Where the exposure sits.
67 liability rules extracted from the EU AI Act, from 116 candidate clauses — who enforces what, which penalties attach, and where the statute stays silent.
▧
Liability rules identified
67
From 116 candidate clauses
⚠
Carry a monetary penalty
7
Of 67 rules total
§
Grant private right of action
5
62 of 67 rules are silent
◆
Enforcement actors named
40+
Led by Commission & AI Office
The penalty ladder
Art. 99 · fixed fines & turnover percentages§ 99Every quantified fine, ranked by ceiling
Tier 1LR0063
€35M
€35,000,000
Prohibited AI practices (Art. 5) — or 7% of worldwide turnover, whichever is higher
Tier 2LR0063 / LR0067
€15M
€15,000,000
Other operator/notified-body non-compliance, and GPAI-model provider infringements — or 3% of turnover
Tier 3LR0063
€7.5M
€7,500,000
Incorrect, incomplete or misleading information supplied to notified bodies or national authorities
SME capLR0066
€1.5M
€1,500,000
Reduced ceiling for prohibited-practice non-compliance where the offender is an SME
Shape of the rule set
67 rules§ ARule type
§ BPrivate right of action
silent62
express5
Where silent, remedy runs through general Union law — e.g. Directive 85/374/EEC on product liability — rather than a right built into the Act itself.
§ CEnforcement actors
§ DPath & confidence
Source: EU AI Act liability-rule extraction · Art. 99 penalty structure
Carver Agents · carveragents.ai
The obligations behind the rules.
963 discrete obligations extracted from 451 eligible clauses — the actual requirements, notifications, and prohibitions that sit underneath the 67 liability rules.
▤
Obligations extracted
963
From 451 eligible clauses
◆
Linked to a liability rule
416
43% of all obligations
⚠
Carry a stated penalty
310
290 fixed · 20 revenue-%
§
No enforcement mechanism stated
538
56% of all obligations
§ AObligation type
§ BWho's addressed
Risk category (top): ai_algorithmic_accountability 261 · transparency_disclosure 56 · privacy_data_protection 43 · health_safety 42 · cybersecurity 21
Full obligation register
963 obligations
Source: EU AI Act obligation extraction · 451 eligible clauses, 963 obligations, 416 liability-linked
Carver Agents · carveragents.ai
Evidence Vault.
Upload the documentation you'd point to if a regulator, auditor, or insurer asked "prove it" — technical files, declarations of conformity, incident reports, training records — tagged against the obligations and liability rules they satisfy.
Where this lives
Files stay in your browser for this session (nothing is uploaded to a server). The evidence log itself — filenames, tags, notes — is saved to this browser's local storage so it survives a reload; the underlying files are not, so they'll show as needs re-attachment if you come back later.
Drop evidence files here, or click to browse
PDF, DOCX, XLSX, images, or any file type — one at a time
▤
Evidence log
0 filesNo evidence uploaded yet — drag a file above to get started.
Evidence log stored locally in this browser only
Carver Agents · carveragents.ai
What would coverage cost?
A parameterized premium model built off the Act's own penalty ladder — adjust any assumption and the pricing rebuilds live.
Not actuarial advice
This is an illustrative pricing model. It converts statutory fine ceilings into an expected-loss estimate using assumptions you control, capped at the modeled policy limit. Real underwriting layers in loss history, sub-limits, jurisdictional spread, and reinsurance terms not modeled here.
§ 01Assumptions
Company classification
Large undertaking
SME
Determines which ceilings apply — turnover-linked for large undertakings, reduced fixed caps for SMEs under Art. 99.
Annual worldwide turnover€50,000,000
Used for the turnover-linked share of fines (7% / 3% tiers) — before the policy cap is applied.
Tier 1 — Prohibited AI practices
Art. 5 · ceiling €35M or 7% of turnover (large) / €1.5M (SME)
Annual probability of triggering0.05%
Tier 2 — High-risk / GPAI non-compliance
Ceiling €15M or 3% of turnover (large) / €750K (SME)
Annual probability of triggering0.15%
Tier 3 — Misleading information to authorities
Fixed ceiling €7.5M, no turnover linkage in the extracted rule
Annual probability of triggering0.25%
Expected payout vs. statutory max25%
Regulators rarely fine at the statutory ceiling. Scales expected loss down to a realistic share.
Annual probability of a regulatory inquiry8%
Covers the 49 non-monetary rules — enforcement actions, complaints, cure requests.
Cost per regulatory inquiry€150,000
Legal defense, documentation response, internal investigation.
Policy retention (deductible)€5,000
Expected loss the insured absorbs before coverage responds.
Policy limit (aggregate cap)€10,000,000
The insurer will not pay more than this per tier, per year. Every exposure amount is capped here before probabilities apply, and it sets rate-on-line.
Market benchmark rate-on-line1.5%
What the broader market charges as a % of limit, for comparison only. Primary-layer tech E&O / D&O rate-on-line typically runs ~1–3% and has been softening. No carrier publishes per-act EU AI Act pricing yet.
Expense loading25%
Insurer overhead — acquisition cost, claims handling, reinsurance ceding cost.
Profit margin10%
Modeled annual premium
—
Modeled rate on line
—
Market benchmark premium
—
Implied loss ratio
Method
For each fine tier, exposure =
min(max(fixed ceiling, % of turnover), policy limit) for large undertakings, or min(SME fixed cap, policy limit) otherwise. Expected loss = exposure × trigger probability × severity factor. Regulatory-inquiry cost adds inquiry probability × cost per inquiry (also capped at the policy limit). Retention is subtracted, then expense loading and profit margin apply multiplicatively.Reading the gap
Default trigger probabilities are set low (0.05–0.25% a year) — actual statutory violations are rare tail events, not routine ones. If the modeled premium sits well under the market benchmark, that's expected for a bottom-up frequency×severity model: carriers writing early-stage lines like this typically anchor closer to benchmark rate-on-line than to a first-principles loss model, since there's no claims history yet to calibrate frequency against.
Source: EU AI Act liability-rule extraction · Art. 99 penalty structure
Carver Agents · carveragents.ai