Private AI for business: what private can mean
“Private AI” is sold as one thing, but it describes at least three different setups. In one, your company’s data goes to an AI provider under business terms that, by default, rule out training. In another, nothing has to leave your building because the model runs on your own hardware. In the third, your files and who may open them live on a team server, yours or one run for you, while a hosted model does the work, so the provider still receives what the AI reads.
The useful question is not “is it private?” but three smaller ones: what leaves your company, who receives it, and what they may do with it. Here is what each setup answers, from the providers’ own pages and our own code, as of November 2026.
1. A business plan from an AI provider
Claude for Work (Team or Enterprise), the Claude API, or OpenAI’s API, under the provider’s business terms.
- Training. Anthropic says that by default it will not use inputs or outputs from its commercial products, such as Claude for Work and the Anthropic API, to train its models. [1] OpenAI says data sent to its API has not been used to train its models since March 1, 2023, unless you opt in. [5]
- The exception people miss. If someone sends Anthropic feedback, such as a thumbs up or down, or a bug report, Anthropic may use that conversation for training and keeps the data tied to the feedback for 5 years. [1][2] The owner of a Team or Enterprise plan can turn off members’ thumbs up/down feedback with the Rate chats setting. [1]
- How long the provider keeps it. Anthropic deletes API inputs and outputs within 30 days, with exceptions; in Claude for Work, chats stay in the product so people can come back to them. [2] OpenAI keeps abuse-monitoring logs of API traffic for up to 30 days by default, and some API data, such as stored conversations and files, until you delete it. [5]
- Zero data retention. Both offer it only by approval. At Anthropic it covers eligible APIs, products that use your commercial API key (including Claude Code through the API) and Claude Code for Enterprise plans, reviewed per organization; Claude for Work chat is not on that list. [3] OpenAI offers zero data retention or modified abuse monitoring to eligible customers after prior approval. [5]
- Consumer plans are different. Anthropic’s 2025 change to its consumer terms, which lets Free, Pro and Max users allow training on their chats, does not apply to Team, Enterprise or the API. [4]
What stays private is, by default, the training question. What doesn’t is the content itself: what a person types or attaches, and every file the AI reads, goes to the provider and is kept for the provider’s retention period, unless you have a zero data retention agreement. For many companies that is acceptable, and it is the least work: there is no server to run.
Fits when your data may go to a provider under a business contract, and you want to start this week.
2. An open model on your own hardware
The only setup where nothing has to leave is a model you run yourself. Ollama’s docs put it in one line: “Download Ollama and run a model on your computer.” [7]
- What it takes. OpenAI’s open-weight gpt-oss models are an example of the size involved: the larger, gpt-oss-120b, fits on a single 80 GB GPU such as an NVIDIA H100, and the smaller, gpt-oss-20b, runs within 16 GB of memory. [6]
- What you give up. It is a different model from the one your team may already use: Claude is offered through Anthropic’s apps and API and through cloud platforms such as Amazon Bedrock and Google Cloud’s Vertex AI. [4] Someone also has to keep the machine, the model and its access up to date.
- “Offline” only until you add tools. The moment the setup searches the web or connects to an online service, those requests leave your network too. Ollama also offers cloud models, which run off your machine. [7]
Fits when a contract or regulation says the data may not leave, and you can run and secure a machine that fits the model you need.
3. Your files on a team server, a hosted model doing the work
A third setup keeps the company’s files, folders, permissions and history on a team server and has a hosted model do the work. It changes where the files live and who may open them, not what the AI provider receives: the provider still gets what the AI reads and, with the features below, more. That is what we build, so read this section as ours.
Trina AX runs Claude Code, and optionally Codex, on one company server: a Linux server or an office Mac that you run (no Codex on an office Mac), or a private server for your company in Seoul that Trina hosts. On a hosted plan Trina runs the server, with encrypted off-site backups, and people who operate it, including Trina, can read stored files and chats. People use it in the browser and in the phone and desktop apps. [8][9][17]
- What the server decides. Who may open each file. Since version 0.5.4, sharing a folder works like Google Drive, with Viewer or Editor for people and groups, and the server checks every file the AI opens as the person asking: files a person can’t open stay invisible to them in files, answers, search, previews, version history and trash, though an answer someone already got keeps what it quoted, and sharing a chat or room warns before people see answers that used files they can’t open. [8][10][12]
- What goes to the AI provider. What the AI reads to answer, under the terms of the account you connect: a company API key, a business plan, or someone’s own subscription, so the training and retention terms are that account’s. Trina provides the workspace, not the AI. [8][11] Each night the team’s AI account also reads new and changed files in the shared, project and group folders and in folders connected from the server’s disk or shares mounted on it to write a file list the AI searches first, so over the following nights the contents of those folders reach that account’s provider; personal folders are not read, and only the server’s host (Trina, on a hosted plan) can turn it off. [14] Once a day it reads the masked first lines of the questions people typed in their own chats over the last 30 days to suggest common questions, and an admin can turn that off. [15] Since version 0.5.6, Find work files on this computer, in the Windows app and the Mac app from the website, sends the names, folder names, types, sizes and dates of up to 80 files on a person’s computer to that person’s AI provider when they start it (files that look personal are left out), and file contents leave the computer only for the files they add, which go to their My files on the team server. [18][19] The experimental API engine can also use another provider, and on a server you install yourself, a local model through Ollama or LM Studio. [8][11]
- What else leaves the server. Web searches, the web pages the AI opens and the services you connect; voice input and the sound of video and audio files when the speech provider is OpenRouter (on Trina hosting both go there by default, and an admin can turn either off); address and update checks; and phone-app notifications through Trina’s push service and then Apple or Google (on by default since version 0.5.4 on servers you install yourself too), as ids, an event code and the phone’s notification token with no message text, which an owner or admin can turn off unless the server’s host has set it. [11][12][13] Where the dev environment is on (by default for new teams where the server has it), a command the AI runs can send any file the person can read to the sites it may reach, such as GitHub, npm or PyPI. [16]
- What it isn’t. Trina AX has no security certification yet. [9]
Installing it yourself is free (AGPL-3.0); Trina hosting has plans from $50 a month for up to 5 people, and AI usage is billed separately by the account you connect. [8]
Fits when you want Claude working on the company’s files with per-person folder access decided on a team server, and you accept that the provider receives what the AI reads and, each night, the team folders’ contents.
Side by side
| Business plan from a provider | Open model on your hardware | Team server, hosted model (Trina AX) | |
|---|---|---|---|
| Where your files live | Your systems; what people send or attach is kept by the provider for its retention period [2][5] | On your machine [7] | On the team server: yours, or Trina’s in Seoul on a hosted plan, with encrypted off-site backups [8][17] |
| What reaches the provider | What people send and every file the AI reads [2][5] | Nothing, until you add web search, online services or cloud models [7] | What the AI reads, the team folders through the nightly file list, the masked question lines, file names from Find work files, and the speech items above [11][14][15][18][19] |
| Training on it | Not by default; Anthropic feedback and OpenAI opt-in are the exceptions [1][5] | Not applicable | The terms of the account you connect [8][11] |
| How long the provider keeps it | Anthropic API: within 30 days, with exceptions; OpenAI API: logs up to 30 days, some data until deleted; Claude for Work: chats stay in the product [2][5] | Not applicable | The terms of the account you connect [11] |
| Who runs it | The provider | You | You, or Trina on a hosted plan [8][9] |
| Which model | The provider’s | An open-weight model, such as gpt-oss [6] | Claude or OpenAI’s models, or other providers through the experimental API engine; a local model on a server you install [8][11] |
How to choose
- The data may not leave your company at all → an open model on your own hardware, with no online tools or cloud models.
- The data may go to a provider under a business contract, and you want the least setup → a business plan, with the feedback exception and the retention period in mind.
- You want the company’s files and who may open them managed on a team server, with Claude doing the work → a company server with a hosted model.
If the goal is to send the provider less, the third setup is not that: it sends at least what a business plan sends, and with the nightly file list, more.
Questions people ask
Is Claude private enough for business data?
On Claude for Work and the API, Anthropic does not train on your inputs or outputs by default, unless someone sends feedback (a Team or Enterprise owner can turn off the thumbs up/down button). [1] It still receives and stores them: API data is deleted within 30 days, with exceptions, and Claude for Work keeps chats in the product. [2]
Does Claude train on my company’s data?
Not by default on commercial products such as Claude for Work and the API. [1] Feedback someone sends is the exception, and Anthropic’s 2025 consumer-terms change applies to Free, Pro and Max, not to business plans. [1][4] If someone connects their own Pro or Max plan to a tool, that plan’s consumer terms apply. [4]
What is zero data retention?
An agreement under which the provider does not store your inputs or outputs, except where it must to comply with law or fight misuse. Anthropic and OpenAI both grant it only after approval, and at Anthropic it covers eligible APIs and Claude Code for Enterprise, not Claude for Work chat. [3][5]
Can a business run AI completely offline?
Yes, with an open model on its own hardware, for example gpt-oss-20b on a machine with 16 GB of memory or gpt-oss-120b on one 80 GB GPU. [6] It stays offline only while it doesn’t search the web, connect to online services or use a provider’s cloud models. [7]
Try it
- An open model on your own machine: docs.ollama.com. [7]
- Trina AX: Security, Files and permissions, Work with team files and hosting plans. Source: github.com/trina-app/trina-ax.
Claude is a trademark of Anthropic. Trina AX is made by Trina Inc. and is not affiliated with Anthropic.
Sources
- [1] Is my data used for model training? · privacy.claude.com/en/articles/7996868-is-my-data-used-for-model-training · read 2026-11-15
- [2] How long do you store my organization's data? · privacy.claude.com/en/articles/7996866-how-long-do-you-store-my-organization-s-data · read 2026-11-15
- [3] I have a zero data retention agreement with Anthropic. What products does it apply to? · privacy.claude.com/en/articles/8956058-i-have-a-zero-data-retention-agreement-with-anthropic-what-products-does-it-apply-to · read 2026-11-15
- [4] Updates to Consumer Terms and Privacy Policy · www.anthropic.com/news/updates-to-our-consumer-terms · read 2026-11-15
- [5] Data controls in the OpenAI platform · developers.openai.com/api/docs/guides/your-data · read 2026-11-15
- [6] gpt-oss · github.com/openai/gpt-oss · read 2026-11-15
- [7] Ollama documentation · docs.ollama.com/ · read 2026-11-15
- [8] Trina AX for LLMs (llms.txt) · ax.trina.app/llms.txt · read 2026-11-15
- [9] Security and how it works — Trina AX · ax.trina.app/security · read 2026-11-15
- [10] Files and permissions (v0.5.4) · github.com/trina-app/trina-ax/blob/v0.5.4/docs/site/use/files-and-permissions.md · read 2026-11-15
- [11] Security and permissions (v0.5.4) · github.com/trina-app/trina-ax/blob/v0.5.4/docs/security-and-permissions.md · read 2026-11-15
- [12] Changelog (v0.5.4) · github.com/trina-app/trina-ax/blob/v0.5.4/CHANGELOG.md · read 2026-11-15
- [13] server/src/notifications/push-native.ts (v0.5.4) · github.com/trina-app/trina-ax/blob/v0.5.4/server/src/notifications/push-native.ts · read 2026-11-15
- [14] server/src/file-list/service.ts (v0.5.4) · github.com/trina-app/trina-ax/blob/v0.5.4/server/src/file-list/service.ts · read 2026-11-15
- [15] server/src/chat/often-asked.ts (v0.5.4) · github.com/trina-app/trina-ax/blob/v0.5.4/server/src/chat/often-asked.ts · read 2026-11-15
- [16] server/src/devenv/setting.ts (v0.5.4) · github.com/trina-app/trina-ax/blob/v0.5.4/server/src/devenv/setting.ts · read 2026-11-15
- [17] server/src/system/cloud-admin.ts (v0.5.4) · github.com/trina-app/trina-ax/blob/v0.5.4/server/src/system/cloud-admin.ts · read 2026-11-15
- [18] Find work files on this computer (v0.5.7) · github.com/trina-app/trina-ax/blob/v0.5.7/docs/site/use/find-work-files.md · read 2026-11-15
- [19] Changelog (v0.5.7) · github.com/trina-app/trina-ax/blob/v0.5.7/CHANGELOG.md · read 2026-11-15