The EU AI Act has moved from statute books to enforcement reality. Brussels has begun writing to AI labs, and the phased obligations are now biting: general-purpose AI model duties applied from 2026, with high-risk system obligations following. California enacted three new AI workplace laws with a 2027 compliance countdown, and the UK's Competition and Markets Authority imposed conduct requirements on Google's AI features in June 2026 while probing xAI's Grok.
For startups, the message is clear: AI regulation is no longer a policy debate — it's an operating constraint. This is a plain-language field guide to what the EU AI Act requires, who it catches, and what to do about it.
The risk-tier system in plain language
The Act's core design is a four-tier pyramid. Every AI feature you ship falls into one of these buckets:
Prohibited (unacceptable risk). A small set of AI uses are banned outright — e.g., social scoring, manipulative AI exploiting vulnerabilities, certain real-time biometric identification by law enforcement. Don't ship these in the EU.
High-risk. Where most regulatory weight falls: AI used in hiring, education scoring, access to essential services (credit, insurance, public benefits), law enforcement, migration, and safety components of regulated products. If your AI influences livelihoods, liberty, or safety, assume high-risk until proven otherwise — these systems face the full apparatus: risk management, data governance, documentation, logging, human oversight, and conformity assessment before deployment.
Limited risk. Transparency obligations rather than full conformity assessment. The headline example: AI-generated content must generally be disclosed as AI-generated, and chatbots must identify themselves. This tier catches anyone shipping generative AI features.
Minimal risk. Everything else — spam filters, game AI, inventory optimization. No new obligations. Most internal productivity AI lives here.
The exercise: map each AI feature to a tier, document the reasoning, and revisit whenever you add capabilities. Tier creep is real — a "minimal risk" internal tool becomes high-risk the moment it screens job applicants.
What "general-purpose AI" obligations actually require
Separately from the risk tiers, the Act imposes duties on providers of general-purpose AI (GPAI) models — foundation models powering downstream applications. These applied from 2026 and are now live: technical documentation of training, architecture, capabilities, and limitations; evaluations including adversarial testing; serious-incident reporting; copyright-compliance policies honoring rights-holder opt-outs; and a detailed summary of training data. The largest frontier models ("systemic risk") face additional duties.
For most startups these duties matter two ways: directly if you train your own models, and indirectly through supply-chain risk — your model providers' compliance becomes yours. Ask providers for their documentation; you'll need it for your own compliance file.
The penalty ladder
Penalties scale with severity and are calculated as a percentage of global annual turnover — the GDPR playbook, chosen deliberately because it worked. The heaviest fines attach to prohibited practices; substantial fines to high-risk and GPAI breaches; lower (but still serious) fines to misleading regulators.
The global picture: California and the UK
The EU isn't acting alone. California's three new AI workplace laws — compliance required by 2027 — start with employment (hiring algorithms, workplace surveillance, automated management). The UK takes a different path — no omnibus AI act, but aggressive use of existing powers: the CMA's June 2026 conduct requirements on Google's AI features and its probe of xAI's Grok show Britain regulating AI through competition law. Functionally, the outcome for companies is similar: new obligations on how AI systems treat content, competitors, and users.
A practical compliance checklist for a US startup serving EU customers
Operational guidance, not legal advice — get counsel for the specifics.
- Inventory every AI feature. List each model, API, and AI-driven decision. You can't comply for what you haven't mapped.
- Tier each feature under the pyramid above, in writing, as a living document.
- Check the transparency tier first. Any generative AI user-facing feature needs AI-disclosure (labels, watermarks, chatbot identification) — the most common obligation and the cheapest to satisfy.
- Audit high-risk exposure. Anything touching hiring, credit, education, or safety-critical decisions needs the full apparatus: risk management, data governance, documentation, human oversight, logging. Conformity assessment takes months — start early.
- Vet your model providers. Request their documentation, evaluation summaries, and copyright policies.
- Build incident-reporting muscle. Define what counts as a "serious incident," who decides, and how you'd report in time.
- Track California separately. The three new workplace laws are a distinct compliance track with a 2027 countdown.
- Document everything. Regulators reward good-faith documentation — the difference between a cooperative and a hostile finding.
Why it matters for founders and investors
AI compliance is becoming a cost of doing business — and a competitive advantage. Startups that treat compliance as product infrastructure (documented, automated, auditable) move faster in enterprise sales, where procurement teams now ask about AI Act readiness the way they once asked about SOC 2. "Regulatory-ready" is becoming an RFP checkbox, and the startups that check it win deals.
For investors, regulatory exposure is now a diligence item: what's the risk-tier mapping? What's the model providers' compliance posture? Any high-risk features? Crisp answers mean de-risked; no answers mean contingent liability surfacing at the worst moment.
The deeper opportunity: compliance tooling. Every regulatory wave creates a picks-and-shovels market — GDPR created privacy tech, and the AI Act is creating AI governance. Startups productizing the checklist above are selling into mandatory demand.
Angeled.com connects founders with investors.
Sources: EU AI Act enforcement reporting (Oct 15, 2026); California AI workplace laws coverage (Oct 10, 2026); UK CMA announcements (June 2026)
Read Next
Open Weights Are Not Open Business Models: Mistral Large 4 and the Licensing Trap
Courts Are Writing AI's Rules Now: Three Rulings Founders Should Read
Delhi High Court Orders 24-Hour Takedown of AI Deepfake Videos Targeting Chief Minister
EU Court Adviser Says AI Training Is Legal — but Gemini's Answers May Still Infringe Copyright
AI Regulation After the Midterms: What Founders Should Do Before the January Bills Arrive
DOJ Launches Super Intelligence Task Force to Police AI Misuse



