Contents
An owner who asks for "private AI" usually wants one assurance: that the company's customer data, contracts and prices stay the company's. The market offers that assurance at several levels, from a clause in a contract to a system the company owns.
Why the term is confusing
Every major vendor now sells a business version of its assistant with stronger data terms than the consumer version. The difference is real. On ChatGPT's Free, Go, Plus and Pro plans, training on conversations is opt-out1; on its Business plan, OpenAI says business data is never used for training2. Anthropic made the same split explicit in 2025, when it began training on consumer Free, Pro and Max accounts whose owners leave the setting on, while excluding Team, Enterprise and API use3.
But a business contract is only the first level of privacy. It settles what the vendor may do with the data. It does not settle who inside the company can reach it through the AI, where it is processed, or what the AI may do with it.
Four levels of private AI
| Level | What it is | What the company controls | Examples |
|---|---|---|---|
| 1. Business plan | A vendor's assistant under business terms | Training excluded by default; admin and SSO | ChatGPT Business, Claude Team24 |
| 2. AI inside the suite | Assistant working over the company's mail and files | Inherits existing permissions, labels and retention | Microsoft 365 Copilot, Gemini in Workspace56 |
| 3. Private deployment | Models hosted in the company's own cloud account | Region, retention, network, and the application around the model | Azure OpenAI, Amazon Bedrock78 |
| 4. Owned AI and data layer | A system the company owns, with AI acting inside it | The data layer, the rules for each AI action and who approves | OpenLazarus9 |
Level 1: a business plan
The quickest step. Claude Team lists model training as none by default and includes SSO4. ChatGPT Business adds encryption, SSO, multi-factor authentication and admin controls for users, usage and billing2. The enterprise tiers add the controls a security team asks for: Claude Enterprise adds SCIM, audit logs and custom data retention4, and OpenAI offers SCIM, role-based access and data residency through its sales team2.
Level 2: AI inside the office suite
Copilot and Gemini work over the documents the company already stores with Microsoft or Google. Microsoft says Copilot's prompts and responses fall under the same contractual terms as the company's email and files, with Microsoft acting as data processor, and that the company's retention policies and audit apply10. Google says interactions stay within the organisation and existing Workspace protections apply6. The condition is housekeeping: both retrieve whatever the user can already open56.
Level 3: models in the company's cloud account
Here the company's engineers build the application and the cloud provider hosts the model. On Azure, prompts and completions are not available to OpenAI or other model providers, are not used to train foundation models, and are processed in the chosen geography unless the company picks a Global or Data Zone deployment7. On Amazon Bedrock, model providers have no access to prompts and completions11, and a zero-retention mode writes no request or response data to durable storage8. Both keep a safety review: Azure may store flagged prompts for human review, and some models on Bedrock require AWS to retain data for up to 30 days78.
Level 4: an owned AI and data layer
At this level the company owns the layer where its data and its AI meet. OpenLazarus describes a single AI and data layer that the company, not software companies, owns, with permissions for data, files, messages and spending set to Always, Ask or Never, and the company deciding who approves; every AI teammate action and every approval is logged9. It can be maintained by the company, by any developer or with OpenLazarus engineering support9. It is an EU company with EU servers, and its AI is hosted in Europe and is not under the US CLOUD Act9. Training terms and certifications were not on the pages we read, so they belong on the list of questions.
What each level costs
| Option | Price on the vendor's page | Billing |
|---|---|---|
| Claude Team, Standard seat | $20 a seat a month annually, $25 monthly4 | 2 to 150 seats |
| Claude Team, Premium seat | $100 a seat a month annually, $125 monthly4 | 2 to 150 seats |
| Claude Enterprise | $20 a seat a month billed annually, plus usage at API rates4 | Minimum 20 seats |
| ChatGPT Business and Enterprise | No price readable on the access date1 | Business monthly or annually from 2 users; Enterprise annually |
| Microsoft 365 Copilot Business add-on | $18 a user a month annually (first-year promotion; $21 list), $25.20 monthly12 | Annual commitment; eligible Business plan required |
| Microsoft 365 Business Standard with Copilot | $23.50 a user a month annually12 | Includes the Microsoft 365 suite |
| Google Workspace Business Standard (Gemini included) | $14 a user a month on an annual plan, at US list price1314 | Gemini Enterprise priced through sales |
| Azure OpenAI | Per input and output token, by model and deployment type15 | Pay as you go, provisioned or batch |
| Amazon Bedrock | Per input and output token, by model and region16 | On demand, batch or provisioned throughput |
| OpenLazarus | Leadership engagement priced per engagement9 | Build it yourself or have it built |
The per-seat figures understate the cost of levels 3 and 4. A private deployment needs engineers to build and maintain the application, and an owned layer needs someone in the company to own its rules. Against that, the seat price of levels 1 and 2 grows with every employee.
How to choose a level
- Start from the data, not the product. List what must never leave the company's control. If the list is short and staff mainly draft and summarise, level 1 or 2 is enough.
- Use the suite you already pay for. A company on Microsoft 365 or Google Workspace gets level 2 without a new vendor contract.
- Move to level 3 when a regulator or customer asks where data is processed. Region and retention are the strongest controls on Azure and Bedrock.
- Move to level 4 when AI starts to act. Once AI sends, pays or changes records, the control that matters is who approves each kind of action, and on which data.
- Write the exit into the plan. Ask each vendor how data and settings leave if the company changes model or supplier.
The end state is a company that can name the level it chose for each kind of data, the price it pays for it and the person who answers for it. The data security checklist turns that into a list of tasks.
Common questions
What does private AI mean for a business?
It means AI whose use of company data is controlled by the company rather than by the vendor’s defaults. Depending on the product, that control comes from a business contract that excludes training, from AI running inside the company’s office suite under its permissions, from models deployed in the company’s own cloud account, or from an AI and data layer the company owns.
How much does private AI cost?
A business plan costs a fixed price per user. On the vendors’ pages, Claude Team starts at $20 a seat a month on annual billing and the Microsoft 365 Copilot Business add-on at $18 a month in a first-year promotion. Private deployment on Azure OpenAI or Amazon Bedrock is billed per token, plus the engineers who build the application.
Is private AI the same as running a model on our own servers?
Not necessarily. Most products sold as private AI run in the vendor’s cloud under contractual commitments. Azure OpenAI and Amazon Bedrock run models in the customer’s own cloud account, which gives control of region and retention without owning hardware.
Sources
- ChatGPT pricing. OpenAI. Accessed 19 Nov 2026.
- ChatGPT Business. OpenAI. Accessed 19 Nov 2026.
- Updates to consumer terms and privacy policy. Anthropic, 28 Aug 2025. Accessed 19 Nov 2026.
- Plans and pricing. Anthropic. Accessed 19 Nov 2026.
- Data, privacy, and security for Microsoft Copilot. Microsoft Learn, 9 Jul 2026. Accessed 19 Nov 2026.
- Generative AI in Google Workspace Privacy Hub. Google. Accessed 19 Nov 2026.
- Data, privacy, and security for Foundry Models sold by Azure. Microsoft Learn, 18 May 2026. Accessed 19 Nov 2026.
- Data retention in Amazon Bedrock. AWS Documentation. Accessed 19 Nov 2026.
- OpenLazarus. OpenLazarus. Accessed 19 Nov 2026.
- Enterprise data protection in Microsoft Copilot and Microsoft Copilot Chat. Microsoft Learn, 29 May 2026. Accessed 19 Nov 2026.
- Data protection in Amazon Bedrock. AWS Documentation. Accessed 19 Nov 2026.
- Microsoft 365 Copilot plans and pricing. Microsoft. Accessed 19 Nov 2026.
- Google Workspace pricing (2026): a complete guide to plans and costs (US list prices). eesel AI. Accessed 19 Nov 2026.
- Google Workspace plans and pricing. Google. Accessed 19 Nov 2026.
- Azure OpenAI Service pricing. Microsoft Azure. Accessed 19 Nov 2026.
- Amazon Bedrock pricing. AWS. Accessed 19 Nov 2026.
Updated 19 Nov 2026: First published, with prices read on the vendors’ own pages.
