ki-richtlinie-entwurf

/ki-governance:ki-richtlinie-entwurf

Purpose

The skill provides a binding internal AI policy (policy) and its annexes as a template – including scope, ethical and legal guardrails, approval process, roles, supplier management, training concept and incident management. The policy is the organizational umbrella of an AI management system within the meaning of ISO/IEC 42001 (Chapter 5 Leadership, Chapter 6 Planning) and the counterpart to the statutory system of obligations of the AI Act (details in ki-vo-compliance).

Inputs

• Organization type (company, group, law firm, professional partnership) and size
• Sector affiliation (finance, health, legal profession, tax advisory, auditing, public sector)
• Existing management system (ISMS according to ISO/IEC 27001, QMS, compliance management)
• Existing tools (Microsoft Copilot, ChatGPT Enterprise, Google Gemini, sector-specific solutions)
• Professional-law obligations (BRAO, BORA, MBO-Ä, StBerG, WPO)

Sub-agent architecture

Researcher provides AI Act anchors (Art. 4, 5, 26, 50), GDPR anchors (Art. 5, 22, 28, 32, 35), professional-law anchors (Sections 43a, 43e BRAO, Section 203 StGB), AGG, standards (ISO/IEC 42001 Chapters 5–8) and commentary references. Drafter creates the policy template with clause numbering, plus annexes (role catalog, approval workflow, supplier checklist, training plan, incident form). Reviewer checks whether Art. 4 AI Act is addressed, the DPA obligation is clear, Section 43e BRAO has been integrated for lawyer clients and the policy does not duplicate AI Act obligations.

Process

1. Scope and definitions

• Spatial (all locations / subsidiaries)
• Personal (employees, trainees, interns, external service providers, clients?)
• Subject-matter (all AI systems within the meaning of Art. 3 No. 1 AI Act; generative AI; embedded AI; agentic systems)
• Reference to AI Act definitions instead of own ones; short glossary as an annex

2. Principles

Principle	Content
Lawfulness	AI Act, GDPR, AGG, professional law, special laws
Security	Confidentiality, integrity, availability (ISMS coupling)
Transparency	Labeling of AI output vis-à-vis data subjects (Art. 50 AI Act as anchor)
Human oversight	no unchecked automation (Art. 14 AI Act; Art. 22 GDPR)
Fairness	Bias avoidance; AGG protected categories + proxy variables
Accountability	clear roles, documented decisions
Data minimization	data minimization Art. 5(1)(c) GDPR; environmental/energy aspect

3. Prohibited uses

• Prohibited practices within the meaning of Art. 5 AI Act (obligation check in ki-vo-compliance)
• Internal taboos (beyond statutory requirements): e.g., no evaluation of employees by AI alone; no client inputs in tools not approved for the law firm; no output in client communication without human approval

4. Approval process (pre-use check)

Threshold model:

1. Triage (every user): rapid risk check (data categories, provider, context of use)
2. Pre-use approval by AI Owner for: personal data, client data, external providers, new fields of use
3. Risk assessment (skill /ki-governance:ki-risikoassessment) for: AI Act high-risk indication, Art. 22 GDPR constellation, third-country transfer, professional secrecy
4. Management board approval for: red risk classification, new system categories, contract conclusion > threshold X EUR

5. Roles and responsibilities

Role	Tasks
AI Owner (organization-wide)	Strategic responsibility, AI inventory, reporting to management board
AI Officer (per system)	Risk, monitoring, supplier, incident management
Risk Owner	Risk treatment decisions, residual risk acceptance
Model Officer	Technical model quality, drift, fairness, audits
Data Protection Officer	DPIA, data subject rights, DPA support
IT Security Officer	TOMs, ISMS linkage
Compliance / Professional Law Officer	Professional law, advertising ban, duty of confidentiality
HR Officer	Art. 4 AI Act training plan, employee data protection

6. Supplier management

Mandatory checks before contract conclusion:

• DPA Art. 28 GDPR with minimum contents (binding instructions, TOMs, sub-processor list, audit right, deletion at the end)
• Third-country transfer Art. 44 et seq. GDPR (adequacy decision / SCC / transfer impact assessment)
• Provider role under the AI Act (provider / authorized representative / importer / distributor) and contractual cascade of obligations
• Professional law: Section 43e BRAO (outsourcing of legal services), Section 203(4) StGB (obligation of persons involved); equivalent for physicians, tax advisors, auditors
• Security/audit evidence: BSI AIC4 self-declaration, SOC 2 Type II, ISO/IEC 27001, ISO/IEC 42001 (once the certification landscape is viable)

7. Training (Art. 4 AI Act)

• Mandatory anchor: Art. 4 AI Act (applicable since 02.02.2025) – personnel must have "sufficient AI competence"
• Target groups: all users (basic training); AI Owners and Model Officers (in-depth); management board (strategic)
• Contents: AI basics, limitations (hallucination, bias), internal policy, data protection, professional law, incident reporting
• Evidence: Participation documentation, repetition at least annually

8. Incident Management

• Reporting path internal (via AI Officer → AI Owner → Management Board if applicable)
• External escalation: GDPR infringements (Art. 33, 34 GDPR), serious incidents within the meaning of Art. 3 No. 49 / Art. 73 AI Act – details in ki-vo-compliance and datenschutzrecht
• Whistleblower protection under the HinSchG for indications of violations of the policy

9. Sanctions, Contact Point, Entry into Force

• Consequences under employment law in the event of a violation (Note: Co-determination under § 87 I No. 1 / No. 6 BetrVG upon introduction)
• Contact point (anonymous or open) for questions and information
• Entry into force, change history, reviewers, follow-up

Sources and Citation Style

Binding: ../../../references/zitierweise.md.

EU and national law

• Regulation (EU) 2024/1689 (AI Act) – in particular Art. 3 No. 1, Art. 4 (training), Art. 5, Art. 26 (operator obligations), Art. 50 (transparency). Detailed cascade of obligations in ki-vo-compliance.
• Art. 5 GDPR (principles)
• Art. 22 GDPR
• Art. 28 GDPR
• Art. 32 GDPR
• Art. 35 GDPR
• § 1 AGG
• § 43a BRAO (confidentiality)
• § 43e BRAO (outsourcing)
• § 203 StGB
• § 87 BetrVG (co-determination)
• HinSchG (whistleblower protection)

Industry Standards

• ISO/IEC 42001:2023 – AI Management System, Ch. 5 (Leadership), Ch. 6 (Planning), Ch. 7 (Support), Ch. 8 (Operation)
• ISO/IEC 23894:2023 – AI Risk Management
• ISO/IEC 22989:2022 – AI Concepts and Terminology
• NIST AI RMF 1.0 (01/2023) – GOVERN function
• BSI AIC4 – requirement groups on governance and supplier management

Commentaries

• Henssler, in: Henssler/Prütting, BRAO, 6th ed. 2023, § 43e marginal no. 1 et seq. [unverifiziert – prüfen]
• Cierniak/Niehaus, in: MüKoStGB, 4th ed. 2022, § 203 marginal no. 1 et seq. [unverifiziert – prüfen]
• Hartung, in: Hartung/Schons/Enders, BRAO, 6th ed. 2022, § 43e marginal no. 1 et seq. [unverifiziert – prüfen]
• Kapoor, in: Keller/Kapoor, KI-VO, 1st ed. 2025, Art. 4 marginal no. 1 et seq. [unverifiziert – prüfen]
• Bertermann, in: Ehmann/Selmayr, DSGVO, 3rd ed. 2024, Art. 28 marginal no. 1 et seq. [unverifiziert – prüfen]

Case law ([unverifiziert – prüfen in Beck-Online / juris / curia.europa.eu])

1. ECJ, judgment of 07.12.2023 – C-634/21, ECLI:EU:C:2023:957 (SCHUFA, Art. 22 GDPR)
2. BAG on co-determination in the introduction of technical equipment, § 87 I No. 6 BetrVG (Senate line)
3. BGH on § 203 StGB and external service providers (Senate line on confidentiality)

Output format

KI-RICHTLINIE
<Organisation> | Version: <X.Y> | Inkrafttreten: <Datum>

§ 1 Geltungsbereich (räumlich, personell, sachlich)
§ 2 Begriffe (Verweis auf KI-VO Art. 3)
§ 3 Grundsätze
§ 4 Verbotene Einsätze
   (Art. 5 KI-VO + interne Tabus)
§ 5 Genehmigungsprozess
   (Triage → Pre-Use-Approval → Risikoassessment → Geschäftsleitung)
§ 6 Rollen und Verantwortlichkeiten
§ 7 Lieferantenmanagement
   (AVV Art. 28 DSGVO; § 43e BRAO; Drittland; BSI AIC4)
§ 8 Schulung (Art. 4 KI-VO)
§ 9 Vorfallsmanagement
§ 10 Sanktionen und Kontaktstelle
§ 11 Inkrafttreten, Wiedervorlage, Änderungshistorie

Anlagen:
  A. Rollenkatalog
  B. Genehmigungs-Workflow (Flowchart-Text)
  C. Lieferantencheckliste
  D. Schulungsplan
  E. Vorfallsformular
  F. Glossar

Risks / typical errors

• Copy AI Act obligations into the policy – maintenance effort and inconsistency; instead a cross-reference to ki-vo-compliance.
• Article 4 AI Act training as a "nice to have" – applicable since 02.02.2025, anchor in the guideline.
• DPA obligation only for "classic" service providers – LLM-as-a-Service is generally data processing on behalf.
• § 43e BRAO / § 203 StGB not addressed for professional secret holders; obligation to bind contributing persons is missing.
• Co-determination obligation § 87 I No. 6 BetrVG overlooked when introducing technical monitoring devices.
• Prohibition clauses too soft ("should be avoided") – fails to serve warning and sanctioning function.
• No obligation to resubmit of the guideline itself (at least annually + upon significant legal changes).