AI Policy
1. Purpose and scope
This policy describes how Kaufmann Sustainability GmbH uses artificial intelligence in client engagements and internal work. It serves as an internal commitment and as a documentation basis for clients, data processing agreements, and audits.
Scope:
- All client projects with AI-assisted workflows
- Internal work (e.g. code generation, research, draft text)
- Sales and marketing materials, where AI was involved
2. Core principles
Materiality, strategy and stakeholder decisions remain a human responsibility. AI is a tool, not a decision-maker.
AI runs in the client's cloud environment (Azure OpenAI EU, Google Vertex EU, AWS Bedrock EU, on-premise). Client data does not leave the client's tenant.
Only providers that contractually guarantee that input data is not used for model training are used. Provider status: see Section 3.
Every productive AI workflow is documented (purpose, input, method, output, sources). Audit trails are required, not optional.
AI output that flows into reports, decision papers or external communication is reviewed by a qualified person before use. Current industry error rates (~23 %) make validation indispensable.
Where possible, smaller specialized models are used instead of large language models. Cloud regions with a high share of renewables are preferred — coordinated with the client's IT team.
3. Provider selection
Whitelist (approved for use)
| Provider | Service | Conditions |
|---|---|---|
| Azure OpenAI Service | GPT-4, GPT-4o, GPT-5 models | Region: West Europe / Sweden Central / France Central. No-training under Service Terms. |
| Google Vertex AI | Gemini models | Region: europe-west4 (Netherlands) / europe-north1 (Finland). Customer data is not used for training. |
| AWS Bedrock | Claude, Llama, Titan | Region: eu-central-1 (Frankfurt) / eu-west-1 (Ireland). Bedrock Data Privacy guarantees no-training. |
| Anthropic API | Claude (Sonnet, Opus) | Direct use: US routing. Where EU residency is required, routed via AWS Bedrock EU. Anthropic API agreement includes a no-training guarantee. |
| OpenAI API direct | GPT models | Internal research / non-personal data only. Not permitted for client projects without explicit client approval. |
Blacklist / use with caution
- Publicly accessible web chatbots (chatgpt.com Free, Claude.ai without login etc.) — not for client data
- Models without a documented no-training policy
Provider change procedure
When adding a new provider:
- Review of privacy notice + DPA availability
- Review of training policy (contractual no-training guarantee)
- Review of EU data residency
- Addition to whitelist + notification of active clients (30 days advance notice)
4. Data classification
Personal data, confidential business secrets, supplier information, financial data.
→ Processed exclusively in client tenant or locally encrypted. Never in public LLM interfaces.ESG data, methodology documentation, audit trail material, aggregated statistics without personal identifiers.
→ Processed in EU regions of approved providers. DPA required.Published reports, press releases, regulatory texts, open datasets.
→ Free to process. Internal tools (e.g. Claude Code for my own research) permitted.5. Audit trail requirements
At minimum, the following is documented per productive workflow:
- Purpose: What is the workflow meant to do?
- Input sources: Which data flows in? (file paths, database sources)
- Model + provider: Which model, which API version, which cloud provider
- Prompt template: Versioned prompts (in Git)
- Output processing: Who reviews the output? Which validation steps?
- Logging: Token consumption, response times, error rates per run
Logging tools (as of 2026): LangSmith, Helicone, OpenLLMetry, Phoenix. At least one per productive pipeline.
6. EU AI Act — Risk classification
The EU AI Act is fully applicable from August 2026. A risk rating is documented per workflow:
| Risk class | Examples | Approach |
|---|---|---|
| Minimal risk | Data extraction, draft text, research summaries | Standard audit trail is sufficient |
| Limited risk | AI-generated client content (e.g. report drafts) | Transparency notice in output that AI was involved |
| High risk | Decision support for personnel evaluations, supplier creditworthiness, lending decisions | Not in standard engagements. Requires separate risk assessment + documented human final decision |
| Prohibited | Social scoring, mass surveillance, manipulative influence | Not offered |
Sustainability reporting workflows typically fall into "minimal" or "limited risk".
7. Incident procedure
If a violation of this policy is identified (e.g. inadvertent use of sensitive data with an unapproved provider):
- Immediate action: Stop the workflow. Isolate the data exposure.
- Within 24 h: Written notification to affected clients.
- Within 72 h: If GDPR-relevant, notify the supervisory authority per GDPR Art. 33.
- Within 7 days: Root cause analysis + mitigation measures documented.
8. Staff and contractors
Current state: Kaufmann Sustainability GmbH operates as a one-person company with Dominik Kaufmann as sole responsible party.
If staff or external contractors are added in the future:
- Written commitment to this policy + data confidentiality (§ 6 BDSG, German Federal Data Protection Act)
- Onboarding training on AI compliance
- No provider access without documented authorization
9. Review and updates
- Annually: Full review of this policy
- Event-driven: Material changes (EU AI Act updates, new providers, new engagement areas)
- Versioning: This file is version-controlled in Git
Owner for updates: Dominik Kaufmann
10. Binding effect
This policy is binding for all activities of Kaufmann Sustainability GmbH from the effective date stated in the version. Deviations are only permitted with documented justification and written approval by the managing director.
Appendix — Quick-reference checklist per new engagement
- DPA signed with client (before data access)
- Cloud region agreed with client IT team
- Provider whitelist aligned with client IT team
- Data classification performed per workflow
- Audit trail tool set up
- EU AI Act risk classification documented
- Validation lead designated (on client side)
- Kickoff protocol signed covering all of the above