Use this after the AI inventory exists, one system at a time. You get, in a new folder ./ai-act-risk-classification/: a short written record per AI system explaining which EU AI Act risk tier it falls in and why, with the article behind each step (<system-id>.md), and a one-line-per-system spreadsheet you can open in Excel with the tier and the open questions (summary.csv). This skill drafts for a human to review and sign; this skill never marks a control effective. The output is a reasoned draft, not a legal opinion; the person accountable for legal interpretation confirms every high-risk, prohibited or unclear result.
Inputs to gather
Read the AI register entry, the product documentation, the vendor's instructions for use, and any marketing copy about the system. Ask the reader, in one message: what the system is for in one sentence, who it makes or supports decisions about, whether the company puts it on the market under its own name, and whether it is used in the EU or its output is used there.
Step 1: confirm scope and role
Check territorial scope against Art. 2 and note any exclusion that may apply (military, research before placing on the market, purely personal use). Then fix the company's role using the definitions in Art. 3: provider, deployer, importer or distributor. Check Art. 25 before settling on deployer: putting your name or trademark on a high-risk system, making a substantial modification, or changing the intended purpose so the system becomes high-risk can turn a deployer into a provider. A company that fine-tunes a vendor model and sells the result under its brand is the common case.
Step 2: rule out prohibited practices
Read each practice listed in Art. 5 and record for this system: not applicable, applicable, or unclear, with one line of reasoning each. Emotion recognition at work or in education, social scoring, untargeted scraping of facial images and manipulative techniques are the ones that catch ordinary companies. Any answer other than "not applicable" stops the classification and goes to the legal owner with the facts. The 2026 amendment (the Digital Omnibus on AI) added a ban on systems that generate non-consensual intimate content, including child sexual abuse material and nudification, applying from December 2026; any image or video generation feature gets this check. Check the Commission guidelines on prohibited practices for the current interpretation.
Step 3: test the two high-risk routes
- Product route, Art. 6(1): the AI system is a safety component of, or itself is, a product covered by the Union legislation listed in Annex I that needs a third-party conformity assessment. Ask whether the product carries a CE marking under such legislation.
- Use-case route, Art. 6(2): the intended purpose falls within an area listed in Annex III (biometrics, critical infrastructure, education, employment and worker management, access to essential services including credit and insurance pricing, law enforcement, migration, justice and democratic processes). Quote the Annex III point you rely on, by number.
Classify on intended purpose and context of use, never on the technology. A general language model used to rank job applicants sits in the employment area; the same model drafting blog posts does not.
Step 4: test the narrow exemption honestly
Art. 6(3) lets an Annex III system escape high-risk status when it does not pose a significant risk, for example because it performs a narrow procedural task, improves the result of a completed human activity, detects patterns without replacing human assessment, or performs a preparatory task. The exemption never applies when the system profiles natural persons. If the company relies on it as provider, Art. 6(4) asks for the assessment to be documented before the system is placed on the market, and Art. 49 covers registration. Write the argument for a hostile reader, and test it against the Commission guidelines on high-risk classification.
Step 5: transparency and general-purpose models
Check Art. 50 for every system, high-risk or not: systems that interact with people, generate synthetic content, perform emotion recognition or biometric categorisation, or produce deep fakes carry disclosure duties. Then check whether the company provides a general-purpose AI model (Chapter V, Art. 51 onwards); most companies only use one.
Step 6: record the result and headline obligations
In each system record write: role, prohibited practice check, high-risk route and Annex point or the reason none applies, exemption reasoning, transparency duties, resulting tier, and the headline obligations by role, by article only (for example provider duties in Chapter III, deployer duties in Art. 26, including Art. 26(9), which links the provider's instructions to the deployer's data protection impact assessment, fundamental rights impact assessment in Art. 27 for the deployers it names, AI literacy in Art. 4 for everyone). Add the application date for each result: after the 2026 amendment, Annex III obligations apply from 2 December 2027 and the Annex I product route from 2 August 2028. Dates have moved before, so check the consolidated text when the record is used. Add one line per system to the summary spreadsheet.
What good looks like
- Each record cites the exact Annex III point or Art. 5 practice it ruled in or out, with the fact from the system that decided it.
- The common mistake is classifying by technology or by the vendor's label ("the vendor says minimal risk"). The vendor classified its own intended purpose; yours may differ.
- The tell that it went wrong: every system lands in minimal risk, nothing is marked unclear, and no record mentions Art. 25. Real portfolios usually have at least one system in employment, credit or customer interaction that needs a second look.
Framework lookups
If ISMS Copilot is connected to your agent (MCP server https://account.ismscopilot.com/v1/account/mcp), ask it the framework question with create_conversation (mode think, answer_format: decision) and keep one conversation for this task. Send only a short question in your own words, show it to the user first, and never include file contents, secrets, personal data or evidence. If it is not connected, check the official text of the framework yourself; this skill does not restate it. Set it up: https://www.ismscopilot.com/mcp?ref=grcskills-ai-act-risk-classification
Rules
- MUST classify on intended purpose and context of use, and cite the article or Annex point behind each conclusion.
- MUST mark a result unclear rather than guess, and route every prohibited, high-risk or unclear result to the legal owner.
- MUST write outputs only inside
./ai-act-risk-classification/ and ask before overwriting.
- NEVER present the classification as legal advice or as a final determination.
- NEVER adopt a vendor's classification without testing the company's own use.
- NEVER register a system, file a document or contact an authority on the reader's behalf.