AI accounts and hosting: what really changes between the plans

The essentials — the product does not tell you where the data goes

Claude Code on a Max account and Claude Code connected to Amazon Bedrock in Europe are the same software. For your data, they are two different situations. Three questions make the difference:

Which contract? Consumer terms, or an enterprise contract.
Where does the data go? The place of processing, and the law that applies to the supplier.
Who can trace it? A log the company can use, or nothing.

Option	Contract	Training on your data	Location	Log	Level covered
Free or personal account	Consumer	Depends on the user’s setting	Not controlled	None for the company	Public
Enterprise account under contract	Enterprise	No, by default	Not guaranteed in Europe	Depends on the plan	Internal
Enterprise, Europe, logging	Enterprise	No, by default	Chosen EU region	Enabled	Confidential
Qualified European cloud	Enterprise	Check the contract	In Europe	Check the contract	Critical
Model on your premises	No third party	Not applicable	Your premises	Your own	Critical

The last column comes from the grid of the tool Which data can you give to an AI?. It is an IndustryHub calibration, not a standard.

Free or personal account

Plans: ChatGPT Free, Plus or Pro; Claude and Claude Code on Free, Pro, Max or Ultra.

Paying does not change the category. A Max subscription is expensive, but it remains an individual account under consumer terms.

Training. At OpenAI, conversations are used for training unless the user opts out. At Anthropic, it is an account setting. If accepted, data is retained for 5 years. If refused, 30 days.

What is missing. Turning off training settles one point only. The company still has no enterprise contract, no account administration and no log. The setting depends on each employee.

Reserve for: public data.

Enterprise account under contract

Plans: ChatGPT Business or Enterprise; Claude and Claude Code on Team, Enterprise or API.

What protects you. OpenAI and Anthropic do not train their models on this data by default. At Anthropic, the data processing addendum is automatically part of the commercial terms. Standard retention there is 30 days. At OpenAI, on ChatGPT Enterprise, administrators set the duration, and a deleted conversation is removed within 30 days unless the law requires otherwise.

What is missing.

Location. Anthropic states that by default it may route traffic to the US, Europe, Asia and Australia, with storage in the US, unless otherwise agreed.
Logging, depending on the plan. On Claude, audit logs are reserved for Enterprise: 180 days exportable, without conversation content. Team does not have them.
Voluntary sharing. Feedback sent to the supplier can fall outside this frame. In Claude Code, /feedback sends the conversation, which is retained for 5 years.

Reserve for: internal data.

Enterprise account, processing in Europe, access logging

Three cumulative conditions: an enterprise contract, processing in Europe, logging enabled. If one is missing, the option falls back to the previous level.

Access	What the supplier says	To check
Claude Code through Amazon Bedrock	Model providers have access neither to prompts nor to completions	`eu-*` region, `eu.` prefix, invocation logging enabled
Claude Code through Google Cloud	`eu` location available	Do not stay on `global`, which guarantees no location
ChatGPT Enterprise or API, data residency in Europe	Content stored in the chosen region, for eligible customers	Place of computation, as well as place of storage; ChatGPT Business is not mentioned
Microsoft 365 Copilot	EU traffic stays within its EU boundary	Anthropic models are currently excluded
Mistral Vibe Enterprise	Hosted in the EU by default, training opted out from the start	Features involving a transfer, to be turned off at organisation level

Three settings are easily lost on Bedrock. With no region set, Claude Code defaults to us-east-1. The eu. prefix is a preference, not a guarantee. And invocation logging is disabled by default: it has to be enabled. In addition, some recent models, such as Claude Fable 5, require AWS to be able to retain exchanges for up to 30 days and review them: check this condition for the chosen model.

A copilot sees what the user sees. Microsoft 365 Copilot surfaces any content the person is entitled to view. An over-broad share becomes reachable through a simple question.

Limit. Hosting in Europe settles the location, not the law. A US supplier remains subject to the CLOUD Act, even for data stored abroad. That is why this option stops at the Confidential level.

Qualified European cloud

In France, the reference is SecNumCloud, the ANSSI qualification. Its framework requires protection against non-European laws. That is what sets it apart from simple hosting in Europe.

The qualification covers a service, not a brand. An AI platform hosted by a qualified provider is not itself qualified. Check that the AI service itself appears in the ANSSI catalogue, updated at least once a month.

In Germany, the BSI’s C5 is of a different kind. There is no C5 certificate, only an audit attestation covering specific services and regions. Read the report: a C5 attestation does not read like a SecNumCloud qualification.

What is missing. A narrower model catalogue and a higher cost. Dual-use items remain a separate case: export control, then a local model.

Model on your premises

Principle. An open-weight model runs on your servers. No AI supplier receives the data.

Claude Code can be connected to it. Ollama exposes a compatible interface: point ANTHROPIC_BASE_URL to http://localhost:11434. Ollama recommends a context window of 64,000 tokens or more. The software remains Claude Code, the model is no longer Claude.

“Local” has to be checked.

Ollama also offers models in its own cloud: choose a local model.
Claude Code keeps a few outgoing connections. The WebFetch safety check sends the hostname to api.anthropic.com, unless skipWebFetchPreflight turns it off. CLAUDE_CODE_DISABLE_NONESSENTIAL_TRAFFIC turns off non-essential traffic.

What is missing. A compute server to size, in-house skills, and often models that lag on the hardest tasks. Licences vary from one model to another, even at the same publisher: read the licence of the chosen model before commercial use.

What remains your responsibility. Access rights, backups, logs. An internal leak remains possible.

One tool, several answers

For an electrical schematic of a routine installation, classified Confidential:

Claude Code with…	Option	Answer
Pro or Max account	Personal account	No
Team, Enterprise or API account	Enterprise under contract	No: location not guaranteed
Amazon Bedrock or Google Cloud, EU region, logging enabled	Enterprise, Europe, logging	Yes
Ollama and a local model	Model on your premises	Yes, subject to the settings above

These “no” answers come from our deliberately cautious grid. A company may judge an Enterprise account sufficient for its routine schematics: that is a decision to take and write down, not to leave to each employee.

Check what you are really using

The account is in the company’s name, not an employee’s
The contract and its data processing addendum are signed or accepted, and archived
Training is turned off for the organisation, and voluntary sharing is controlled
The place of processing is written in the contract, not just the place of storage
In Claude Code, /status shows the expected provider and region
Logging is enabled, and someone knows how to read it
File access rights are reviewed before connecting an assistant

Limits

These plans change often: this text reflects the supplier pages consulted on 31 October 2026. The signed contract prevails. Prices are not covered, as they move too quickly. This article is not legal advice.