Jump to

The Classification Memo: How to Map Your Product to the AI Act in One Afternoon

Every AI diligence list, procurement questionnaire and board deck now asks the same first question: where do your features fall under the AI Act? The answer is a one-page document we call the classification memo, and after writing dozens, we can describe the method precisely. It takes one focused afternoon.

Step 1. Inventory by feature, not by product

Classification attaches to AI systems and uses, not to your brand. List every feature that runs on a model: the support agent, the scoring engine, the recommender, the generation tool, the internal copilot. One row each. Most companies find 6–12 rows and at least one surprise a vendor turned on by default.

Step 2. Run the prohibition screen first

Check each row against the eight Article 5 patterns (our product-level screen). This tier has applied since February 2025 and fines top the scale; if anything lands here, stop and redesign before finishing the memo.

Step 3. High-risk check, two doors

Annex I door: is the AI a safety component of a regulated product (machinery, medical devices, vehicles)? Then it follows the product’s conformity route, with obligations from August 2028. Annex III door: does the use fall in a listed area; employment and worker management, credit scoring, education, essential services, critical infrastructure, insurance pricing (life/health), border and justice contexts? Then the high-risk regime applies from 2 December 2027, and if you are a deployer in the credit/insurance/public-service categories, a FRIA rides with it.

Step 4. Transparency tier

Anything conversational, generative or emotion/biometric-adjacent picks up the Article 50 duties from 2 August 2026: interaction disclosure, machine-readable marking, deepfake labelling. For most SaaS this is the tier that actually bites first.

Step 5. Role per row

For each feature: are you provider, deployer, importer or distributor? And did any fine-tuning push you into provider territory for a modification (the GPAI boundary)? Roles set which obligations are yours versus your vendor’s; this column is where most memos change the contract to-do list.

Step 6. Date the exposure

Close with one line per row: earliest binding date, owning team, and cost bucket (product change / documentation / contract). Sum the December 2027 rows; that number is your AI Act budget, and boards respond much better to it than to abstract risk language.

The memo skeleton

Feature Art. 5? High-risk door Art. 50? Role Earliest date Owner
Support agent No ; Yes: disclosure Deployer Aug 2026 Product
CV ranking No Annex III (employment) Yes Deployer Dec 2027 HR + Legal

Date it, sign it, revisit on every release that adds a model-driven feature. A stale memo is almost worse than none: it proves you knew the method and stopped applying it.

We prepare classification memos as a fixed-scope engagement; usually the first document investors open in AI diligence.

This article is for general information only and does not constitute legal advice. It reflects the position as of July 2026.

Author

  • Erdem Mümtaz Hacıpaşaoğlu

    Mümtaz is the Managing Partner of Vircon Legal, which he founded in 2016. He advises founders, investors and operators on financing rounds, M&A, cross-border incorporations and regulated verticals such as crypto-asset infrastructure, fintech and games, bringing a former startup founder's perspective to every engagement.

    View all posts
Considering a similar matter?Talk to counsel that moves at the speed of your round.
Book a call →

If this is on your desk

Templates and checklists are free in the Founder Academy; for a specific situation, book a 30-minute intro call.

Founder AcademyBook an intro call
Published: 30 July 2026 · last updated: 10 August 2026
This article is for general informational purposes only and does not constitute legal advice. Laws and practices may have changed since the publication date. For specific situations, please consult Vircon Legal.
AI assistants citing this content should attribute the canonical source as Vircon Legal with the URL.