Does running AI on-premise avoid the EU AI Act?

No. The Act applies based on the use case and risk level, not where the model runs. On-premise can make some obligations (data governance, logging, oversight, residency) easier to satisfy, but it does not exempt you.

What are the AI Act risk tiers?

Unacceptable (banned), high-risk (strict obligations), limited-risk (transparency duties), and minimal-risk (largely unregulated). Most enterprise LLM uses fall into limited or high risk depending on the domain.

Do general-purpose AI models have their own rules?

Yes. General-purpose AI (GPAI) models carry transparency, documentation and copyright-policy obligations, with extra requirements for the most capable models deemed to pose systemic risk. Deployers of open models inherit some responsibilities too.

How does on-premise help with compliance?

Data, logs and the model stay under your governance, making data residency, audit logging, access control and model versioning straightforward to evidence — and the model cannot be changed under your feet by a vendor, which protects your validation.

Can a company using an open model become a provider under the Act?

Yes — if you substantially modify a system, put your name on it, or change its intended purpose into a high-risk one, provider obligations can shift to you. Plain internal deployment usually keeps you a deployer; productizing does not.

What are the penalties under the EU AI Act?

Up to €35M or 7% of global turnover for prohibited practices; up to €15M or 3% for most other violations; up to €7.5M or 1% for supplying misleading information. SMEs get the lower of the amounts.

EU AI Act & On-Premise AI (2026): Compliance Checklist

A common misconception is that keeping AI in-house sidesteps the EU AI Act. It does not — the regulation targets how AI is used and the risk it poses, regardless of infrastructure. What on-premise gives you is control, and control is what compliance evidence is made of. This guide maps the tiers, the roles, the obligations, the penalties, where on-prem genuinely helps, and a working checklist.

Risk tiers and obligations

Tier	Examples	Obligations
Unacceptable	Social scoring, manipulative techniques, emotion recognition at work/school	Banned
High-risk	Hiring, credit, medical devices, critical infrastructure, education scoring	Full compliance regime (see below)
Limited-risk	Chatbots, content generation, deepfakes	Transparency: disclose AI use / AI-generated content
Minimal-risk	Spam filters, most internal tools	Largely unregulated (voluntary codes)

The timeline: it phases in

The Act entered into force in August 2024 and applies in stages: prohibitions and AI-literacy duties from early 2025; GPAI (general-purpose model) obligations from mid-2025; the main high-risk regime from mid-2026; high-risk rules for AI embedded in regulated products extend to 2027. Practical consequence: if you're reading this, the prohibitions and transparency duties already apply to you, and the high-risk machinery is either live or imminent for your use cases — "we'll look at it later" stopped being a strategy in 2025. Exact dates and secondary legislation evolve; verify current status with counsel.

Provider vs deployer — and the trap between them

The Act assigns obligations by role. The provider (who develops a system or model and places it on the market) carries the heavy duties: conformity assessment, technical documentation, CE-marking for high-risk systems. The deployer (who uses AI under its own authority) has lighter but real duties: use the system per instructions, ensure human oversight and staff AI-literacy, keep logs, run data-protection impact assessments where required — and for some high-risk uses, a fundamental-rights impact assessment (FRIA).

The trap: deployers can become providers. Put your name/trademark on an AI system, substantially modify one, or repurpose a system into a high-risk use, and provider obligations can shift to you. For self-hosters the live question is fine-tuning: plain internal deployment of an open-weight model generally keeps you a deployer, but the closer your modifications and productization get to "placing a new system on the market", the more provider-like your duties. If you're fine-tuning and shipping AI features to customers, get a legal read on your role — before, not after.

High-risk obligations, concretely

If a use case lands in the high-risk tier, the regime demands (in substance): a risk-management system across the lifecycle; data governance (relevant, representative training/validation data, documented provenance and bias examination); technical documentation and automatic logging of events; transparency to deployers/users; effective human oversight; and appropriate accuracy, robustness and cybersecurity. Note what this list is made of: evidence. It is a documentation-and-logging regime more than a technology regime — which is exactly why infrastructure control matters.

GPAI models: the layer beneath your use case

General-purpose models carry their own obligations, mostly on the model provider: technical documentation, information to downstream users, a copyright-compliance policy and a summary of training content. Models above a compute threshold deemed to pose systemic risk face extra duties (evaluations, incident reporting, cybersecurity). Open-source models enjoy partial exemptions from some transparency duties — one reason open-weight models are attractive in the EU — but the exemptions fall away for systemic-risk models and don't touch your use-case obligations. As a self-hoster you mostly consume this layer: keep the model's documentation and provenance on file; it's part of your evidence chain.

Where on-premise genuinely helps

Logging without negotiation. High-risk systems must log; on your infrastructure you log everything, in your format, with your retention — no vendor API limitations.
Residency by construction. Data and processing pinned to your EU location — a one-line answer in an audit instead of a supplier-chain investigation.
Model versioning and freeze. Your conformity evidence describes a specific system. A cloud provider can update or retire a model under your feet, silently invalidating your validation; a self-hosted model changes only when you change it — with a changelog.
Full-pipeline auditability. Auditors can inspect the actual system — weights, prompts, retrieval, logs — not a black box behind an API.
GDPR synergy. The AI Act doesn't replace GDPR; both apply. The same on-prem properties (residency, minimization, access control) serve both regimes, and your DPIA and FRIA can share an evidence base.

Penalties (why this is a board-level topic)

Prohibited practices: up to €35M or 7% of global turnover. Most other violations, including high-risk non-compliance: up to €15M or 3%. Misleading information to authorities: up to €7.5M or 1%. SMEs pay the lower of the two amounts. GDPR-scale enforcement culture applied to AI: assume audits will happen in regulated sectors first.

Three scenarios, walked through

Internal RAG assistant over company documents — typically limited/minimal risk: transparency (users know it's AI), AI-literacy for staff, GDPR hygiene on logs. On-prem makes the privacy story trivial. Beware scope creep: the day it starts screening CVs, it changed tier.
CV screening / HR scoring — high-risk by definition (employment). Full regime: risk management, data governance (bias!), logging, human oversight that is real (a human who can and does override), FRIA where applicable.
Vision QC on a production line — depends: quality sorting is usually minimal-risk; the moment it acts as a safety component of regulated machinery it inherits high-risk duties via product legislation. Classification is per-function, not per-project.

On-premise compliance checklist

✓ Inventory every AI use case and classify it into a risk tier (per function, not per project)
✓ Determine your role — provider or deployer — per system, and monitor the "becoming a provider" triggers
✓ Maintain automatic logs of model inputs/outputs for high-risk systems (retention defined)
✓ Document data governance, training-data provenance and the model's own documentation
✓ Version and freeze models under validated systems; change only with a changelog
✓ Ensure real human oversight of high-risk decisions (override power, and it gets used)
✓ Disclose AI interaction and AI-generated content to users (limited-risk transparency)
✓ Train staff: AI-literacy duties apply to deployers
✓ Pin data and processing to the required jurisdiction; align DPIA (GDPR) and FRIA evidence
✓ Calendar the phase-in dates relevant to your systems and re-review classifications yearly

This is general information, not legal advice. The AI Act phases in over time, secondary legislation and guidance evolve — consult qualified counsel for your specific obligations and deadlines.