One-Pager · Risk & Governance

Working Safely with AI.

The practical one-pager for everyone using AI in their daily work. Which data can go in, which cannot. When you have to label. What you are personally responsible for. When to escalate. Nothing academic, just a stand-by reference for your desk.

DECAID Academy
As of: October 2026
Language: English
1

The Three-Question Check Before Every AI Use

Which tool?

Is the tool approved (tool registry), GDPR-compliant, and covered by a data processing agreement? Public free-tier versions without a workspace license typically have no training opt-out.

Which data?

What am I feeding into it? Public, internal, customer-related, or sensitive? The data classification determines the permitted tool, not the other way around.

What risk does this create?

Who sees the output? Is it used externally? Does it make decisions about people? The greater the impact, the more rigorous the review.

2

The Data Traffic Light: What Can Go Into Which AI?

Not every tool is suitable for every data class. The traffic light is the default. When in doubt, go one level stricter.

Public & fictional Works everywhere

Examples: Press releases, public websites, brainstorming, generic subject-matter questions, fictional personas, code without business logic.

Tools: ChatGPT, Claude, Gemini, Copilot, Langdock. Free-tier versions are fine as long as there are no internal references in the prompt.

Internal Approved tools only

Examples: Strategy documents, internal communications, roadmaps, meeting notes, anonymized analyses, code with business logic.

Tools: Approved workspace/enterprise versions with a DPA only. Prefer EU hosting and actively verify data residency, because some providers also route traffic outside the EU during peak load and add third-party models with their own retention rules. Do not assume that someone has to switch this on first: in some suites, adding third-party models is enabled by default in newly created EU tenants, runs outside the EU data boundary and then follows the third-party provider's retention periods instead of the suite's. Check rather than assume. Never in private free accounts.

Customer-related & sensitive Private only

Examples: Customer data, personnel files, contracts, financial data, salaries, sick notes, health data.

Tools: Private/local models or workspace with verified data residency only, no training logging. When in doubt: consult the data protection officer.

Prohibited Never

Examples: Social scoring of individuals, emotion recognition in the workplace, real-time biometric identification, manipulative practices (EU AI Act Art. 5). Newly prohibited (from 2 December 2026): systems that generate non-consensual intimate imagery of a person or child sexual abuse material. For you as a user, this prohibition applies to intentional use.

Rule: If you are not sure whether it is prohibited: do not do it. Better to ask the AI champion once too often.

3

Output Responsibility & Labeling

You are responsible for what you publish

  • Fact check: Cross-check figures, quotes, names, and sources. AI hallucinates, even when it sounds convincing.
  • Bias check: Are groups being stereotyped? Would those affected sign off on this text?
  • Verify sources: Does the AI reference real, existing sources? Are the links correct?
  • Provenance seals are not proof of authenticity: Content Credentials on an image only tell you what the metadata says. They are often missing because platforms strip them on upload, and they can be forged: a method documented in August 2026 via manipulated smartphones is considered unfixable by the manufacturer. Verify the source, not the seal.
  • Four-eyes principle for external-facing content: Never publish press texts, customer communications, or legal content without a second review.

Labeling requirement (EU AI Act Art. 50)

These transparency obligations have been binding across the EU since 2 August 2026. The Digital Omnibus package confirmed this deadline and only postponed the high-risk deadlines (standalone Annex III systems to December 2027, systems embedded in products to August 2028). For interpretation, the European Commission published final guidelines on Art. 50 on 20 July 2026: not binding, but the main reference for national authorities. For the machine-readable marking of systems already on the market, a grace period applies until 2 December 2026. It only covers providers; your disclosure duties as a deployer apply without any transition period. Violations can be fined with up to EUR 15 million or 3 percent of global annual turnover. The Commission and national authorities began enforcement on 2 August 2026. As of mid-October 2026, no fines against deployers are known; circulating reports of first million-euro fines come from blogs without a primary source and are unsubstantiated.

  • Chatbots: Make it clearly identifiable as AI. Do not impersonate a human.
  • Images, video, audio: If realistically generated or altered: label visibly. Major providers now embed C2PA Content Credentials and invisible watermarks (such as SynthID). However, the guidelines explicitly require labeling that is perceptible to humans, not just machine-readable metadata, because that gets lost when editing, taking screenshots, and uploading to platforms that re-encode files.
  • Who labels what: Art. 50(2) obliges providers to apply machine-readable marking, Art. 50(4) obliges you as a deployer to disclose in a way perceptible to humans. Do not rely on the tool having already taken care of it: for text, Claude and Gemini now mark machine-readably, and ChatGPT has been rolling this out for EU users since October 2026 (via the API only if switched on). Hardly anyone can check such invisible text watermarks themselves yet, and even light rewording weakens them. Visible labeling remains your job in every case.
  • Texts on topics of public interest: Label if not editorially reviewed and approved.
  • Recommendation: When in doubt, be transparent. "Content created with AI support." never hurts and builds trust.
4

The Most Important Do's and Don'ts

DO
  • Only use approved tools. New tool? Get it approved first.
  • When in doubt, go one level stricter than the traffic light suggests.
  • Keep prompts for critical workflows (audit trail).
  • Label AI outputs for external use, if not editorially approved.
  • Human in the loop for decisions that affect other people.
  • Train the team: Actively promote AI literacy and document the training (Art. 4 EU AI Act, enforceable by authorities since 2 August 2026).
  • Ask rather than assume: The AI champion, data protection officer, IT, and Legal are there for exactly this.
DON'T
  • No customer data into public free-tier tools (ChatGPT, Claude, Gemini Free).
  • No health, application, or personnel data into unapproved AI.
  • No fully automated decisions about applicants, customers, or employees (Art. 22 GDPR: in October 2026 the Dutch data protection authority fined Uber around EUR 825 million for this, not yet final).
  • No copyrighted works uploaded to AI without proper rights.
  • No shadow AI: using tools covertly and submitting the results.
  • No blind acceptance of outputs, not even "because it has to be quick".
  • No skipping the label because "nobody will notice anyway".
5

When to Escalate and to Whom?

Since 29 July 2026, the Bundesnetzagentur has been Germany's central market surveillance authority for the AI Act, with its own complaints office. Since 2 August 2026, the EU has additionally operated a complaints tool and a whistleblower tool through which employees and competitors can also report anonymously. Escalating internally is therefore not just good practice but the way to resolve a problem before it ends up outside the company.

Escalate when ...

  • Suspected data protection incident (data ended up in the wrong AI).
  • New use case involving customer-related or sensitive data.
  • Output with significant external impact (press, customer, authority).
  • Suspected discrimination or unfair bias effects.
  • Uncertainty about data classification or tool suitability.
  • Suspected prohibited practice (EU AI Act Art. 5).

To whom?

  • AI champion on the team: first point of contact.
  • AI Manager / AI Project Manager: for tool approvals and use-case assessments.
  • Data Protection Officer (DPO): for personal data matters.
  • IT / Security: for technical anomalies or suspected data leaks.
  • Legal: for copyright, licensing questions, and contract matters.
  • Management: for incidents involving reputational or compliance risk.
6

Mini-Checklist Before Sending

I will not release anything until I can check off all of the following:

  • Tool is approved and appropriate for the data class (traffic light)
  • No personal, customer-related, or sensitive data in public AI
  • Output reviewed for facts, sources, figures, and quotes
  • Bias check completed: would I personally sign off on this?
  • For external-facing content: a second person has reviewed it
  • If externally visible: AI involvement disclosed (Art. 50)
  • For critical workflows: prompt and source documented
  • When uncertain: contacted AI champion / DPO / Legal instead of guessing