama. Open WebUI is the easiest front door for a non-tech office.

Security basics (do not skip)

Keep it on the office network only. Do not port forward it on your router. If people need it from home, use a VPN (such as WireGuard or Tailscale), not an open port.
Turn on logins. In Open WebUI, the first account is the admin. Turn off open sign-ups and add each person yourself.
Do not expose Ollama directly. Ollama has no password of its own. Let only Open WebUI talk to it, and use the Windows firewall to allow just your office network.
Strong admin password and updates. Keep Windows, the GPU driver, Ollama and Open WebUI updated monthly.
Back it up. Back up the Open WebUI data folder (users, chats, uploaded documents) like any other business file.
Set a simple rule. Tell staff what is OK to paste in. Local is private, but chats are still stored on the server, so treat it like the office file share.

Want help setting it up?

Start with the free Check My PC tool. It tells you in plain words whether a PC you already own can be the office AI server, and the one part to add if it cannot. Nothing to download. Nothing leaves your PC.

If you are in Dallas Fort Worth and would rather have someone do it, Blake has spent decades troubleshooting PCs and can help set up a private AI server for your office. Get in touch.