Describe a hypothetical AI use case in six clicks. A deterministic scoring engine (gates first, arithmetic second) returns the risk tier and the governance controls it needs, and shows every number and citation behind the result. Runs entirely in your browser.
An educational triage heuristic: a deliberate simplification of these frameworks, built for scoping conversations. Not legal advice, not a compliance determination.
Regulation is categorical, not continuous. The EU AI Act never says a hiring system is 7.4 risky. It says employment is Annex III, and Article 5 practices are banned outright. So the model screens for prohibited practices first, applies category floors second, and only then does arithmetic. Nothing here is a weighted average.
Domain, data sensitivity, consequence, and people reached, plus autonomy times consequence. A sum plus a product, nothing is divided by anything, so no dimension can average another one down. Range 0–26.
Consequence is the only axis that appears twice, because it's the one thing that makes every other axis matter. Autonomy counts only as a multiplier of consequence: an autonomous meeting-notes bot is harmless, an autonomous rejection engine is the worst thing in the catalog, and addition would treat those as the same increment. And because the four additive axes max out at 14, the Critical tier (18+) is unreachable without meaningful autonomy over a consequential output. That comes from the shape of the formula, not from a tuned threshold.
On top of the score, six categorical gates set floors the score can exceed but never undercut. An Annex III hiring pilot stays High even when its arithmetic says Limited. Gates only raise; no gate can lower a tier the score earned.
| Gate | Condition | Effect | Source |
|---|---|---|---|
| G0 | Any Article 5 prohibited pattern | Prohibited: absolute override, score suppressed | EU AI Act Art. 5 |
| G1 | Annex III domain and the output influences or decides an outcome about a person | Floor: High | EU AI Act Art. 6(2) + Annex III |
| G2 | Fully automated and determinative | Floor: Critical | GDPR Art. 22 + internal escalation policy |
| G3 | Special-category data shaping outcomes about people | Floor: High | GDPR Arts. 9, 35; EU AI Act Art. 10 |
| G4 | Human-facing interaction or synthetic content | Floor: Limited | EU AI Act Art. 50 |
| G5 | Safety- or health-critical output | Floor: High | NIST AI RMF MAP 5.1; product-safety logic of the Act |
| Tier | Score range | EU AI Act equivalent |
|---|---|---|
| Minimal | 0–3 | Minimal risk: no specific obligations beyond voluntary codes; Art. 4 AI literacy still applies |
| Limited | 4–9 | Limited / transparency risk: Art. 50 obligations |
| High | 10–17 | High-risk: Art. 6 + Annex III; Arts. 9–15, 26, 27 |
| Critical | 18–26 | No EU equivalent: a deliberate internal escalation tier above the regulation, per NIST AI RMF GOVERN 1.3 and ISO/IEC 42001 Cl. 6.1.2–6.1.3 |
| Prohibited | override | Art. 5 prohibited practices. Not a point on the scale |
An average lets a low axis pull the result down. That's the failure mode where low data sensitivity launders a hiring system into medium risk. A sum can only be raised, and the sum doesn't decide the tier on its own: category gates set floors the score can exceed but never undercut. Try the "Hiring shortlist pilot" example above. The arithmetic lands at Limited (8/26), and the Annex III gate forces High anyway. The number orders the review queue, and the gates do the classifying.
Small integers express a rank order, which is all the underlying frameworks support. Domain gets the widest range (0–4) because the EU AI Act's primary classifier is domain. The moment you write 0.25 you've claimed a precision that doesn't exist and invited "why not 0.3?". With integers the honest answer is available: this is a rank order, and here's the rank.
Because autonomy on its own isn't risk. It's risky in proportion to what it's autonomous over, which is also what the Act encodes: Art. 6(3) exempts narrow preparatory tasks, and Art. 14 requires effective oversight for high-risk systems. The multiplication has a useful side effect. It makes autonomy the biggest single contributor to the score, so the model keeps telling product teams the same thing: keep a human in the loop and the governance burden drops a tier.
Correct, and that's why this is a triage heuristic that borrows the Act's taxonomy rather than a compliance determination. The Act's classification logic (domain, consequence, autonomy, data) is the best-articulated public consensus on what makes an AI use case risky, so it makes a strong intake vocabulary. Whether you're a provider or a deployer, and whether a given system is in scope at all, are legal questions this tool deliberately doesn't answer. For a US-only company the same axes survive; you'd re-source the gates to NYC Local Law 144, Illinois' Human Rights Act amendment, Colorado's AI Act, and sectoral regimes like ECOA and HIPAA. The axes barely change, because domain, data, autonomy, consequence, and reach are how every one of these regimes stratifies risk.
The Act is a floor written for market placement, and an operating committee needs a distinction it doesn't make. Among high-risk systems, the ones running fully automated over determinative outcomes need named executive accountability, independent evaluation, and a tested kill-switch, while the ones with a human approving every decision don't. Collapsing both into "high-risk" means either over-governing the safe pattern until the business routes around you, or under-governing the dangerous one. Setting internal criteria stricter than regulation is exactly what NIST AI RMF GOVERN 1.3 and ISO/IEC 42001 Cl. 6.1.2–6.1.3 ask an organization to do.
Five things. Model security (prompt injection, jailbreaks, data poisoning), which is a security review and separate from use-case risk. Third-party and supply-chain risk (vendor due diligence, GPAI provider obligations). Environmental footprint and cost. Copyright and training-data provenance. And the provider-vs-deployer question under the Act, which is the most consequential legal question here and not answerable from six chips.
Article and clause references were checked against primary sources in July 2026. One timing note: under the EU's Digital Omnibus on AI (approved June 2026), obligations for Annex III high-risk systems now apply from December 2, 2027, while most Article 50 transparency obligations took effect August 2, 2026. The axes and gates are stable either way. The references live as data in the model object and can be updated without touching the scoring logic.