ChatGPT, Copilot, Claude: what to settle first

Almost every company already has them in use: ChatGPT, Microsoft Copilot, Claude. Clients rarely ask "May we?", but rather "What do we need to settle first?"

Personal account or business account

The decisive difference lies not in the tool but in the contract. Anyone using a free or personal account concludes the contract as an individual. The company receives no data processing agreement, and the inputs may be used to train the models unless the user opts out in the settings. That is what OpenAI states for the Free, Plus and Pro plans, and what Anthropic states in its consumer terms for Claude.

Business plans – ChatGPT Business, Enterprise and the API, Claude under the Commercial Terms, Microsoft Copilot in a company account – are set up differently: the provider offers a data processing agreement under Art 28 DSGVO, and inputs are not used for training by default. For customers in the EEA, the contracting party is OpenAI Ireland Ltd in the case of OpenAI and Anthropic Ireland, Limited in the case of Anthropic; Copilot is governed by the Data Protection Addendum and the Product Terms, under which Microsoft acts as processor and does not train its foundation models on inputs. These are the providers' own statements as of October 2026, to be checked against the contract text before signing; no recommendation is implied.

GDPR: the company remains the controller

As soon as an employee pastes a customer email, a job application or a set of minutes into an AI tool, the company processes personal data and needs a legal basis under Art 6 DSGVO, usually legitimate interest or performance of a contract. The privacy notice and the information given to staff must name the provider as a recipient, or at least the category of recipient (Art 13, 14 DSGVO), and the record of processing activities under Art 30 DSGVO needs an entry. A data protection impact assessment under Art 35 DSGVO is required only where a high risk is likely – for instance where individuals are systematically evaluated – and as a rule not for drafting texts.

Transfers to the United States, which all three sets of contract terms provide for, are permitted on the basis of the adequacy decision (Implementing Decision (EU) 2023/1795) where the provider is certified under the Data Privacy Framework, and otherwise on the basis of standard contractual clauses. The General Court of the EU dismissed the action against the decision (T-553/23); the appeal C-703/25 P is pending (as at October 2026).

Works council: works agreement or not?

Under § 96a Abs 1 Z 1 ArbVG, the introduction of systems that collect, process or transmit employees' personal data by automated means requires the works council's consent as soon as the data go beyond general personal details and details of professional qualifications; data used solely to fulfil obligations under statute, a collective bargaining agreement, a works agreement (Betriebsvereinbarung) or the employment contract are exempt. In addition, § 96a Abs 1 Z 2 ArbVG covers systems for evaluating employees where they are used to collect data not justified by the employees' assignment within the business. In our assessment, a Copilot embedded in Microsoft 365 that generates usage reports and logs per user will usually fall within these provisions; a works agreement is then a condition of validity. The literature on AI assistants in the employment relationship draws the same line: what matters is the intensity of the intrusion, and that depends on licence, configuration and permissions. Analytics functions such as Viva Insights or Copilot Analytics that give the individual personal pointers intrude less than evaluations that give managers breakdowns of small groups or of the whole organisation. If the works council refuses consent, the conciliation board can substitute its decision for that consent (§ 96a Abs 2 ArbVG). § 96 Abs 1 Z 3 ArbVG is stricter: monitoring measures and technical monitoring systems that affect human dignity have no legal effect without the works council's consent, and that consent cannot be substituted. According to the Supreme Court, human dignity is affected where the intensity of monitoring exceeds what the purpose of the monitoring requires (8 ObA 288/01p, call logging; 9 ObA 120/19s, GPS tracking). Anyone evaluating AI logs for performance monitoring may reach that threshold.

Without a works council, § 10 AVRAG requires the consent of the individual employee for such monitoring measures; absent a written agreement on its duration, the employee may withdraw it in writing at any time. Where there is no works council, there is no substitute rule for the data systems covered by § 96a ArbVG; the GDPR remains the benchmark there.

Trade secrets and professional secrecy

Under § 26b UWG, a trade secret is protected only as long as its holder takes reasonable steps to keep it secret. Anyone entering calculations, source code or draft contracts into a personal account whose inputs may be used for training puts that protection at risk. Those bound by professional secrecy are subject to further obligations: lawyers are bound to confidentiality under § 9 Abs 2 RAO and must make sufficient arrangements to ensure that their support staff reliably comply with that duty; doctors and their assistants are subject to § 54 Abs 1 ÄrzteG 1998, and tax advisers and auditors to § 80 WTBG 2017. The Austrian Bar (Österreichischer Rechtsanwaltskammertag) says so expressly for lawyers in its guide on AI in law firms (September 2025): entering client-related or otherwise confidential information into public or unsecured AI systems breaches professional secrecy; the provider must undertake in writing not to use inputs for training and must be prepared to enter into an agreement under § 40 Abs 3 RL-BA. In our view, this standard applies to everyone bound by professional secrecy: information entrusted in confidence may be entered into an AI tool only where access by third parties – including the provider, for training purposes – is contractually excluded and technically secured. A personal account is therefore ruled out.

AI Act: a short list for deployers

Anyone using ChatGPT, Copilot or Claude professionally is a deployer within the meaning of Art 3 Nr 4 AI Act. Under Art 4 AI Act as amended by Regulation (EU) 2026/1744, it must take measures to support the AI literacy of its staff, without having to guarantee any particular level. Under Art 50 Abs 4 AI Act, it must disclose deep fakes and AI-generated text on matters of public interest that is published without review. The obligations for high-risk systems, such as AI in recruitment, apply to standalone systems under Annex III from 2 December 2027. For details, see our article on the AI Act.

What to settle first

1. Approved list: which tools and plans are permitted – and which are not.
2. Business accounts instead of personal accounts: centrally administered access through the company; prohibit personal logins for business data.
3. Conclude a data processing agreement, checking the standard contractual clauses, sub-processors and storage location.
4. Switch off training or obtain confirmation that the plan excludes it, and document the setting.
5. Internal AI policy: what may be entered and what must never be, who checks results, when labelling is required. It also serves as the measure under Art 4 AI Act.
6. Involve the works council before the tool is rolled out – where logging or evaluation is per user, by way of a works agreement under § 96a ArbVG.
7. Privacy notice and record of processing activities: add the provider; check whether an impact assessment is needed.

This information is general in nature and does not replace advice on an individual case.