"Nine ways to use AI on your own data, from ChatGPT Enterprise to self-hosted Llama, compared on training, retention, region and encryption, with US, UK, Australian and EU rules."
Key Takeaways
- 1Every business AI plan we checked on 22 October 2026 (OpenAI, Microsoft, Anthropic, Google, AWS) says it does not train on your data by default. The real differences are retention, region and who can see what.
- 2Check five things for each option: training use, retention, data region, encryption, and access control. "Private" on a sales page covers some of them, rarely all.
- 3Business seats (ChatGPT, Copilot, Claude, Gemini) are the fastest route. A model in your own AWS, Azure or Google Cloud account is the most private route most companies need. Self-hosting is for data that cannot leave your network.
- 4RAG connects the AI to your own documents without retraining it, respects permissions and shows sources. It sits on top of any of the options.
- 5No AI tool is HIPAA compliant by itself. You need a business associate agreement (BAA) for the specific product. OpenAI names the API and ChatGPT for Healthcare, not ChatGPT Enterprise.
- 6Outside the US, UK firms often need a data protection impact assessment (DPIA) and an International Data Transfer Agreement (IDTA). Australian firms must meet Australian Privacy Principle 8 (APP 8) for overseas vendors. EU AI Act Article 50 transparency has applied since 2 August 2026.
- 7Market build ranges for a private RAG system run from about $10,000 for a basic setup to $150,000 or more for enterprise scale (ScalaCode, July 2026).
The short answer
Private AI for business means the vendor does not train on your prompts and files, keeps them only as long as you allow, stores them in your chosen region, encrypts them, and shows them only to people you authorise. Most companies get there with a GPT, Claude or Gemini model inside their AWS, Azure or Google Cloud account, linked to their files with RAG. Self-host an open model only when data cannot leave your network.
We design, build and support private AI assistants and agents, so weigh this page with that in mind. Every vendor term below was read on the vendor's own page on 22 October 2026 and is linked where it appears. Terms change, so check the link before you sign. This is general guidance, not legal advice or a compliance opinion.
The usual buyer is an operations, IT or compliance lead who has found staff pasting customer data into personal AI accounts. Most vendor pages describe one route. This guide compares nine, from ChatGPT Business (formerly ChatGPT Team) to models on your own servers, on the same five tests.
What "private" actually means: the five tests
Vendors use "private" and "secure" loosely. Ask these five questions of every option, and get the answer from the vendor's terms, not the sales deck.
- Training use. Are your prompts, files or answers used to train or improve the vendor's models? By default, and after any setting change?
- Retention. How long does the vendor keep prompts and answers, who sets that, and is there a zero data retention (ZDR) option, meaning nothing is stored after the answer is returned?
- Data region. Where is data stored and where is it processed? Can you pin both to the US, UK, EU or Australia? This is what AI data residency means in a contract.
- Encryption. Is data encrypted at rest and in transit, and can you hold your own encryption keys (customer-managed keys)?
- Access control. Can the AI only see what each user is already allowed to see, with single sign-on and a log of every question and answer?
The fifth test depends on your own file permissions, not on the vendor. If every employee can open the HR folder, an AI assistant connected to that folder will happily summarise salaries for anyone who asks.
The nine options compared (vendor terms read 22 October 2026)
Swipe sideways to see every column.
| Option | Trains on your data by default? | Retention | Data region | HIPAA BAA |
|---|---|---|---|---|
| ChatGPT Enterprise | No | Admin-set; deleted chats gone within 30 days | At rest in US, Europe, UK, Australia and 6 more for eligible customers; in-region processing US or Europe | Not named (OpenAI names ChatGPT for Healthcare and API) |
| ChatGPT Business (was Team) | No | Deleted or unsaved chats gone within 30 days | Not on OpenAI's residency list | Not named |
| Microsoft Copilot (Microsoft 365) | No (not used to train foundation models) | Admin-set in Microsoft Purview | EU Data Boundary for EU traffic; Anthropic models excluded | HIPAA listed; check your Microsoft agreement |
| Claude Team and Enterprise | No, unless you opt in | Deleted within 30 days; ZDR not available for chat | Not stated for chat plans | Enterprise, once HIPAA mode is switched on |
| Gemini for Google Workspace | No, not outside your domain without permission | Admin-set, 90 days to indefinite | Not stated on the privacy hub | Google says it can support HIPAA workloads |
| Amazon Bedrock | No; not shared with model providers | Set per AWS Region; zero retention possible | The AWS Region you choose | HIPAA eligible |
| Azure OpenAI (Azure AI Foundry) | No, without your permission | Stored data deletable by you | Your chosen geography, or a US or EU Data Zone | Check your Azure agreement |
| Google Vertex AI | No, without your prior permission | Cache up to 24 hours; ZDR with set-up steps | US, EU, UK (europe-west2), Australia (australia-southeast1) and more | Check your Google Cloud agreement |
| Self-hosted Llama or Mistral | Not applicable: no vendor sees data | You decide | Wherever you run it | Your own safeguards |
The table covers options 1 to 8: (1) ChatGPT, (2) Copilot, (3) Claude, (4) Gemini, (5) Bedrock, (6) Azure OpenAI, (7) Vertex AI and (8) self-hosted models. Option 9, RAG over your own documents, sits on top of any of them and has its own section below.
1 to 4: business AI seats (fastest, least control)
ChatGPT Enterprise and Business. OpenAI's enterprise privacy page says: "By default, we do not use your business data for training our models." Enterprise adds AES-256 encryption at rest, TLS 1.2 or higher in transit, customer-held keys, SOC 2 Type 2 and ISO/IEC 27001, and data at rest in ten regions including the US, UK, Europe and Australia for eligible customers. The business data page offers BAAs to "ChatGPT for Healthcare and API healthcare customers", so do not assume Enterprise is covered for patient data.
Microsoft Copilot. Microsoft's Copilot privacy page says prompts, responses and data accessed through Microsoft Graph "aren't used to train foundation LLMs". Copilot respects the permissions users already have, which is both its strength and its risk. EU calls stay in the EU Data Boundary, with the exception Microsoft names for Anthropic models used inside Copilot. If you are deciding between Copilot's own agent builder and a custom build, see Copilot Studio vs custom AI agents.
Claude for Work (Claude Team and Claude Enterprise). Anthropic's privacy centre says it will not use inputs or outputs from commercial products, including Claude for Work, to train models unless you explicitly opt in. Zero data retention covers the API and Claude Code on Enterprise, not chat, and a BAA covers Claude Enterprise only after the primary owner switches on HIPAA mode.
Gemini for Google Workspace. Google's Workspace privacy hub says content is not used for model training outside your domain without permission, lets admins keep Gemini prompts from 90 days to indefinitely, and lists SOC 1, 2 and 3, ISO/IEC 27001 and FedRAMP High.
5 to 7: models inside your own cloud account (the usual sweet spot)
Here you call GPT, Claude, Gemini, Llama or Mistral models through the cloud you already pay for, under the contract, region, keys and audit logs you already have. This is what most companies mean when they ask for a "private ChatGPT" or a "private GPT".
- Amazon Bedrock. AWS's Bedrock FAQ says your content "is not used to improve the base models and is not shared with any model providers", is encrypted in transit and at rest, and can use your own keys. AWS lists Bedrock as HIPAA eligible and in scope for ISO and SOC. It runs Claude, Llama, Mistral and others.
- Azure OpenAI (Azure AI Foundry). Microsoft's data privacy page says prompts and completions are not used to train any generative AI foundation models without your permission, encrypts with AES-256 by default with a customer managed key option, and keeps data in your chosen geography unless you pick a Global or Data Zone deployment.
- Google Vertex AI. Google's data governance page says Google won't use your data to train or fine-tune models without prior permission. Its residency list includes the UK (europe-west2) and Australia (australia-southeast1), which matters for the compliance section below. A
globalendpoint carries no residency promise, so pick a regional one.
Direct APIs sit between seats and cloud. The OpenAI API does not train on API data, keeps abuse-monitoring logs up to 30 days, and offers regional storage including the UK and Australia. Outside the US, zero data retention or modified abuse monitoring needs OpenAI's approval, and newer models cost 10% more there. The Anthropic API offers US-only or global processing, with US-only priced at 1.1 times standard, so UK, EU and Australian buyers who need data held in-country usually reach Claude through Bedrock or Vertex instead.

8: a private LLM on your own servers (self-hosted Llama or Mistral; most control, most work)
An open-weight model is a model file you download and run on your own servers or cloud machines. Nothing goes to an AI vendor. In exchange you own the hardware, security patches, upgrades and evaluation, and open models often trail the best API models on hard reasoning.
Teams usually serve these models with open-source tools such as vLLM or Ollama. Llama and Mistral are the two families covered below. Qwen, DeepSeek and Google's Gemma are other open-weight families, each under its own licence, so read the licence before you build on one.
- Llama 4 (Meta). The licence allows commercial use, requires companies over 700 million monthly active users to request a licence from Meta, and asks you to show "Built with Llama" where you use it.
- Mistral. Mistral says its Mistral 3 models are "released under the Apache 2.0 license", a permissive licence with no user cap. Some other Mistral models carry different licences, including one non-commercial voice model, so check each model card.
Option 9, RAG (retrieval-augmented generation): how private AI answers from your own documents
RAG is how an AI answers from your contracts, policies, product sheets or tickets without being retrained on them. Here is the business version, in five steps:
- Collect. Pull documents from SharePoint, Google Drive, your ERP, CRM or help desk, and drop duplicates and stale versions.
- Index. Split each document into short passages and store a searchable version (a vector) in a database inside your own cloud account. Common choices are pgvector on PostgreSQL, Azure AI Search, Vertex AI Vector Search, Amazon OpenSearch Service and Pinecone.
- Retrieve, with permissions. When someone asks, fetch only the passages that person is allowed to see. Skip this filter and the assistant can quote files the person asking could never open.
- Answer. Send the question and those few passages to the model you chose above. The model sees snippets, not your whole library.
- Cite and log. Show which document each fact came from, and log the question, the passages and the answer for audit.
RAG beats fine-tuning (retraining a model on your data) for most business knowledge: it updates the moment a file changes, a deleted document stops appearing in answers, and permissions carry through. We build RAG systems and model integrations as part of our LLM development services; when the assistant also takes actions, such as drafting an order or updating a record, it becomes an agent (see AI agent development).
Not sure which of the nine fits your data rules?
Bring your data types, your regions and the systems the AI must read. In a 30-minute call with founder Bhavesh Barot we will point to the option that meets your rules with the least work, and we show working software on your own data before you sign anything.
Talk to the FounderCompliance by country: US, UK, Australia, EU
Each of these rules asks two things: know where the data goes, and hold the right contract with whoever stores it. The laws differ by country, but UK GDPR and EU GDPR turn on the same two points.
Swipe sideways to see every column.
| Country | Rule that bites | What it means for private AI |
|---|---|---|
| US (healthcare) | HHS cloud computing guidance under HIPAA | A cloud provider holding patient data is a business associate even if it stores only encrypted data and lacks the key. You need a signed BAA for the exact product. |
| US (B2B buyers) | SOC 2 (AICPA) | An audit report, not a law, covering security, availability, processing integrity, confidentiality and privacy. Enterprise customers will ask for your vendor's report. |
| UK | ICO AI guidance, DPIA rules, transfer rules | AI is listed as an innovative technology trigger for a DPIA alongside one other factor. Data stored outside the UK needs a transfer mechanism such as the IDTA plus a transfer risk assessment. The ICO notes its AI guidance is under review after the Data (Use and Access) Act. |
| Australia | APP 8 and OAIC AI guidance | Before personal information goes to an overseas vendor, take reasonable steps so it is handled to Australian standards. The OAIC says not to enter personal information into publicly available AI tools. |
| EU | GDPR Article 28, EU-US Data Privacy Framework, AI Act Article 50 | A processor contract with the vendor; US transfers can rely on the Data Privacy Framework if the vendor is certified. Article 50 transparency (tell people they are talking to AI) has applied since 2 August 2026. |
On the EU dates: the AI Act Service Desk shows the Act applying from 2 August 2026 under Article 113. The 2026 Digital Omnibus, Regulation (EU) 2026/1744, in force since 27 July 2026, moved the high-risk deadlines to 2 December 2027 and 2 August 2028 but left Article 50 on 2 August 2026. A private internal assistant is rarely high-risk; one used for hiring or credit decisions can be.
HIPAA compliant AI: which products sign a BAA
HHS treats a cloud provider that holds patient data as a business associate, so a US healthcare company needs a signed BAA for the exact product it uses. Per the vendor pages above, OpenAI offers BAAs for the API and ChatGPT for Healthcare, not ChatGPT Enterprise. Anthropic covers its API and Claude Enterprise once HIPAA mode is on. AWS lists Amazon Bedrock as HIPAA eligible. Google says Gemini in Workspace can support HIPAA workloads. For Copilot, Azure OpenAI and Vertex AI, check the BAA terms in your existing Microsoft or Google Cloud agreement.
Which option fits your company size
Swipe sideways to see every column.
| Company size | Usual best fit | Why | Watch out for |
|---|---|---|---|
| Under 50 staff | Business seats in the suite you already use | Live in a day; no-training default; admin controls | Staff still using personal accounts; no written AI rule |
| 50 to 250 staff | Seats for general use, plus one RAG assistant in your cloud for your own documents | Answers from your files with sources; your region and keys | Shared-drive permissions that are too open |
| 250 to 1,000 staff | Models in your AWS, Azure or Google Cloud account, RAG with single sign-on and role-based access | One contract, one audit trail, region pinned | Several teams buying separate tools; no owner |
| Regulated, any size (health, finance, legal) | Cloud models under your existing BAA or regulator-approved contract; self-host only if rules demand it | Contract coverage matters more than the model | Assuming a chat plan is covered by a BAA when only the API is |

Private AI cost: what a build costs (market ranges)
Seats are priced per user by each vendor. On 31 October 2026, Microsoft's pricing page listed Microsoft 365 Copilot Business at $21 per user a month on a yearly plan ($18 for the first year under a promotion running to 31 December 2026), and Anthropic's pricing page listed Claude Team and Claude Enterprise seats at $20 a month billed yearly, with Enterprise usage billed on top at API rates. A custom private assistant over your documents is a build project. These are published market ranges, not our prices:
Swipe sideways to see every column.
| Source | Scope | Published range |
|---|---|---|
| ScalaCode (July 2026) | Basic RAG | $10,000 to $25,000 |
| ScalaCode | Production RAG | $25,000 to $60,000 |
| ScalaCode | Enterprise agentic RAG | $60,000 to $150,000+ |
| RaftLabs (August 2025) | Simple document Q&A | $12,000 to $30,000 |
| RaftLabs | Enterprise RAG platform | $70,000 to $120,000+ |
Both are development firms pricing their own market, so treat the ranges as a sense of scale. What moves a project up the range: the number of source systems, per-user permissions, the accuracy bar, and whether the assistant only answers or also acts. FactoryJet quotes the build, and any support you choose, as fixed prices in writing after a short scoping call. For agent build ranges, see our enterprise AI agent cost guide.
Eight red flags when a vendor says "private AI"
- They cannot link to the clause that says your data is not used for training.
- "Private" means a password on a shared consumer account.
- No answer on retention, or retention only the vendor can change.
- A
globalor unspecified region when your rules need the UK, EU or Australia. - A BAA promised for a product the model maker does not list as covered.
- The AI sees every document regardless of who is asking.
- No log of questions, sources and answers you can export.
- You do not own the code, prompts, index and test set at the end of the project. Our AI development contract checklist covers the clauses.
Seven questions to ask before you sign
- Which model provider and which product (seat, API, or cloud platform) will process our data?
- Where is data stored and processed, and can both be pinned to our region?
- How long are prompts and answers kept, and can we turn retention off?
- Who holds the encryption keys?
- How are our existing file permissions enforced in answers?
- Which agreement covers us (BAA, Article 28 contract, IDTA), and for which features?
- How will accuracy be tested on our own questions before staff use it?
For the wider security checklist once an assistant can take actions, read AI agent security: the OWASP Top 10 risks and controls, and put the answers into your AI governance policy.
Other private AI vendors and guides, and how they differ
A US buyer searching Google for private AI on 22 October 2026 found these business pages on the first page. We list them so you can compare. We have not reviewed their services, and the list leaves out forums.
Swipe sideways to see every column.
| Site | What they sell | Page title shown in Google |
|---|---|---|
| coresite.com | Data centre colocation | Private AI: The Smart Choice for Many Enterprises |
| digitalrealty.com | Data centre colocation | What is Private AI? The Next Step in Privacy-First Innovation |
| cloudera.com | Data platform software | What Is Private AI? |
| wezom.com | Software development agency | Private AI for Business: How to Protect Sensitive Data in AI |
| hpe.com | Servers and hardware | What is Private AI? | Glossary |
| teradata.com | Data analytics platform | What Is Private AI? The Enterprise Playbook |
| suse.com | Linux and container platform software | Private AI: Securing Innovation for the Future of Enterprise |
Most of these come from data-centre, hardware or platform vendors, and several define private AI as running it on their infrastructure. This guide compares the nine routes a mid-sized company actually chooses between, quotes each vendor's own terms with the date we read them, and adds the US, UK, Australian and EU rules on one page.
How FactoryJet builds private AI on your data
We design, build, implement and support private AI assistants and agents: model choice, RAG over your documents with your permissions, single sign-on, logging, and testing on your own questions. AI and hosting accounts can run in your name, at cost, in the region you need, and we keep managing the servers, models, APIs and updates after launch. You own the code, prompts, index and test set. We show working software on your own data before you sign.
See LLM development, AI agent development and AI integration services.
Sources
Pages opened and read on 22 October 2026 unless marked otherwise.
- OpenAI, Enterprise privacy and Business data privacy, security and compliance
- OpenAI, Data controls in the OpenAI platform
- Microsoft Learn, Data, privacy and security for Microsoft Copilot
- Anthropic privacy centre: model training, zero data retention, BAAs; Claude API data residency
- Google Workspace, Generative AI in Google Workspace privacy hub
- AWS, Amazon Bedrock FAQs and Bedrock security and compliance
- Microsoft Learn, Data, privacy and security for Azure OpenAI
- Google Cloud, Vertex AI data governance and data residency
- Meta, Llama 4 Community License; Mistral AI, Introducing Mistral 3
- US HHS, Guidance on HIPAA and Cloud Computing; AICPA, SOC 2
- ICO, Guidance on AI and data protection, When do we need to do a DPIA?, International transfers
- OAIC, APP guidelines chapter 8; OAIC, Guidance on privacy and commercially available AI products
- GDPR Article 28; European Commission, EU-US data transfers; AI Act Service Desk, Article 50 and Article 113
- ScalaCode, RAG development cost; RaftLabs, RAG development cost
- EUR-Lex, Regulation (EU) 2026/1744 (Digital Omnibus on AI)
- Microsoft 365 Copilot pricing; Claude pricing (both read 31 October 2026)
- vLLM documentation; Ollama; pgvector
- Google first-page results for the comparison table, checked through DataForSEO on 22 October 2026
Private AI on your own data, for US, UK and Australian companies
See a private assistant answer from your own documents before you commit
Tell us your data rules, your region and the systems it must read. We will recommend the route and show it working on your own data before you sign. FactoryJet quotes a fixed price in writing after a short scoping call.
Frequently Asked Questions
Is there an AI that is completely private?
Can I have my own private AI?
What does private AI mean?
Does ChatGPT Enterprise train on your data?
Is ChatGPT Enterprise safe for confidential information?
Will ChatGPT keep my data private?
What is the difference between ChatGPT and ChatGPT Enterprise?
Is Microsoft Copilot private for company data?
Does Claude use my business data for training?
Is Gemini for Google Workspace private?
What is the most private way to use GPT or Claude models?
How do I set up a private LLM?
Is an LLM private?
Is self-hosting an open-source LLM worth it for a business?
Can businesses use Llama and Mistral models commercially?
What is RAG in AI?
Can a private AI only show staff the files they are allowed to see?
Is RAG better than training a model on our own data?
Is there a HIPAA compliant AI?
Do UK companies need a DPIA before using AI on personal data?
Can Australian businesses put customer data into ChatGPT?
What does the EU AI Act require for a private AI assistant?
How much does a private AI system cost to build?
How much do business AI seats cost per user?
Which private AI option is best for a small business?
Bhavesh Barot
Founder & CEO
Founder & CEO of FactoryJet, which he started as a B2B ecommerce marketplace. FactoryJet has served 500+ businesses and builds online stores, B2B portals, AI agents and AI search.



