Data privacy in AI automation: your data and your inventions stay with you.

In the agency model, work is often spread across several people, sometimes subcontractors. In my way of working, your contact and the only outside person with access to your data are the same person. This page explains how I protect your data, trade secrets and inventions when setting up AI automation.

Is my data used to train AI models?

Not in a properly chosen setup. Free chat apps and business services are different. With business APIs and enterprise plans, providers do not use customer data for model training by default. Free tiers often do.

I share in writing which provider is used, under which contract terms and in which region. Company data is never put into free consumer tools.

Where is my data processed?

We choose one of three options together, depending on how sensitive the data is:

• Business cloud: models under contracts that exclude training, in an EU data region where possible. For everyday correspondence and reporting.
• Your own server: open-source workflow tools (for example n8n) run on your company's server. Only the smallest piece of text the task needs goes out.
• Local model: for the most sensitive work, such as drawings, costs, personal data and R&D, the model runs in-house. Data never leaves.

What are the patent and trade secret risks?

This is the least discussed topic, and potentially the most expensive later. Some examples:

• An invention not yet filed: typing its details into an uncontrolled tool can put its confidentiality and future protection at risk. R&D documents are therefore processed only in a local model.
• Customer drawings and specifications: NDAs with automotive and defence customers often restrict uploading these to third-party clouds. I read the agreement first and set up accordingly.
• Generated content: text, code or images produced by AI can resemble someone else's protected work. Every output that leaves the company is approved by a person and checked for originality where needed.
• Open-source model and tool licences: not every open-source model allows commercial use in the same way. The licences of the tools used are documented.

What is done for GDPR and KVKK?

• Data inventory: which personal data is processed by which agent, for what purpose, is written down.
• Data minimisation: the agent sees only the fields its task needs. Identity and contact details are masked where needed.
• International transfers: if a cloud is used, a transfer route that complies with GDPR and with KVKK's cross-border transfer rules, updated in 2024, is chosen.
• Access and logging: access is limited, and every agent step is logged.

Who has access to my data?

You and your team. From outside, only me. There are no subcontractors. I sign a non-disclosure agreement before work starts. When the work ends, access is closed and working copies are deleted.

Who owns the system?

You do. Workflows, agent instructions, integrations and the accounts used are set up in your own accounts. When the work ends, everything stays with you, documented; you do not become dependent on me. Monthly continuity support is available if you want it, but it is not required.

Could an agent send confidential information out by mistake?

The setup is designed to prevent it. Every output to a customer, supplier or anyone outside is approved by a person. The agent's permissions match its task; steps such as sending emails, payments or contracts never run without approval. Every step is logged.

What security risks are specific to AI agents?

Systems built on large language models have risks of their own on top of classic information security. I use the OWASP top 10 risk list for these applications as a checklist. The ones that matter most in my setups:

• Prompt injection: an instruction hidden in an email or document the agent reads can try to steer it. Outside content is treated as data, not as instructions, and critical steps stay under human approval.
• Excessive agency: giving an agent more access and tools than its task needs. Each agent gets only the permissions its job requires.
• Sensitive information disclosure: an agent passing on information it should not have seen. Access is limited at document level, and personal data is masked where needed.
• System prompt leakage: no passwords, keys or confidential details are written into an agent's instructions.
• Unbounded consumption: an agent stuck in a loop driving up cost and load. Request and budget limits are set.

Why does working boutique make a difference for privacy?

Privacy is more a matter of responsibility than of tool settings. In a large team, your data passes through many people and systems, and tracking who accessed what gets harder. When I work one-on-one, your contact, the person doing the setup and the person with access to your data are the same. Responsibility sits in one place.

I learned privacy discipline in the automotive supply chain.

In Hyundai and Ford vehicle programmes, customer drawings, costs and quality data were protected by strict confidentiality rules. I saw the same care in supplier audits. I apply this discipline to every AI agent setup. Details are on the About page.