Private AI

Private AI for business: what it is and how to check it

Private AI means your organisation controls where its documents, prompts and answers go. What that means in practice, the ways to deploy it, and what to ask a vendor.

By Marella AI · Updated 23 Oct 2026

Short answer

Private AI is AI where your organisation controls what happens to its data: where documents and prompts are stored, which model processes them, who can see them, and whether they are used for training. It is a description, not a standard or a certification, so ask how each of those is handled in the configuration you would buy.

What makes AI private

Five questions decide it. If a vendor cannot answer all five in writing for the configuration you would buy, the word private is not telling you much.

Where your documents, the text extracted from them, your prompts and the answers are stored, and in which country.
Which model processes them, who runs that model, and whether its provider keeps anything.
Whether any of it is used to train a model, according to the contract rather than the marketing page.
Who can see it: your own staff by role, the vendor's support team, and any subprocessors.
What is logged, for how long, and who can read the logs.

Private AI and public AI assistants

General-purpose assistants such as ChatGPT and Microsoft Copilot run on their provider's infrastructure under the provider's terms. Their business plans add contractual limits: OpenAI says it does not train on business data by default, and Microsoft says prompts from work accounts are not used to train its foundation models. Your prompts and files still go to that provider, though, and OpenAI's terms let its staff access business conversations for limited purposes, such as investigating abuse or where the law requires it.

For a lot of work that is acceptable. For client files, health records and regulated data, organisations usually want processing to stay inside a boundary they have set and can check.

Is ChatGPT private? Plan by plan

The ways to deploy it

Private AI is not one architecture. These are the usual routes, from least to most work for your own team. Where the model runs is a separate choice from where the software runs.

Managed private deployment: the vendor runs it, with storage, model provider and access agreed in writing.
Your own cloud account: the software runs in a cloud environment you control, under your existing agreements.
On-premise or self-hosted: the software runs on servers you operate, and you take on patching, backups and monitoring.
Local models: the model runs on hardware you control, so prompts are not sent to an outside model provider.

On-premise and self-hosted AI

What private AI does not fix

Keeping data private does not make the answers correct. A private model can still misread a scan, miss the passage that matters or state something its source does not say. For anything a person will rely on, the answer needs to show where it came from, so someone can check it.

How cited answers work

How Marella does it

Marella is private AI for an organisation's own documents. It answers questions over the documents you approve, cites the passage behind each answer, and fact-checks answers against their sources, flagging inaccurate citations and unsupported claims.

It can be managed by us, deployed into a cloud environment you control after an architecture review, or run from a self-hosted package on infrastructure you choose. The model provider is configurable. Access is limited to your organisation and set by role and department, and our legal terms state that customer data is not used to train Marella or its configured AI providers.

Compare the deployment options

Questions to ask any private AI vendor

Ask these before a demo, and expect the answers in writing for the configuration you would buy.

Draw the data flow for our configuration, including logs, backups and support access.
Name every subprocessor that would see our data, and where it is.
Show us the contract terms on training and retention.
Tell us which model runs, and whether we can change it.
Explain what happens to our data when we leave.
Run it on documents like ours before we sign.

Enterprise AI data-flow checklist

Public AI assistants compared with a private deployment

Typical patterns. The contract and configuration you sign decide the details.

Decision factor	Public AI assistant	Private AI deployment
Where it runs	The provider's infrastructure.	A boundary you agree: managed, your cloud or your own servers.
Training	Personal plans may train on conversations unless you opt out. Business plans say they do not train on your data.	Set in your contract for the deployment.
Who can see it	Provider staff for limited purposes, under the provider's terms.	Your users by role, plus any support access you approve.
Which model	The provider's choice.	Chosen and written down for your deployment.
What it answers from	Whatever a user pastes or uploads.	The documents you connect and approve.

Scroll to see more

Caveats

B1Private is a description, not a certification. Check each data path by name.
B2Self-hosted does not automatically mean on-premise, air-gapped or local models.
B3OpenAI and Microsoft terms were checked on 2 October 2026 and change often.
B4Marella's deployment details are agreed at architecture review, not assumed from this page.

Frequently asked questions

What is private AI?

Private AI is AI where your organisation controls where its data is stored, which model processes it, who can see it and whether it is used for training. It covers managed private deployments, deployments in your own cloud account and fully self-hosted setups.

Is there a private alternative to ChatGPT for business?

Yes. ChatGPT's business plans add contractual limits, but your prompts and files are still processed by OpenAI. A private deployment keeps documents, prompts and answers within a boundary you agree in writing. Marella is one: it answers from your organisation's own documents and cites the passage behind each answer.

Is ChatGPT Enterprise private AI?

It is a business service with contractual protections, and OpenAI says it does not train on business data by default. Your data is still processed on OpenAI's infrastructure under OpenAI's terms. Whether that is private enough depends on the data and on your obligations.

Is private AI the same as on-premise AI?

No. On-premise is one way to get it, with the software on servers you operate. A managed deployment, or one in your own cloud account, can also be private if storage, model provider and access are agreed and written down.

Can private AI run without an internet connection?

Only if the software and the models both run on your own hardware and updates arrive by a route you control. Most private deployments still call an outside model provider, so ask for the network diagram.

Is private AI more secure?

Not automatically. It moves control, and the work that comes with it, to you or to a vendor whose controls you have checked. A badly run private server is less safe than a well-run managed service.

Plan an evaluation

See Marella on your own documents

Tell us about the workflow and your data rules, not the confidential details. We agree how documents are handled before any are shared.

Book a demo Explore the platform