Private AI for business: keep your data out of public chatbots
What happens to business data typed into AI chatbots, the four levels of private AI from business plans to self-hosted models, costs, and a one-page staff policy.
Private AI for business means using AI without handing your business's information to someone else's training data, storage or staff. For most small businesses it takes three things: a business plan of the AI assistant your team already likes (not personal accounts), the right settings switched on in it, and a one-page rule for staff on what goes where. Businesses with stricter needs can go further: keep data in a chosen country, run models inside their own cloud account, or run an open-weight model on their own hardware. Below are the four levels, what each costs, what to check in a provider's terms, and a staff policy you can copy.
- The biggest risk isn't the AI company. It's staff using personal accounts: Cyberhaven's 2026 report found a third of employees reach AI tools through personal accounts.
- Business plans of ChatGPT, Claude, Gemini and Copilot don't train on your content by default. Personal plans may.
- Four levels of privacy: business plan, your own cloud account, a chosen country, your own hardware. Each costs more effort than the last.
- Most small businesses need level 1 done properly, plus a private knowledge base for their own documents.
- Write the rule down. A one-page staff policy does more than any setting.
What actually happens to what you type into a chatbot?
It depends on the plan, not the product. The same ChatGPT can treat your words in two different ways.
On OpenAI's personal plans (Free, Plus and Pro), OpenAI says it may use your content to train its models unless you turn off "Improve the model for everyone" under Settings, Data controls. Even after you opt out, pressing thumbs up or down on an answer lets OpenAI use that whole conversation for training. On ChatGPT Business, Enterprise and Edu, and through its API, OpenAI says it does not use inputs or outputs for training by default. Anthropic draws the same line: its consumer terms don't apply to Claude for Work or the API, and its Team plan has no model training on your content by default. Google's Workspace privacy hub and Microsoft's "enterprise data protection" for Copilot make similar promises for business accounts, with the provider acting as your data processor under a data protection agreement.
Training is only one question. The others are how long your words are kept, whether people at the provider can read them, which country they're stored in, and who in your own business can see them. That's what the four levels below are about.
Why personal accounts are the real leak
In April 2023, Samsung engineers pasted internal source code into ChatGPT to help fix it. Samsung banned generative AI tools on company devices the following month, worried that what staff had sent was hard to retrieve or delete. Banning didn't work for long, anywhere: people simply use AI on their phones.
THE LEAK IS THE PERSONAL ACCOUNT
Sensitive data goes into AI daily, often through accounts the business can’t see.
- AI interactions that involve sensitive data39.7%
About once every three days for the average employee.
- Employees reaching AI tools through personal accounts1 in 3
- Claude users on personal accounts58%
- Perplexity users on personal accounts60%
Show the data as a table
| Measure | Share |
|---|---|
| AI interactions that involve sensitive data | 39.7% |
| Employees reaching AI tools through personal accounts | 1 in 3 |
| Claude users on personal accounts | 58% |
| Perplexity users on personal accounts | 60% |
Cyberhaven's 2026 AI Adoption and Risk Report, based on data movements inside companies, found that 39.7% of AI interactions involve sensitive data, which works out at the average employee putting sensitive information into an AI tool about once every three days. A third of employees reach AI tools through personal accounts, including 58% of Claude users and 60% of Perplexity users. A personal account is invisible to the business, carries personal-plan terms, and leaves with the employee when they go.
The fix isn't a ban. It's giving people an approved tool that's as good as the one on their phone, and telling them plainly what goes in it.
The four levels of private AI
| Level | What it is | Where your data goes | Typical cost | Effort | Good for |
|---|---|---|---|---|---|
| 1. Business plan | ChatGPT Business, Claude Team, Gemini in Workspace or Copilot with enterprise data protection, set up properly | The provider's servers, under a data processing agreement, not used for training by default | About $20 to $25 per user a month | An afternoon | Almost every small business |
| 2. Your own cloud account | Models used through Amazon Bedrock, Google Vertex AI or Azure, inside the cloud account you already have | Your cloud account; Bedrock, for example, says inputs and outputs aren't shared with model providers or used for training | Pay per use, charged by the amount of text processed, plus a build | Days to weeks, with a developer | Assistants and automations built into your own systems |
| 3. A chosen country | Data residency: storage, and sometimes processing, kept in a named region | Your chosen region, such as the UK, the EU or the US | An enterprise plan or an approved API account, or level 2 in a chosen cloud region | Paperwork plus setup | Regulated work, public-sector clients, contracts that name a country |
| 4. Your own hardware | An open-weight model such as gpt-oss running on your own machine or server | Nowhere: it never leaves your building | Hardware, power and someone to look after it | Ongoing | Highly confidential material, or no internet allowed |
Level 1 is the one most businesses skip and most need. Claude Team, for example, costs $25 a seat a month, or $20 billed annually, for teams of 2 to 150, as of November 2026; ChatGPT Business was the same price when we last checked its pricing in October. The work is in the setup: turn on single sign-on if you have it, remove personal accounts from company devices, decide who can connect the assistant to email and files, and switch off anything you don't need.
Level 2 suits AI that's built into your business rather than used in a chat window: a chatbot on your website, an assistant that reads invoices, an agent that books appointments. You call the models through the cloud account you already pay for, so the data stays under the agreements you already have. Amazon says Bedrock content is not used to improve the base models and is not shared with any model providers.
Level 3 is about geography. OpenAI offers data residency in Europe, the UK, the US, Canada, Japan, South Korea, Singapore, India, Australia and the UAE, but for ChatGPT Enterprise, Edu and approved API customers, not ChatGPT Business. The big clouds let you pick a region for most models. Ask for it only if a law, a regulator or a client contract actually requires it, because it narrows your choice of models and plans.
Level 4 is the only option where nothing leaves the building. OpenAI's open-weight gpt-oss-20b runs in about 16 GB of memory, and gpt-oss-120b needs a single 80 GB graphics card, both under the permissive Apache 2.0 licence. The trade-offs are real: answers are less capable than the best hosted models, the machine needs updates, backups and security, and every new model is your job to install. We suggest it for a narrow job, such as summarising confidential files, not as everyone's daily assistant.
Private AI for your own documents
Most businesses don't want a private chatbot that knows the world. They want one that knows their price list, procedures, contracts and past jobs, and answers staff questions from them with the source attached. That's a private knowledge base, built with a method called retrieval-augmented generation (RAG). Our plain-English guide to RAG explains how it works.
A private knowledge base can run at any of the four levels. The parts that make it private are the same each time: it only reads the folders you choose, each person only gets answers from documents they're allowed to open, nothing is used for training, and every answer shows the document it came from. That's how we build private knowledge bases for clients.
What to check in a provider's data terms
Before you sign up, or before you trust the plan you already have, check these ten things in the provider's own terms and admin settings. Write down the date you checked, because they change.
- Training. Is your content used to train models by default? Can anyone in your business switch that on?
- Feedback buttons. Does rating an answer share the conversation? On OpenAI's personal plans it does.
- Retention. How long are conversations and files kept, and can you shorten it?
- Deletion. When you delete something, when does it actually go? Are there exceptions for legal holds or safety reviews?
- Human review. Can the provider's staff or contractors read your content, and when?
- Location. Where is data stored and processed? Can you choose?
- Agreement. Is there a data processing agreement, with the provider as your processor?
- Subprocessors. Which other companies handle your data, and in which countries?
- Connectors. When the assistant is linked to email, files or a CRM, what can it read, and can you limit it?
- Admin controls. Can you see who uses it, remove leavers' access, require single sign-on and export or delete everything?
If a provider can't answer the first seven clearly in writing, don't put customer information into it.
A one-page AI policy for staff
The most useful privacy measure costs nothing. Copy this, fill in the brackets, and pin it where people will see it.
AI AT WORK
Use AI as much as it helps. Use [OUR APPROVED TOOL] for work, never a personal account, and check what it writes before it goes to anyone.
FINE IN ANY AI TOOL
- Public information: our website, published prices, news
- Drafts with names and numbers taken out
- General questions: how do I write a formula, word a reply, plan a week
ONLY IN [OUR APPROVED TOOL], SIGNED IN WITH YOUR WORK ACCOUNT
- Customer and staff names, emails, orders and notes
- Contracts, quotes, financial figures and plans
- Anything a client sent us in confidence
NEVER IN ANY AI TOOL
- Passwords, keys, codes and card numbers
- Health information, unless [OWNER] has approved the tool for it
- Anything we’ve signed a confidentiality agreement about, unless it allows it
Pasted something you shouldn’t have? Tell [OWNER] the same day. Nobody gets into trouble for saying so; we only want to fix it.
SIGNED: ____________ (OWNER)REVIEWED: [DATE]
Keep it to one page. Long AI policies don't get read, and the people most likely to paste a customer list into a free chatbot are the busiest ones, not the careless ones. Review it every six months, or whenever you add a new tool.
The rules behind it
Data protection law doesn't stop you using AI. It asks you to do what the levels above already do. Under the UK GDPR and the EU GDPR, an AI provider handling customer or staff data on your behalf is a processor, so you need a contract with it, a lawful reason to use the data, and to tell people in your privacy notice. In the US there's no single federal law, but state privacy laws, sector rules such as HIPAA for health information, and your own client contracts often say where data can go. If you handle health, legal or financial records, or work for the public sector, check those first. This is general information, not legal advice.
What we would do
For a business of 2 to 50 people:
- Pick one assistant your team likes and buy its business plan for everyone who uses AI at work.
- Set it up properly: single sign-on, training off, connectors limited, personal accounts off company devices.
- Pin the one-page policy and talk it through in ten minutes at a team meeting.
- Build a private knowledge base for the documents people keep asking about, so they stop pasting them into chat windows.
- Go to levels 2 to 4 only for a reason: a law, a client contract or a genuinely confidential job.
Our AI consultancy can set up level 1 and write the policy with you; private knowledge bases and AI integration cover the rest. If you're still choosing an assistant, our comparison of ChatGPT, Claude, Gemini, Grok and Muse sets out what each plan costs and what it does with your data.
Next Step
Talk to us about private AI
A free 20-minute call: which level you need, what to switch on, and a written quote if there's something to build.
Frequently Asked Questions
What is private AI for business?
AI set up so your business's information stays under your control: it isn't used to train anyone's models, is kept only as long as you choose, is stored where you choose, and is seen only by the people who should see it. It ranges from a business plan of a public assistant with the right settings, through models run inside your own cloud account, to an open-weight model on your own hardware.
Is ChatGPT safe for business data?
On a business plan, with care. OpenAI says it does not use inputs or outputs from ChatGPT Business, Enterprise, Edu or its API to train its models by default. On Free, Plus and Pro, content may be used for training unless you turn off 'Improve the model for everyone' in Settings, Data controls, and feedback you give with the thumbs buttons can still be used even after you opt out. Personal accounts are the wrong place for customer, staff or financial information.
Does ChatGPT or Claude train on my data?
On personal plans it depends on your settings. On business plans, no by default: OpenAI says so for ChatGPT Business, Enterprise and its API, and Anthropic says Claude Team has no model training on your content by default. Check the current terms on the day you sign up, because they change.
What is a private LLM?
A large language model that runs somewhere you control, such as your own server or your own cloud account, instead of a shared public service. Open-weight models such as OpenAI's gpt-oss can be downloaded and run this way. They are less capable than the best hosted models, but nothing you type leaves your machines.
Can a small business run AI on its own computer?
Yes, for lighter jobs. OpenAI's gpt-oss-20b runs on a machine with about 16 GB of memory, and tools such as Ollama install it in minutes. The larger gpt-oss-120b needs a single 80 GB graphics card. Expect slower, less polished answers than a paid business plan, plus the work of keeping the machine updated and backed up.
Is private AI more expensive?
The first level isn't: a business plan costs about the same as a personal one, around $20 to $25 per user a month. Data residency and enterprise controls usually mean an enterprise contract or a developer build. Self-hosting swaps the subscription for hardware and someone's time. Most small businesses need the first level done properly, plus a private knowledge base for their documents.
What should staff never put into an AI tool?
Into a personal or free AI account: anything about a named customer or employee, passwords and keys, bank and card details, health information, unreleased prices or plans, contracts and anything under a confidentiality agreement. Into the business's approved tool, most of that is fine for the jobs it was approved for, except passwords, keys and card numbers, which never go into any AI tool.
Next Action
Want help implementing this in your business?
Tamkoz builds AI phone receptionists, bespoke chatbots, and business automations. All work is quoted in writing before starting.