The three options compared
| Business plan of a public chatbot | Your own chat app over model APIs | Self-hosted open-weight models | |
|---|---|---|---|
| Examples | ChatGPT Business, Claude Team, Microsoft 365 Copilot | Open WebUI or similar, connected to an AI provider’s API | Ollama or LM Studio on your own hardware; AdminOnsite |
| Training on your data | Off by default on the major vendors’ business plans | Off by default on the major model APIs | No vendor involved |
| Who holds the prompts | The chatbot vendor | Whoever runs the app, plus the model provider | You |
| How you pay | Per seat | Per use (tokens), plus hosting | Hardware, or one monthly price for a managed box |
| Model quality | Frontier | Frontier | Open-weight; strong for office work |
| Upkeep | The vendor | You or your IT provider | You, your IT provider, or AdminOnsite |
What it costs for a 5, 10 or 20-person office
| Office size | Business plan at $20–25 per seat | Your own app over APIs | AdminOnsite (managed on-site) |
|---|---|---|---|
| 5 people | $100–125 a month | Depends on usage | Practice: $249 a month + $299 setup |
| 10 people | $200–250 a month | Depends on usage | Office: $399 a month + $399 setup |
| 20 people | $400–500 a month | Depends on usage | Firm: $699 a month + $499 setup |
US list prices from each vendor’s own page, checked November 13, 2026. Prices change; check the vendor before you buy. Per-seat figures use standard seats; premium seats cost $100–125 each.
Which kind fits law, CPA and insurance offices?
If your clients, insurer or regulator ask where their files are processed, the first two options give the same answer: on a vendor’s servers, under contract. For many firms that’s acceptable, and a business plan is the cheapest way to frontier models.
Where the answer needs to be “in our office”, only self-hosting gives it. The cost is model quality at the top end and the work of running a server; AdminOnsite takes the second part off your hands for a flat monthly price. See the guides for law firms, accounting firms and insurance agencies.
Six questions to ask any “private AI” provider
- Where are conversations processed and stored?Which company, which cloud, which country, including any model provider behind the app.
- For how long?Default retention, who can change it, and when deleted data leaves backups.
- Who can read them?Your admins, the provider’s staff, contractors, and for what reasons.
- Does anything train a model?Prompts, files or feedback, and whether that’s off by default.
- What can your admin control?Sign-in, roles, folder access, retention and logs.
- What does the contract say?The terms that make each answer binding.
Questions
Can I run ChatGPT on my own servers?
No. ChatGPT runs only as OpenAI’s hosted service. To keep prompts on your own hardware you run open-weight models instead.
What is private AI?
AI used in a way that keeps your data under your control: not used for vendor training, with retention you decide, and, in the strictest sense, processed only on systems you control.
Is ChatGPT Business private?
It’s private in the sense that OpenAI doesn’t train on workspace data by default and your admins control access. It isn’t on-site: OpenAI processes and stores the data, and admins can view members’ conversations.
Is a private LLM as good as ChatGPT?
For drafting, summarizing and searching your own files, open-weight models are very capable. On the hardest reasoning they trail the newest cloud models.