
www.simplify-data.de
Many companies have switched on Microsoft 365 Copilot, ChatGPT or similar tools without defining what they may be used for. The AI Act draws a sharp line between everyday use (drafting emails, summarising text, searching documents) and high-risk purposes such as recruiting or performance evaluation. This article complements our introductory article “EU AI Act: What companies need to consider when using AI” by looking at what it means in day-to-day operations (as of November 2026).
Please note: this article is general information and not legal advice. The AI Act applies alongside the GDPR, and both need to be checked.
Deployer or provider: your role day to day
The AI Act assigns duties by role. Two are enough for practical purposes.
Deployer
A deployer is someone who uses an AI system professionally under their own responsibility. If your staff use Copilot in their mail client or ChatGPT in the browser, you are usually a deployer. The provider of the system is then Microsoft, OpenAI or another company. Deployers mainly carry duties on AI literacy (Art. 4), transparency (Art. 50) and, for high-risk systems, on using the system as intended and ensuring human oversight (Art. 26).
When you become a provider
If you configure or repurpose a general-purpose AI system so that it serves a high-risk purpose, you can be treated as a provider under Art. 25. One example is a Copilot Studio agent that pre-sorts job applications. The duties are then considerably broader. That is why every self-built agent deserves a look before it goes live.
Training duty under Art. 4 in practice
Since 2 February 2025, companies have had to take measures so that their staff can use AI adequately. The Digital Omnibus softened the wording: what is required is measures that promote AI literacy, not a guaranteed level for every individual. No certificate is mandated. A useful record answers three questions:
- Who may use which tool?
- Which limits and typical errors are known, such as invented sources or wrong figures?
- Who received which training, and when?
A short, role-based briefing with a date and attendee list is often enough to start. Managers are explicitly part of it.
Transparency under Art. 50 in the office
The transparency duties have applied since August 2026. For deployers they mainly concern content that goes outside the company: deepfakes must be disclosed as such, and AI-generated text published to inform the public on matters of public interest must be labelled unless it has been reviewed and held editorially responsible. Internal drafts or a translated customer email are usually not affected. Providers of generative systems must additionally make output machine-readable as AI-generated; for systems already on the market before August 2026, the Digital Omnibus names a transition period until 2 December 2026. Decide internally which content is reviewed before publication and who is responsible for that review.
When Copilot slides into high-risk territory
Everyday drafting, translating and summarising of existing business documents is not listed in Annex III. Use becomes high-risk when the system materially contributes to decisions about people. For Annex III purposes, the Digital Omnibus names 2 December 2027 as the date from which high-risk duties apply.
| Use | Classification |
|---|---|
| Drafting emails, summarising text, meeting notes | Everyday use, not high-risk |
| Internal knowledge search over your own documents | Everyday use, check data protection and permissions |
| Pre-screening or scoring job applications | Annex III, employment: high-risk |
| Evaluating performance, preparing promotion or dismissal | Annex III, employment: high-risk |
| Allocating tasks based on personal traits | Annex III, employment: high-risk |
If you already use such applications today, do not wait until December 2027. Pause them now or prepare them properly.
Allow list and block list for operations
One page of rules helps more than a long policy document:
- Allowed: email drafts, minutes, internal summaries, research followed by review.
- Only with approval: customer data, contracts, financial data, texts for publication.
- Not allowed: personnel files, applications, health data, confidential content in private consumer accounts.
- Always: check the results. No decision about people based solely on AI-generated text.
Works council and GDPR
Where a works council exists, involve it early. In Germany, AI tools that can evaluate behaviour or performance are regularly subject to co-determination. For high-risk systems in the workplace, Art. 26 also provides for informing workers' representatives in advance. The GDPR remains unaffected: you need a legal basis, a data processing agreement with the provider, transparency towards data subjects and, for high risk, a data protection impact assessment.
Checklist
- Build an inventory: which AI tools run in the Microsoft tenant, in the browser and on mobile devices?
- Record the purpose of each tool in one sentence.
- Deliver and document Art. 4 training for users and management.
- Publish the allow list and block list.
- Check your own agents and Copilot Studio projects for Annex III purposes.
- Define a review process for AI content that goes outside the company.
- Involve the works council and review the data processing agreement.
- Start a data protection impact assessment as soon as personal data is processed at scale.
How Simplify Data can help
We help introduce AI in a way that documents purpose, data flows and training: from the initial inventory and usage rules to custom document search where your data is not used for training. Read more on our Artificial Intelligence page, on planning as part of our IT consulting, or get in touch directly.
Sources
- European Commission: AI Act, Regulatory Framework
- EUR-Lex: Regulation (EU) 2024/1689 (AI Act)
- EUR-Lex: Regulation (EU) 2026/1744 (Digital Omnibus on AI)
- EUR-Lex: Regulation (EU) 2016/679 (GDPR)
This article is for general information only and is not legal advice; for your specific case, please consult a lawyer.