Brief guide for AI chatbots, assistants, RAG systems, and automations
|
As of 2 August 2026, the transparency requirements of Article 50 of the EU AI Act apply. For businesses, this primarily means clear notification of interaction with AI, proper labeling of certain types of content, and the ability to demonstrate how the system is controlled.
First, determine: is this a standard or high-risk project?
Stop the one-day scenario and schedule a separate assessment if AI affects hiring, lending, access to education or essential services, uses biometrics/emotion recognition, or is related to safety and medicine.
- Standard support chatbot or knowledge base search: most often, start with transparency, GDPR, security, and human oversight.
- AI agent that sends emails, updates CRM, creates payments, or deletes data: add confirmation for critical actions and a detailed log.
- High-risk case: deeper classification, risk management, technical documentation and possibly conformity assessment will be required.
Plan for one working day
| Time | Task | Result |
| 09:00–10:00 | Inventory | Map: AI features, users, data, integrations, actions, and responsible person. |
| 10:00–11:00 | Transparency | Notification of interaction with AI is shown before or at the beginning of the first contact. |
| 11:00–12:00 | AI content | Rules for labeling text, images, audio, video, and deepfakes. |
| 12:00–14:00 | GDPR | Data minimization, legal basis for processing, retention periods, DPA, and privacy notice. |
| 14:00–15:00 | Logging | Logs of agent actions, errors, human confirmations, and system versions. |
| 15:00–16:00 | Security | Least privilege, key protection, limits, human approval, and emergency stop. |
| 16:00–17:00 | Documentation | System passport, data schema, roles, risks, changes, and incident procedures. |
| 17:00–18:00 | Testing and fixing | Screenshots, test cases, a sample log, and a signed decision on readiness/further development. |
What exactly to do: a short instruction
1. Inform the person that they are communicating with AI
Show the message from the start of the first interaction. It should be clear, noticeable, and accessible. Do not hide the information only in the user agreement.
|
2. Set rules for labeling AI content
| Situation | What to do |
| AI chatbot / agent | Disclose AI interaction at the start of the conversation. |
| Deepfake: realistic image, audio, or video | Add a clear visible label no later than the first display. |
| Text on a socially significant topic | Label if it was created/substantially altered by AI and has not undergone meaningful human review. |
| Editorial content after real testing | A separate label is usually not required if a person can modify/reject the text and bears editorial responsibility. |
| Your own generative AI service | Check machine-readable labeling and discoverability of output content. |
3. Minimize the use of personal data
- Do not send names, phone numbers, documents, or correspondence to the model if the task can be solved without them.
- Document the purpose and legal basis for processing, update the privacy notice, and review contracts/DPAs with vendors.
- Set retention periods for conversations and logs; restrict access and account for data transfers outside the EEA.
- If the processing is likely to pose a high risk to individuals, separately assess whether a DPIA is required.
4. Enable AI agent activity logging
For high-risk AI, automatic logging is a direct requirement. For a regular agent, it is a practical measure: without logs, it is difficult to investigate an error and prove that a human confirmed a critical action.
| Block | What to record |
| Minimum | Time; operation ID; workflow/model version; tool or action; result/error; human confirmation. |
| Do not store unnecessarily | Full passwords, API keys, payment details, unnecessary personal data, and the full prompt. |
| Control | Retention period, access rights, protection against changes, and clear incident search. |
5. Limit permissions and keep people in control
- Grant the agent only the necessary permissions; store keys in secure storage, not in the prompt or code.
- Require confirmation before sending messages, publishing, making payments, deleting, or changing important data.
- Add rate limits, an action allowlist, a fallback scenario, a stop button, and an error notification.
- Check for prompt injection, data leakage, incorrect tool, repeated action, and model unavailability.
6. Gather the minimum required documents
| Document | Contents |
| AI system passport | Purpose, users, owner, model, suppliers, integrations, and usage boundaries. |
| Data schema | What data is included, where it is transferred, where and for how long it is stored. |
| Risk Register | Error, impact, likelihood, control, owner and residual risk. |
| Human oversight | Which decisions are reviewed by a human and how to stop/cancel an action |
| Change & incident log | Versions, changes, tests, failures, actions taken, and responsible parties. |
Final readiness checklist
Mark an item only when there is evidence: a screenshot, setting, document excerpt, test, or log entry.
| Check | Proof | |
| ☐ | An AI project owner and incident manager have been assigned. | Name and role |
| ☐ | The purpose, users, model, providers, and integrations are described. | System passport |
| ☐ | Checked whether the scenario involves high-risk or prohibited practices. | Classification result |
| ☐ | The user sees an AI notification from the start of the first interaction. | Screenshot |
| ☐ | There is a clear way to reach an agent for complex and disputed cases. | Dialogue test |
| ☐ | Cases of visible and machine-readable content labeling have been defined. | Publication rule |
| ☐ | Substantive review and editorial responsibility have been assigned to socially significant texts. | Editor / order |
| ☐ | Only necessary personal data is collected. | Field list |
| ☐ | Privacy notice, legal basis for processing, DPA, retention periods, and cross-border transfer have been reviewed. | GDPR check |
| ☐ | Logs record the action, version, result, error, and human approval. | Sample log |
| ☐ | Secrets never end up in the prompt, logs, or repository. | Secrets check |
| ☐ | The agent has minimal permissions; dangerous actions require confirmation. | Access matrix |
| ☐ | Limits, emergency shutdown, and a fallback scenario are configured. | Failure test |
| ☐ | Errors, prompt injection, repeated actions, and data leaks have been tested. | Test report |
| ☐ | Procedures for changes, retesting, and incident logging have been defined. | Change/incident log |
| ☐ | Employees working with the system have received AI training. | Training record |
What should be ready by the end of the day
- user notification and content labeling rules;
- system passport, data schema, risk register, and human oversight procedures;
- sample log, access matrix, test results, and a list of improvements with priorities.
|
How CenterAI can help
CenterAI conducts a practical AI audit: reviews workflows, data and access, identifies obvious risks, and recommends transparency settings, logging, human approval and a minimal documentation package. After the audit, the necessary improvements can be implemented separately.
Discuss an AI project and book a consultation: https://centerai.eu/