All news

Europe's draft AI rules make compliance part of the product

A steel caliper measures a small white cube tied with a cobalt ribbon on a graphite table in a band of soft light

The European Commission proposed the first broad AI law, with fines up to 6% of global turnover. High-risk systems will need documented data and oversight.

Fines of up to €30 million or 6% of worldwide annual turnover anchor the first comprehensive legal framework for artificial intelligence, which the European Commission proposed on 21 April. The draft regulation sorts AI systems by risk and reaches providers outside the EU whose systems are used in the Union.

In the high-risk categories, compliance becomes a fixed cost of entering the market, much like safety certification for machinery, and fixed costs favour scale. A large vendor spreads testing and audits across thousands of customers, while a small team selling to banks carries the same burden on a fraction of the revenue. Regulatory readiness now belongs in the price of every investment in hiring, lending, education and health, and the strongest companies in those sectors will sell their audit trail as a feature.

The draft draws three lines. A short list of uses is banned, including social scoring by governments and, with narrow exceptions, live remote biometric identification by police in public spaces. A longer list is classed as high-risk, from CV sorting in hiring and credit scoring to exam marking and safety components of products. Most uses, from video games to spam filters, face no new duties, so the cost lands on a defined set of sectors rather than on the whole market.

What a high-risk system must carry is the real price list: risk management, high-quality training data, logs, documentation, human oversight and a conformity assessment with a CE mark before sale. Like the EU data protection rules, these duties travel with the product to any company that sells into Europe, wherever it is based.

Time favours the teams that start now. The legislative process will take a while, and the draft allows 24 months after entry into force before the rules fully apply, a window to build compliance in rather than bolt it on. Teams in high-risk areas should document data, test for bias, log decisions and design human review from the first version, because buyers will ask before the law does. The tooling is a market of its own: data documentation, model monitoring, bias testing and conformity workflows that every high-risk provider will need.

Scope is still open, and scope decides the bill. The European Parliament and member states will rework the list of high-risk uses and the definition of AI, which settles how many companies fall inside, and whether the final text eases the load on small providers decides how heavy the fixed cost is for a young company. The other test sits outside Europe: whether the UK and the US answer with rules of their own or present themselves as lighter places to build, given that Washington so far treats AI as a matter of funding and security.

Regulation rarely closes a market; it changes who can serve it. Fanaura's mandate covers AI-first companies in the EU as well as in the US and Latin America, and its work with founders includes planning for rules like these from the first customer.

For founders

Building tools that make machine learning systems auditable and compliant? Send us the project.

Fanaura invests its own capital and the capital of its limited partners (LPs), usually through a separate SPV for each deal. If we see a fit, we get in touch.

Apply in five minutes
Subscribe

Get the news by email

One short note each weekday. No marketing, and you can unsubscribe at any time.