It is the first question a Chief Information Security Officer (CISO) asks when shown a ChatGPT license to roll out to 500 employees: “before I sign, I need to know where the data my teams type into the tool is going.”

The right question. Because depending on the model and the plan, the answer ranges from “nowhere, it’s encrypted and deleted at the end of the session” to “into the model’s collective memory, and it won’t come back”.

The problem doesn’t come from the tools. It comes from the fact that five mainstream AIs are now present in most companies (ChatGPT, Claude, Gemini, Mistral, Copilot), each with three or four subscription tiers, and rules that change every six months.

Data confidentiality and AI: the 5 security criteria to examine

Is your data used to train the model?

This is the question that changes everything. If it is, what you type goes into what the model learns, and a later user could potentially find a trace of it in a generated answer. On most consumer plans (free or Plus/Pro), training is on by default: you have to turn it off manually in the settings. On business plans (Business, Team, Enterprise), the rule is reversed: no training on your data, and this guarantee is written into the contract.

How long is it kept?

Even without training, your data sits somewhere. Retention periods range from a few hours (temporary mode) to several years. Business plans often let you configure this retention, or switch it off completely with a no-retention agreement (Zero Data Retention).

Where is it stored?

The United States? Europe? Your own servers? For a European company subject to the GDPR, this is no detail. Some vendors now offer regional hosting in Europe; others stay in the United States by default, with standard contractual clauses for the transfer.

Which security certifications?

Three standards carry authority: SOC 2 Type II (an independent audit of security controls), ISO 27001 (information security management) and ISO 42001 (the most recent, dedicated to AI management systems). HIPAA comes on top for healthcare. A serious vendor publishes them on a compliance portal, accessible under a confidentiality agreement.

What governance to manage use across the company?

This is what lets a CIO or a CISO stay in control: single sign-on (SSO) through your company directory, audit logs (who did what, and when), a data processing agreement (DPA) you can sign as the legal basis under the GDPR, admin roles, access controls by team, and automatic creation and deletion of accounts from your HR directory (SCIM). Without these tools, you are handing out a license you can’t manage.

To go further, find out what European regulation requires regarding data.

Comparing the main AIs

ChatGPT (OpenAI)

On ChatGPT Business, Enterprise, Edu and the API, OpenAI commits by contract not to use customer data to train its models, and data is encrypted in transit and at rest. Enterprise goes further: you control the retention period, and OpenAI has passed a SOC 2 audit.

On governance, the Enterprise plan offers the full toolkit you would expect: single sign-on that works with your company directory, automatic creation and deletion of user accounts, admin roles, programmatic access to audit logs, and the option of using your own encryption keys. Data residency is available in ten regions: Europe, the United Kingdom, the United States, Canada, Japan, South Korea, Singapore, India, Australia and the United Arab Emirates.

Point to watch: following a court order in the New York Times v. OpenAI lawsuit, some past conversations (April to September 2025) are kept for the duration of the litigation for ChatGPT Free, Plus, Pro, Team and API subscribers. ChatGPT Enterprise, however, is explicitly excluded, as are conversations originating in the EEA, Switzerland and the United Kingdom.

Claude (Anthropic)

On the Claude Team and Enterprise plans (Claude for Work) and the Anthropic API, Anthropic does not reuse your conversations to train its models. Data is encrypted in transit and at rest, and by default no Anthropic employee has access to your conversations, unless you explicitly agree to it or during an investigation into a breach of the usage policy.

On certifications, Anthropic publishes four on its compliance portal: SOC 2 Type I & II, ISO 27001:2022 (information security), ISO/IEC 42001:2023 (AI management systems) and a HIPAA-compliant configuration for healthcare. For the most sensitive data, the Zero Data Retention option is available through a dedicated agreement.

The point to watch for European companies: Anthropic’s direct API doesn’t yet offer hosting exclusively in Europe (only the United States and “global” regions are available). To host your data in Europe, there are two documented solutions: going through the cloud platforms of Amazon (AWS Bedrock) or Google (Vertex AI), which offer the Claude models in their European regions.

Gemini (Google)

On Google Workspace Business/Enterprise (Google’s collaboration suite for businesses) and on Vertex AI (its cloud platform dedicated to AI), your content isn’t used to train the public models and isn’t examined by human reviewers outside your domain without your permission. Your content stays within your organization. On certifications, the official list is long: SOC 1/2/3, ISO 9001, ISO 27001, 27017, 27018, 27701, ISO/IEC 42001 and FedRAMP High. Regional hosting is possible (in EU regions in particular).

A built-in advantage if you already work in the Workspace ecosystem: Gemini inherits your existing Drive and Gmail permissions, and administration goes through the same console. Point to watch: for consumer accounts (free Gemini, personal Gmail account), the rules are different: activity is kept for 18 months by default, and you can change this setting or turn activity off. On a business Workspace with a signed data processing agreement, the processing is governed by contract.

Mistral (Le Chat)

The only serious European alternative. Mistral AI is a French company headquartered in Paris. The data processed on Le Chat and on La Plateforme is hosted by several cloud providers, including Azure in Sweden and Google Cloud Platform in Ireland, which places the offer within the scope of the GDPR, with on-premises deployment possible for highly regulated sectors.

Choose your plan carefully: on the free Le Chat, conversations may be used to train the models (you can object by contacting [email protected]). On Le Chat Pro, you object directly from your account. On Le Chat Team, Le Chat Enterprise and the paid La Plateforme API, your data is not used for training by default. On La Plateforme, retention is 30 days (for moderation), with a Zero Data Retention option available.

Copilot (Microsoft 365)

No training on your data: it is contractual and explicit. “Prompts, responses, and data accessed through Microsoft Graph aren’t used to train foundation LLMs.” A key advantage for European users: the data processed by Copilot stays hosted in Europe.

The governance is the Microsoft 365 governance you already know: Microsoft Entra (Microsoft’s identity service) for single sign-on, Microsoft Purview (Microsoft’s data governance tool) for retention policies and sensitivity labels, data kept in a space dedicated to your company and separated from other customers, and full encryption of data at rest and in transit. Published certifications: GDPR, ISO 27001, ISO/IEC 42001, HIPAA.

Note: since January 2026, Microsoft has been adding Anthropic models to Copilot. But Microsoft explicitly states that these models are outside the EU Data Boundary and outside its in-country LLM processing commitments: turn them on knowingly if your regulatory framework requires it.

With security settled, you still have to choose your AI. Performance, price, sovereignty, ecosystem: Claude, ChatGPT, Gemini, Mistral or Copilot: which AI should your company pick?

How to keep your data safe when your company uses AI

Three reflexes:

  1. No employee should use a personal version (free or Plus) to handle company data. The security gap with a business plan is considerable, and the extra cost (a few dozen euros per user per month) is minimal compared with the risk of leaks or data exfiltration flagged by regulators.
  2. Sign a DPA (Data Processing Agreement) with the vendor you choose, even if your company has only 10 people. It is the basic requirement of the GDPR, and the first document a customer or an auditor will ask you for.
  3. Turn on single sign-on (SSO) and audit logs as soon as the tool goes live, without waiting six months. It is the only way to answer an auditor or a customer who asks who accessed which data.

Conclusion

The security of an AI tool at work is checked against five concrete criteria: training, retention, hosting, certifications and governance. And it is written into a DPA. The five major vendors have all published solid commitments on their business plans. The real issue is the personal accounts your employees use alongside them, without telling you.

The only way to stay in control is to give them an official alternative, governed and paid for by the company.

FAQ: your questions about data security and AI

Do AI vendors use my data to train their models?

It depends on the plan. On most consumer plans (free or Plus/Pro), training on your conversations is on by default and has to be turned off manually in the settings. On business plans (Business, Team, Enterprise), the rule is reversed: no training on your data, and this guarantee is written into the contract. The five major vendors (OpenAI, Anthropic, Google, Mistral, Microsoft) have published commitments to this effect.

Which criteria should you check before choosing an AI for your company?

Five criteria: training (is your data used to train the model?), retention (how long is it kept?), hosting (the United States, Europe or your own servers?), certifications (SOC 2 Type II, ISO 27001, ISO 42001) and governance (single sign-on, audit logs, a data processing agreement). All of it is written into a DPA signed with the vendor.

Can you use ChatGPT with company data?

Not with a personal version, free or Plus: the security gap with a business plan is considerable, for an extra cost of a few dozen euros per user per month. On ChatGPT Business, Enterprise, Edu and the API, OpenAI commits by contract not to use your data to train its models, and data is encrypted in transit and at rest.

How do you keep AI use secure in your company?

Three reflexes: forbid the use of personal versions for company data, sign a DPA (data processing agreement) with the vendor you choose, even for a 10-person company, and turn on single sign-on and audit logs as soon as the tool goes live. The only way to stay in control is to give employees an official alternative, governed and paid for by the company.