Private AI is a set of choices, not a product. Compare five privacy levels, the three layers every setup has, and what ChatGPT's own terms say.

Private AI is an approach to using artificial intelligence in which your organization controls where the model runs, where your data and knowledge are stored, and who can access both. It is not a single product: it ranges from a business AI plan with no-training terms to an open-weight model running on hardware you own.
Key takeaways
In consumer apps, "private AI" usually means a chat app that promises not to log or train on your conversations. This guide covers the team meaning: a setup you can explain and defend in a security review.
You will find the five levels of private AI, the three layers every private AI system has, what ChatGPT's data terms say as of October 2026, and a checklist for deciding which level your team needs.
| Public AI (consumer tier) | Private AI | |
|---|---|---|
| Data use | Content may be used to improve models, depending on settings | Contractual no-training terms, or no third party involved |
| Model location | Vendor's shared infrastructure | Vendor's business service, your cloud account, or your own servers |
| Access control | Individual accounts | Organization-managed identity, roles, and permissions |
| Cost profile | Per-user subscription, low entry cost | Higher seat prices, cloud usage fees, or hardware and operations cost |
| Effort | Sign up and start | From an admin setup to a dedicated infrastructure project |
The right column is wide on purpose. "Private" covers several different architectures, so it helps to treat it as a spectrum.
There are five common levels. Each one moves more control to your organization and adds more work. The goal is to match the level to the sensitivity of the data you plan to use, not to pick the highest level everywhere.
The private AI spectrum: control and effort both rise from left to right.
| Level | Example | Model runs on | Who controls data use | Typical effort |
|---|---|---|---|---|
| 1. Consumer chat apps | ChatGPT Free, Plus, Pro | Vendor infrastructure | Individual user settings | None |
| 2. Business tiers of public AI | ChatGPT Business or Enterprise | Vendor infrastructure | Business contract | Low |
| 3. Dedicated cloud deployments | Azure OpenAI in Microsoft Foundry, Amazon Bedrock | Your cloud provider | Your cloud agreement | Medium |
| 4. Self-hosted open-weight models | An open-weight model served with vLLM | Your servers or cloud account | You | High |
| 5. Isolated or air-gapped | Level 4 with no internet connection | Your isolated hardware | You | Very high |
What it is: Free or individual paid plans of public AI assistants.
What stays private: Very little by default. OpenAI states that for its services for individuals (Free, Plus, Pro), it may use your content to train its models. Users can opt out by turning off "Improve the model for everyone" in their data controls.
What does not: Prompts and files are processed and stored on the vendor's infrastructure under consumer terms. Your company does not manage the accounts.
Typical owner: Individual employees, often without IT involvement.
Typical effort: None. That is also the risk: this is where unmanaged "shadow AI" use starts.
What it is: Team or enterprise plans of the same assistants, with business terms.
What stays private: Training and, depending on the plan, retention. OpenAI's enterprise privacy page (updated January 8, 2026) says OpenAI does not train its models on business data by default for ChatGPT Business, ChatGPT Enterprise, ChatGPT Edu, and its API Platform. It also says Enterprise and Edu customers control how long their data is retained.
What does not: The model still runs on the vendor's infrastructure, and your data still leaves your network to be processed there.
Typical owner: IT or procurement, with a signed agreement.
Typical effort: Low. Buy seats, connect identity, set usage policies.
What it is: Models consumed through a cloud provider you already use, under your existing cloud agreement, such as Azure OpenAI in Microsoft Foundry or Amazon Bedrock.
What stays private: Microsoft states that prompts, completions, and embeddings sent to models sold by Azure are not available to other customers or to OpenAI, and are not used to train generative AI foundation models without your permission. AWS states that Amazon Bedrock runs models in deployment accounts that model providers cannot access, and that Bedrock traffic can stay on private paths through Amazon VPC and AWS PrivateLink.
What does not: The model is still a managed service. You rely on the cloud provider's controls and contracts rather than your own hardware.
Typical owner: Platform or cloud engineering.
Typical effort: Medium. You need cloud accounts, networking, quota requests, and an application layer on top.
What it is: You download an open-weight model and run it on servers you control, on premises or in your own cloud account, using an inference server such as vLLM.
What stays private: Prompts, outputs, and logs never go to a model vendor. You decide what is stored and for how long.
What does not: Nothing leaves by default, but you now own patching, scaling, monitoring, and model updates. Model licenses vary, so review each one before commercial use.
Typical owner: Infrastructure or ML engineering.
Typical effort: High: GPUs, capacity planning, and evaluation work. Our guide to private LLMs covers the deployment trade-offs.
What it is: Level 4 with no internet connection. Models, updates, and dependencies are moved in through controlled processes.
What stays private: Everything, as long as the isolation holds.
What does not: Convenience. You lose easy updates, hosted tools, and anything that expects an outbound connection.
Typical owner: Security and infrastructure teams in defense, critical infrastructure, or similar settings.
Typical effort: Very high.
Most teams use Level 2 or 3 for general work and reserve Level 4 or 5 for specific high-sensitivity workloads. Some consumer services also market "private AI" built on confidential computing, where prompts are processed inside hardware-isolated environments (for example, Privatemode). That is a useful option for individuals, but the team questions below still apply.
The spectrum describes where the model runs. But the model is only one part of the system. A working private AI setup has three layers, and each needs its own privacy decision.
Each layer can be changed without rebuilding the other two.
This is the large language model and the service that runs it. The questions: who operates the model, does the operator keep or train on your inputs, and where is it processed?
This is the material the AI works with: specs, SOPs, customer notes, contracts, roadmaps, and meeting notes. It is often the most sensitive part. Many teams secure the model carefully and then paste that material into a tool they do not control.
Keep the knowledge layer in a system where you control storage, backups, and retention. It should also be structured well enough for AI to retrieve the right context. Our article on building an LLM wiki covers how to organize internal knowledge for AI use.
This decides who can use the AI and which content it can see. Look for single sign-on with your identity provider, roles, and per-document permissions. An AI assistant should never let someone read a document through the AI that they could not open directly.
Separating the layers has a practical benefit: you can change one without rebuilding the others. You can switch model providers without moving your knowledge, or tighten access without changing the model.
It depends on the plan. Here is what OpenAI's own pages say as of October 2026.
Free, Plus, and Pro (individual plans). According to OpenAI's data usage article, OpenAI may use your content to train its models. You can opt out under Settings > Data controls by turning off "Improve the model for everyone." Even after opting out, if you give feedback on a response (such as a thumbs up or down), the entire conversation tied to that feedback may be used for training.
ChatGPT Business, Enterprise, Edu, and the API Platform. OpenAI's enterprise privacy page states that it does not train on business data by default, that customers own their inputs and outputs where allowed by law, and that Enterprise and Edu customers control retention. The same page describes encryption at rest (AES-256) and in transit (TLS 1.2+).
So the honest answer:
The same pattern applies to most major AI assistants: read the terms for your specific plan, not the general marketing page.
Match the level to the data, not to the team. One company can use several levels.
You likely need something beyond Level 1 if any of these apply:
If you checked one or two, a business tier (Level 2) with clear policies may be enough. If you checked several, or regulated data is involved, plan for Level 3 or higher and treat the knowledge layer as part of the project.
A practical first project:
For a full rollout plan, including vendor questions and policy templates, read private AI for business.
Here is one way to combine the three layers: a self-hosted workspace for knowledge and a swappable model layer.
AFFiNE is an open-source, local-first workspace that combines docs, whiteboards, and databases. You can self-host AFFiNE with the official Docker Compose setup. Your content then lives in your own PostgreSQL, Redis, and file storage (local disk, S3-compatible storage, or Cloudflare R2).
Because AFFiNE is local-first, docs and whiteboards stay editable offline in the desktop app and sync when the connection returns. This applies to editing, not to AI features.
AFFiNE is free to self-host up to 10 seats per workspace. Larger teams need a paid Team Plan or license, as described in the team license docs.
AI in AFFiNE is optional, and AFFiNE does not run models itself. The model layer stays swappable:
The settings are documented in the AFFiNE AI administration guide.
An important caveat: model behavior and feature coverage depend on the provider and model you connect. The docs already note that some features need specific providers; for example, transcription and workspace indexing require a server-stored Gemini key or an AFFiNE AI plan fallback. A custom endpoint is not guaranteed to support every AFFiNE AI feature, so test each model with your real tasks before rollout.
For agent workflows, AFFiNE also includes a workspace MCP server that lets AI clients you approve search and read workspace documents. Credentials are read-only by default; at the time of writing, write tools are limited to development and canary builds.
AFFiNE supports OIDC single sign-on, plus Google and GitHub OAuth. Multi-factor authentication is enforced at your identity provider. Inside the workspace, you assign workspace roles and set access levels per document.
The result: knowledge stays on your infrastructure, the model can move between Levels 2 and 4 as needs change, and access follows your existing identity provider. For larger rollouts and licensing, see AFFiNE for enterprise.
AFFiNE covers the knowledge layer. It is not the right choice if:
Private AI is a way of using AI where your organization controls where the model runs, where your data is stored, and who can access both. It ranges from business AI plans with no-training terms to self-hosted models on your own infrastructure.
Individual plans (Free, Plus, Pro) may use your content for training unless you opt out, according to OpenAI. OpenAI does not train on Business, Enterprise, Edu, or API data by default. In all cases, processing happens on OpenAI's infrastructure.
Usually, yes, but it depends on the level. Business tiers cost more per seat than consumer plans, dedicated cloud deployments add usage fees, and self-hosted models add hardware and operations costs. Matching the level to the data, instead of running everything at the highest level, keeps cost down.
Yes, if you self-host an open-weight model in an isolated environment, but this takes significant setup. Our guide to offline AI covers the options and limits.
A private LLM is a large language model that runs in an environment you control, or under terms that keep your data out of the vendor's training. It is the model layer of a private AI system, not the whole system.
Yes. You can self-host both models (with an inference server) and the tools around them, such as chat interfaces, knowledge bases, and workspaces. Our self-hosted AI guide compares the common stacks.
Self-host free up to 10 seats; Team $10/seat/month billed annually; offline Installable License via sales.