--- title: "What the EU AI Act requires of a company AI policy" canonical: "https://nordvec.com/blog/does-your-company-need-an-ai-policy" lang: "en" claims: "https://nordvec.com/api/trust/claims" --- # What the EU AI Act requires of a company AI policy Source: https://nordvec.com/blog/does-your-company-need-an-ai-policy This file is rendered from the same message catalogue as the page, at request time; the HTML page is the canonical document. Statements about the processing chain are bound to the claim record at https://nordvec.com/api/trust/claims, which says which of them the register currently supports. Rasmus Kjær Damgaard, Founder · Published August 3, 2026 · Updated September 27, 2026 This article was produced with AI assistance, and no review of it by a person is on record. Digital source type (IPTC): https://cv.iptc.org/newscodes/digitalsourcetype/trainedAlgorithmicMedia No law mandates a document called an AI policy. Here is what Articles 4 and 50 of the EU AI Act require, what a policy contains, and why the approved-tools register is the section that does the work. **TL;DR:** No EU law mandates a document called an "AI policy". What the law does require, since 2 February 2025, is that companies **promote AI literacy** in the staff who operate AI (Art. 4, softened from "ensure" to "promote" by the Digital Omnibus in July 2026), and, since 2 August 2026, that they meet the [Article 50 transparency duties](/blog/eu-ai-act-article-50) that apply to their use. A written policy is the practical instrument for both, and the section that carries the weight is the approved-tools register. You can assemble a policy with a legal citation behind each section using our free [AI policy generator](/tools/ai-policy-generator). ## What the law requires, precisely [#what-the-law-requires-precisely] Three legal anchors matter for an ordinary company using AI tools: 1. **Art. 4 AI literacy** (in force since 2 Feb 2025): promote a sufficient level of AI literacy in staff operating AI systems, considering their technical knowledge and the context of use. The Digital Omnibus changed the operative verb from "ensure" to "promote" in July 2026; mandatory oversight training remains for high-risk deployers. There is no prescribed method: a policy plus training is the standard way to evidence it. 2. **Art. 50 transparency** (since 2 Aug 2026): the four duties covered in [our Article 50 guide](/blog/eu-ai-act-article-50), applying by role and use, not to everyone equally. 3. **GDPR**, which never went away: pasting personal data into an AI tool is processing, processor agreements are required (Art. 28), and a breach can trigger the 72-hour notification duty (Art. 33). Everything else commonly found in AI policies (ethics statements, review boards) is governance choice, not legal requirement. A good policy separates the two honestly. ## What a policy contains [#what-a-policy-contains] The sections that carry weight: | Section | Why it matters | |---|---| | Approved-tools register | The operative control against shadow AI: which tools, what approval state, what data may enter each | | Data classification + prohibited inputs | A traffic-light model people can apply in seconds, plus the never-list | | Human review rules | AI output is a draft until a person has checked it; four-eyes for anything published | | Transparency duties | The Art. 50 findings for your specific use, not boilerplate | | Private accounts + reporting | A reporting path that is protective, not punitive, so shadow use surfaces instead of hiding | | Training | The Art. 4 hook, with a cadence | | Incidents | Who to tell, how fast, and why speed beats certainty (GDPR's 72 hours) | | Owner + review triggers | A named owner and concrete triggers, not just an annual date | A policy that says "use AI responsibly" governs nothing; a register that says "these five tools, these data tiers, requests for new tools go here" is where the shadow-AI problem gets solved. ## Generate one instead of starting from a blank page [#generate-one-instead-of-starting-from-a-blank-page] Our [AI policy generator](/tools/ai-policy-generator) assembles the policy above from structured answers: your role under the Act (including the white-label trap), your tools with per-tool data limits, your rules, your governance. Every section cites the specific obligation behind it, from the AI Act and GDPR to ISO/IEC 42001 control ids and NIST AI RMF GOVERN subcategories. It runs entirely in your browser, nothing you enter leaves your device, and no AI model writes any part of the document. We used it to generate [our own AI policy](/ai-transparency). ## Sources [#sources] - [Regulation (EU) 2024/1689, Article 4](https://eur-lex.europa.eu/legal-content/EN/TXT/?uri=CELEX%3A32024R1689#art_4) - [Regulation (EU) 2024/1689, Article 50](https://eur-lex.europa.eu/legal-content/EN/TXT/?uri=CELEX%3A32024R1689#art_50) - [Commission FAQ on Article 50](https://digital-strategy.ec.europa.eu/en/faqs/transparency-obligations-under-article-50-ai-act) - [Digitaliseringsstyrelsen FAQ on the AI Regulation](https://digst.dk/tilsyn/ai-forordningen/faq-om-ai-forordningen/) ## Frequently asked questions ### Is an AI policy legally required in the EU? No law names an 'AI policy' as a mandatory document. Article 4 of the AI Act requires promoting AI literacy in staff, and Article 50 imposes transparency duties; a written policy is the practical way to discharge both and to evidence accountability. ### What should an AI policy contain at minimum? Scope, an approved-tools register with data limits per tool, prohibited inputs, human-review rules, the transparency duties that apply to your use, private-account rules with a reporting path, training, incident handling, and a named owner with a review cadence. ### Why is the approved-tools list the most important section? Because shadow AI is where the risk sits: employees pasting company data into unapproved tools. A policy without a concrete, maintained tool register with per-tool data limits governs nothing. ### Can we generate a usable policy instead of writing one from scratch? Yes. Our free generator assembles one deterministically from structured answers, with every section citing the legal obligation behind it, and no AI model writes any part of the document. [All articles](/blog)