ChatGPT, Microsoft Copilot, Claude and Gemini: privacy for business use compared

Which differences between ChatGPT, Copilot, Claude and Gemini matter for privacy?

The tier matters most. All four providers separate consumer products from business products, and the promises on training, contracts and data location differ more between those two groups than between the providers. An organisation that processes personal data from its business will therefore usually end up on a business tier. Which one is then a question of data location, administration features and price.

This page has one job: it puts the contract position of four widely used assistants side by side. For ChatGPT in depth, see Is ChatGPT GDPR compliant?. The table reflects 13 November 2026 and reports only what the providers write on their own pages. If you are drafting rules for staff, the 12 rules for using AI at work are the companion piece.

What does the comparison look like in one table?

The table separates consumer tiers from business tiers. Each cell names the source that supports it; the numbers point to the source list at the end. Where the pages we checked say nothing, the cell says so instead of assuming.

One reading note: “no training” does not mean “no processing”. Business tiers also process your inputs to answer you and store them under their retention rules. A DPA provides for this to happen only on your instructions (Article 28(3)(a) GDPR)^[14], but it does not replace a legal basis.

Why are consumer tiers the weaker starting point for a company?

Because the contract you need for personal data is missing, and because training often runs through settings that individual employees control. The German data protection authorities recommend that employers provide work devices and accounts so staff do not have to use private ones, and that accounts are set so that inputs are not used for training^[12].

An opt-out is not a substitute. On ChatGPT, switching off data sharing does not delete saved chats, and a thumbs rating can send the whole conversation to training again^[3]. On Claude, a thumbs rating stores the conversation for up to five years^[16]. On Gemini Apps, a subset of chats is read by reviewers and kept for up to three years^[11]. These are properties of consumer products, not defects: they are built for private use.

How do the business tiers differ from each other?

Mainly in data location and in how far the promises reach. All four name a contract and no default training; beyond that, four differences count when you choose.

• Data location in Europe. Microsoft describes an EU Data Boundary but excludes Anthropic models and web search queries^[4][5]. OpenAI offers European storage to eligible Enterprise, Edu, Healthcare and API customers; Business is not listed^[1]. Anthropic names Europe through cloud partners^[9]. For Google, some features lack a stated regional control^[10].
• Access to your documents. Microsoft 365 Copilot only surfaces content through Microsoft Graph that the user has at least view permission for. That is an advantage in daily work and a risk if permissions are too wide^[4].
• Administration. Anthropic lets owners switch off the rating feature^[7]. Microsoft says interaction data is stored according to the commitments that apply to the tenant^[4].
• Contract mechanics. Anthropic incorporates its DPA automatically, while OpenAI asks you to request one through a form^[8][2]. Ask to see the agreement that applies to you and file it.

Where does the data go, and how much of that can you control?

With all four assistants the same limit applies: after you send a prompt, you cannot see where it is processed, copied or stored. The table reports what the providers commit to; it cannot show you that a commitment is kept in every case, and each provider documents exceptions of its own. Microsoft excludes web search queries and Anthropic models from the EU Data Boundary^[4][5]. OpenAI’s residency is storage at rest for eligible customers, and Business is not listed^[1]. Anthropic names Europe only via its cloud partners^[9]. Google’s page describes no regional control for some Workspace features, and in the consumer Gemini app a subset of chats is read by reviewers and kept for up to three years^[10][11]. In every case a processor may engage further processors under conditions, so you deal with a chain^[14].

These are documented limits, not evidence that anyone breaks a promise. But choosing the best tier does not close the gap, and it cannot be undone: what has been sent cannot be recalled, and deleting a chat or changing a setting acts on the future. The one control you hold yourself is what leaves your device. Sending less and replacing identifying details is data minimisation^[15]. The detail behind this is in where your data goes when you use an AI assistant.

What does the table not decide?

The table answers one question: what does the provider say about its plan? Four questions stay with you whichever provider and tier you choose.

Legal basis

Every processing step with personal data needs a legal basis, say the German data protection authorities^[12]. The Baden-Württemberg authority describes a “double door” model: provider and user each need their own basis, and a contract with the data subject generally does not cover third parties’ data typed into prompts^[13]. That is an authority’s view, not a ruling of the highest courts. A DPA does not answer this question.

Information for the people concerned

If customer, applicant or employee data goes into a tool, the circle of recipients you must tell people about can change. That belongs in your privacy notices and your record of processing activities. The table changes nothing there.

Professional secrecy

Some professions are bound by confidentiality rules that go further than the GDPR. In Germany, for example, lawyers, tax advisers and doctors can face criminal liability for unauthorised disclosure of secrets under § 203 of the criminal code^[17]. A DPA does not satisfy this by itself. Check your professional body’s guidance before using any assistant for client matters.

Employee representation

Where works councils or similar bodies exist, involve them early. In Germany, for example, the DSK recommends considering a works or service agreement and notes that co-determination applies in some cases^[12].

How do you choose among the four assistants?

Start from the data and your existing environment, not from the brand name. The decision aid below is a sequence of questions, not legal advice.

Two notes. First, running several assistants in parallel is costly to administer, because each needs a contract, a record entry, settings and training. Second, even on the best business tier the volume question remains: data minimisation means sending only what the purpose needs^[15]. Names, numbers and addresses that the prompt does not need are better replaced before sending. One tool for this is the Ainym browser extension, in closed beta with pilot teams, which detects personal data on your own device and shows every match for you to confirm before anything is sent; no automatic detection finds everything, so a human check stays necessary. The method is explained in anonymise data before ChatGPT.

What should you check before deciding?

Next step: take your three most common AI tasks, note which tier each would need, and compare with the table. For staff-facing rules, use the 12 rules; for ChatGPT contract detail, Is ChatGPT GDPR compliant?. This text is general information, not legal advice.

This article is general information and not legal advice. It does not replace an assessment of your own case by your data protection officer or a lawyer. Last updated: 13 November 2026.