Private AI for Business: 9 Ways to Use AI on Your Own Data, Compared (2026)

In the US, "private ai" is searched about 40,500 times a month and "private ai for business" 2,400 times, with advertisers paying $45.23 a click for the second; "hipaa compliant ai" adds 6,600 searches (DataForSEO, 22 October 2026). The buyers behind those searches are usually an operations, IT or compliance lead who has found staff pasting customer data into personal AI accounts and needs a safe option by next quarter. They want secure AI for business, and most private AI solutions on the market describe one route only. This guide compares nine options on the same five tests, from ChatGPT Team (now ChatGPT Business) to on premise AI.

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.

1. Training use. Are your prompts, files or answers used to train or improve the vendor's models? By default, and after any setting change?
2. 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?
3. 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.
4. Encryption. Is data encrypted at rest and in transit, and can you hold your own encryption keys (customer-managed keys)?
5. 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 is the one most companies fail, and it is not the vendor's fault. 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)

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 pagesays: "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 pagesays 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 (Team and 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. Two details matter: 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 FAQsays 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 pagesays 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. One trap: a global endpoint carries no residency promise.

Direct APIs sit between seats and cloud. The OpenAI APIdoes not train on API data, keeps abuse-monitoring logs up to 30 days, and offers regional storage including the UK and Australia (non-US regions need OpenAI's approval for zero data retention or modified abuse monitoring, and newer models carry a 10% uplift 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: self-hosted open models such as Llama and 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, which is why "on premise llm" and "self hosted llm" draw 2,900 and 3,600 US searches a month. In exchange you own the hardware, security patches, upgrades and evaluation, and open models often trail the best API models on hard reasoning.

• Llama 4 (Meta). The licenceallows 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 modelsare "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: how private AI answers from your own documents

RAG (retrieval augmented generation) is how an AI answers from your contracts, policies, product sheets or tickets without being retrained on them. "rag" alone draws 74,000 US searches a month, mostly from people learning the term. Here is the business version, in five steps:

1. Collect. Pull documents from SharePoint, Google Drive, your ERP, CRM or help desk, and drop duplicates and stale versions.
2. Index. Split each document into short passages and store a searchable version (a vector) in a database inside your own cloud account.
3. Retrieve, with permissions. When someone asks, fetch only the passages that person is allowed to see. This is where most DIY builds leak.
4. Answer. Send the question and those few passages to the model you chose above. The model sees snippets, not your whole library.
5. 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).

Compliance 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. AI data privacy law differs by country, and GDPR AI questions in the UK and EU turn on the same two points.

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, 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.

Which option fits your company size

Private AI cost: what a build costs (market ranges)

Seats are priced per user by each vendor. A custom private assistant over your documents is a build project. These are published market ranges, not our prices:

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"

1. They cannot link to the clause that says your data is not used for training.
2. "Private" means a password on a shared consumer account.
3. No answer on retention, or retention only the vendor can change.
4. A global or unspecified region when your rules need the UK, EU or Australia.
5. A BAA promised for a product the model maker does not list as covered.
6. The AI sees every document regardless of who is asking.
7. No log of questions, sources and answers you can export.
8. 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

1. Which model provider and which product (seat, API, or cloud platform) will process our data?
2. Where is data stored and processed, and can both be pinned to our region?
3. How long are prompts and answers kept, and can we turn retention off?
4. Who holds the encryption keys?
5. How are our existing file permissions enforced in answers?
6. Which agreement covers us (BAA, Article 28 contract, IDTA), and for which features?
7. 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.

What else ranks for "private ai for business", and how this guide differs

These are the business sites on Google's first page for "private ai for business" in the US when we checked on 22 October 2026 (DataForSEO), with the title Google showed. We list them so you can compare; we have not reviewed their services, and the list leaves out forums.

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.

• 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
• Our own measurement: US search volumes, click prices, People Also Ask questions and the first-page results, DataForSEO, 22 October 2026