AI-assisted, human-edited, EU AI Act compliant.
We use AI in our day-to-day work. We're also a Dutch B.V., which means the EU AI Act applies to us as a deployer. Here's exactly what we use, how we use it, and how that interacts with the work we ship to clients.
What our compliance looks like
- Anthropic Claude (Team) · no-training
- OpenAI ChatGPT (Team / Enterprise) · no-training
- Named human editor on every output
- Per-engagement AI usage log (24 mo)
- 5 working-day regulator response SLA
What we use AI for
AI is a tool in our stack. Nothing we ship has been generated unattended. Below is the honest list.
Research and analysis
Large-scale SERP and citation analysis across Google AI Overviews, ChatGPT, Perplexity, Claude and Gemini. Pattern detection inside log files and crawl exports.
Drafting
First-pass drafts of long-form briefs, schema templates and code. Every draft is edited by a named human before anything ships to a client. No raw AI output goes out the door.
Engineering
AI pair-programming for our own internal tools, including the systems that power the Key Argument Method (citation tracking, passage extraction, schema validation).
What we don't use AI for
Final published copy, client communication, strategic recommendations, anything that requires accountability. AI doesn't sign off on work. People do.
Client data
We don't feed client data into consumer chat tools. Enterprise plans with no-training agreements only. Confidential or PII data is processed in EU-hosted environments with documented retention.
Disclosure to your audience
When client content is AI-assisted and human-edited, disclosure is usually not required under Article 50. When it is (purely AI-generated published as-is), we label it. We never label work we didn't actually use AI on.
The AI vendors we actually use
Two providers, both on team or enterprise plans with no-training-on-our-data terms. Personal accounts on consumer plans are not used for client work.
Anthropic Claude
Plan: Pro and Team. Anthropic's Team plan includes a no-training-on-customer-data clause by default. Used for drafting, code review, research synthesis, and internal tooling.
OpenAI ChatGPT
Plan: Team and Enterprise. Both plans include a no-training-on-customer-data clause by default. Used for drafting, code review, research synthesis, and structured data work.
We also use the corresponding APIs (Claude API, OpenAI API) for internal automation. API access is governed by the same no-training defaults. New vendors are added to our internal AI policy only after the no-training clause and data-handling terms are confirmed in writing.
What "human-edited" means in practice
The Article 50(4) human-review carve-out has no fixed legal definition yet. The standard we apply on every client deliverable is below, so there is no ambiguity.
Substantive editing
Not just typo fixes. Restructuring, claim verification, removing weak passages, adding original insight.
Fact-checking
Every number, date, citation and named source is verified against primary sources before delivery.
Original layer
Our analysis, our examples, our take. If the piece does not survive removing the AI baseline, it does not ship.
Named sign-off
Each deliverable is signed off by a named editor and recorded in the engagement log, retained for 24 months.
EU AI Act, plain-language
Our status
Key Arg B.V. is a deployer of AI systems under the EU AI Act. We are not a provider. We use general-purpose AI tools (OpenAI, Anthropic, Google) on enterprise terms, integrated into our internal workflows.
Article 50 transparency
Article 50 transparency rules become enforceable on 2 August 2026, with the Digital Omnibus moving some deadlines to 2 December 2026. AI-generated content meant to inform the public must be labelled as such. AI-assisted content that goes through meaningful human editorial review generally does not require labelling. We err on the side of labelling when in doubt.
AI literacy and human oversight
Everyone on our team uses AI tools as part of normal work and has formal training on responsible use. Every output is reviewed before delivery. Decisions about strategy, scope and budget are made by humans with names and accountability, not by models.
If a regulator asks
We keep an internal log of the AI systems we use, the data categories processed, and the human-review step for each engagement. If you need this documentation as part of a procurement or audit process, we can supply it within five working days.
Questions before we start working together
If your procurement team needs specific assurances, or if your internal AI policy needs to be reconciled with ours, just ask. We've seen most of the questions and we have written answers.