The 30-second answer
Two things decide everything: the plan you subscribe to, and where processing takes place. The first is settled in five minutes; the second can be a deal-breaker.
On individual plans, model training on your content is turned off user by user. On the Team and Enterprise plans, it is off by default. However, direct access to Claude does not offer a European processing region: if your requirements demand one, the documented route goes through a third-party cloud provider.
Time for deleted conversations to be erased from storage systems, on consumer plans.
Retention in de-identified form if the user left training turned on.
European regions where Claude models are available through a third-party cloud provider: Paris, Frankfurt, Ireland, Stockholm.
Information checked on August 31, 2026. This article describes the state of the offering, not legal advice: have your data protection impact assessment validated by your counsel.
The difference nobody looks at: opt-out versus default
This is the biggest gap between the plans, and it does not show in daily use: the interface is identical. You can only see it in the plan comparison table.
The operational consequence is clear. As long as your employees use personal subscriptions, you can neither check nor enforce this setting — and you have no visibility into what has been submitted. Moving to an organization plan is therefore less a purchase of features than a way to regain control. That is the argument to make to leadership if they find the budget high.
How long content is retained
On consumer plans, the vendor documents the following periods. A deleted conversation disappears immediately from the visible history, and from storage systems within thirty days. If the user left training turned on, the content concerned is retained in de-identified form for up to five years. Conversations flagged for violating the usage policy are retained for up to two years, and the associated safety scores for up to seven years. Feedback you send voluntarily is kept for five years.
These periods apply to consumer products: the terms for business plans are covered in separate documentation, and the Enterprise plan adds configurable retention controls. If your industry requires a specific retention period, that is the point to get written into the contract in black and white, not to infer from a help page.
The cheapest reflex
No retention rule protects data that should never have been submitted. The most effective measure is upstream: a policy that names precisely the categories forbidden in conversations. It costs one page and avoids most of the risk.
The hard point: where processing takes place
This is the topic where you need to be precise, because it sometimes decides whether the project is feasible at all.
Direct access to Claude — the app, the Team and Enterprise plans — does not currently offer a European processing region. The only geographic choice highlighted in the commercial offering is US-only processing, for workloads that require it. We did not find any documentation on the vendor’s help pages announcing an equivalent European region.
The route that exists today goes through a third-party provider: Claude models are available from the European regions of the major cloud providers — notably Paris, Frankfurt, Dublin and Stockholm — as well as through the European regions of a second provider. This route serves a different use: it covers an application integration built by your teams, not the chat window rolled out to your employees.
The practical consequence, and it is fundamental
If your requirements mandate processing within the European Union — because a public-sector client demands it, or because your impact assessment concluded so — standard conversational use does not meet that requirement, and no setting will change that. You then need either to restrict the categories of data submitted, or to go through an integration hosted in Europe. Better to settle this before buying licenses than after.
This reasoning is the same as for any outsourced processing: we detailed it in the GDPR guide for small businesses. Obligations specific to AI systems fall under another law, covered in our EU AI Act compliance obligations.
Connectors: the blind spot in access rights
This is the fastest-growing risk, and the least discussed. Connecting the assistant to your email, your document repository or your management software transforms its use — and moves the security question onto ground the policy does not cover.
A connector inherits the rights of the person who authorizes it. If an employee has, through years of accumulated permissions, read access to a leadership folder they never open, that folder becomes searchable in natural language. Nothing was hacked: a dormant right became an active one, because the barrier was no longer the difficulty of searching.
Three precautions, in order of effectiveness:
- Clean up access rights before connecting, not after. Plugging in an assistant is the best excuse you will ever get to audit permissions accumulated over ten years. Seize it.
- Keep control over which connectors are allowed. Organization plans let an administrator decide which connectors, remote and local, can be used. Letting everyone plug in whatever they want amounts to opening an unsupervised exit channel.
- Treat retrieved content as untrusted. A document or web page pulled in may contain instructions designed to hijack the assistant. The principle to teach is simple: whatever comes from an external source is data to examine, never an instruction to execute.
What to ask for before you sign
If your company answers vendor questionnaires, or has its own suppliers fill them in, here are the documents to request. They exist: the vendor provides a trust center that includes, in particular, a SOC 2 Type 2 report and certification to ISO 27001.
The distinction not to blur
A vendor certification attests to its own security. It says nothing about yours, or about the compliance of your use. A perfectly certified supplier does not stop an employee from pasting a customer file into a conversation. The two topics are independent, and only the second depends on you.
Shadow AI: what is already happening without you
Start from a realistic assumption: in a small business with more than twenty people, usage began before the decision. Employees are already using a generative assistant on company documents, from a personal account, without anyone having authorized or forbidden it.
This situation is riskier than either alternative — a supervised rollout, or an explicit ban — because it combines their drawbacks: data leaves the company, and the company gets no organized benefit from it.
How to find out without launching a witch hunt. An open question in a team meeting, asked without threat, gives a more reliable result than analyzing network logs: “who is already using it, and for what?”. The answers come, as long as the question is not seen as a trap. The conversation that follows gives you both the current picture and the list of use cases to cover first.
The reasoning to present to leadership
The budget for an organization plan is not weighed against “spending nothing”: it is weighed against uncontrolled use that is already costing you, in the form of risk. Framed this way, the trade-off changes nature — and that is usually the argument that unlocks the decision.
The usage policy, on one page
This is the deliverable to produce before access is rolled out to anyone. Four sections are enough, and they fit on one side of a page.
What never gets pasted. Name categories, not principles. “Personal data of identifiable customers”, “contract documents marked confidential”, “credentials, keys and technical secrets”, “health data”, “information covered by an active non-disclosure agreement”. A named prohibition is respected; a general principle is not.
What must be checked before it goes out. Every figure, every date, every regulatory reference, every standard name and every quotation. This rule is not a matter of style: it is what keeps an error from ending up in a document handed to a customer.
Who decides when in doubt. A named person who can be reached. Without an identified contact, doubt gets resolved by usage — that is, in favor of whatever is easiest.
What is disclosed to the customer. Specify whether the company mentions the use of AI in its deliverables, and in which cases. It is as much a commercial question as a regulatory one, and it should be settled once and for all rather than case by case.
The order of operations
Policy first, organization plan second, training third. This order is not cosmetic: the policy is the only one of the three steps whose absence causes irreversible damage, since it concerns data that would already have left the company.
For the rollout itself, see our guide to getting started with Claude for business. To place this work within an overall security policy, our complete guide to cybersecurity for businesses.