Skip to content

Pickles GmbH — AI Governance Framework

An open-source AI governance framework for German legal AI providers, covering EU AI Act, GDPR, BDSG, and BRAK professional standards.

Full documentation: Documentation site

This framework exists because there weren't clear worked examples of what EU AI Act compliance documentation looks like in practice for a legal AI company. Rather than write documents manually, I designed a multi-agent system to produce them — with regulatory research, assumption tracking, and quality controls built into the generation process.

The result is 22 documents across five stages, a worked example showing a fictional contract review tool walked through the complete framework, and a reusable methodology that other organisations can adapt.

Licence: CC BY 4.0 — use and adapt freely, attribution required. Status: Framework proposal. Not a legal compliance certification. Legal review: Required before operational use.


Contents

Stage First document Count
1 — Regulatory Orientation Regulatory Orientation Note 1
2 — Governance Foundation AI System Inventory 4
3 — Regulatory Alignment EU AI Act Risk Mapping 6
4 — Monitoring & Controls AI Monitoring Framework 3
5 — Commercial Packaging Governance Information Pack 2
Worked Example (Phase 2) System Profile — Vertrag.AI 6

See ASSUMPTIONS-LOG.md for all unverified assumptions embedded in the framework. See DISCLAIMER.md for full scope and limitations notice.


Framework Architecture

The framework is organised across four governance layers:

Layer Coverage
AI System Foundation System inventory, risk classification, intake workflow, human oversight policy
EU AI Act & Regulatory Alignment Risk mapping, technical documentation, transparency, GDPR/DPIA, vendor risk
Monitoring & Operational Controls Monitoring framework, incident response, model change management
Client-Facing Assurance Governance information pack, enterprise sales enablement

Stage dependency map

Stage 0 ──▶ Stage 1 ──▶ Stage 2 ──▶ Stage 3 ──▶ Stage 4 ──▶ Phase 2
 Setup       Reg.         Gov.         Reg.              │    Worked
             Orient.      Found.       Align.            │    Example
                                                    Stage 5
                                               (parallel with 4)

Worked Example — Vertrag.AI

Phase 2 takes a fictional contract review tool, Vertrag.AI, through the complete framework end to end. Every key template is populated with realistic content, and the risk classification reasoning is made explicit at each decision gate.

The worked example demonstrates: - How the risk classification decision tree produces traceable, defensible outputs - Why the EU Act classification (limited risk) and the internal classification (Medium-High) diverge, and why that gap matters - What the human oversight gate looks like in a professional-liability context - How the monitoring framework handles legally privileged material

See the System Profile and the Reflection Note for the full worked example and a first-person account of what the exercise revealed about the framework's design.


Methodology

The framework was produced by a multi-agent Claude Code system with five defined agent roles: Orchestrator, Research Reader, Document Drafter, Assumptions Tracker, and Run Summariser. The system runs autonomously from regulatory source documents placed in an input folder, producing audit-ready Markdown documentation with explicit assumption flagging throughout.

The documents are the output. The system design is the methodology.

The next stage is testing whether this approach transfers to a different jurisdiction and regulatory architecture.

Estimated build cost for Phase 1 (16 documents): $4–10 USD using Claude Sonnet with prompt caching.


Regulatory Coverage

Regulation Coverage
EU AI Act (Regulation (EU) 2024/1689) Risk classification, Article 50 transparency obligations, Annex III mapping, Article 11 technical documentation structure
GDPR (Regulation (EU) 2016/679) Data flow mapping, DPIA, sub-processor obligations, data subject rights
BDSG (Bundesdatenschutzgesetz) German national data protection overlay
BRAK professional standards Lawyer obligations when using AI tools, client confidentiality

Using This Framework

This framework is designed to be adapted, not copied verbatim. Before operational use:

  1. Replace all fictional Pickles GmbH / Vertrag.AI content with your actual system data
  2. Resolve all [ASSUMPTION] flags against real company information
  3. Have the framework reviewed by a qualified lawyer in your jurisdiction
  4. Update regulatory references if the EU AI Act or related guidance has been updated since this was produced

The ASSUMPTIONS-LOG.md lists every unverified assumption across the framework. Start there.


Licence

CC BY 4.0 — you are free to use, adapt, and distribute this framework, including commercially, provided you give appropriate credit.


Author

Built by David Cockson — compliance and governance professional with a background in financial services and igaming RegTech.

Pickles GmbH does not exist. All company characteristics, systems, and scenarios in this framework are fictional and produced for educational and demonstration purposes only.


  • Sable AI Ltd AI Governance Framework — UK jurisdiction equivalent covering UK GDPR, Data (Use and Access) Act 2025, Equality Act 2010, and ICO AI in Recruitment guidance. Built for a fictional UK HR/recruitment AI startup.