When to use
Use this at first contact with a proposed or in-flight AI use case — an intake form, a new vendor pilot, an internal build request — when you need a fast, defensible sort before committing deeper effort. It answers one question: how much scrutiny does this deserve? The output routes the case to approve, approve-with-conditions, assess-further, or prohibit, and decides whether a full impact assessment is warranted.
Not for: the deep assessment itself once triage says "assess further" (use aigov-aia-generation), reviewing a vendor's contract and DPA (use aigov-vendor-ai-review), building the register of what you already run (use aigov-ai-inventory), mapping regulatory obligations to controls (use aigov-reg-gap-analysis), first-time scoping interviews (use aigov-cold-start-interview), or standing up policy (use aigov-policy-starter).
This skill is decision-support, not legal advice. Triage is a screening judgment; any high-stakes or rights-affecting use case, any "prohibited?" call, and anything to be relied on or filed must be reviewed by a qualified attorney before you act on it.
Method
- Capture the case in one line: what the system does, who/what it acts on, the decision or output it produces, and whether a human is meaningfully in the loop.
- Screen for prohibited practices first (EU AI Act Article 5 — social scoring, subliminal/manipulative techniques, exploitation of vulnerabilities, untargeted facial-image scraping, most real-time remote biometric ID in public, emotion recognition in workplace/education, certain predictive policing/biometric categorization). A hit stops triage.
- Screen for high-risk categories: Annex III domains plus safety-component/regulated-product uses. Decision: any match, or any decision that materially affects a person's rights, safety, access, or livelihood → route to assess-further and trigger aigov-aia-generation.
- If not high-risk, screen for limited-risk transparency duties (chatbots, AI-generated or manipulated content / deepfakes, emotion/biometric categorization outside prohibited contexts).
- Otherwise classify minimal-risk. Decision: minimal-risk with no sensitive data and no automated adverse decision → approve or approve-with-conditions.
- Map the case to NIST AI RMF MAP (context, purpose, affected people, benefits/harms) to record why the tier fits.
- Emit tier + route + rationale + the one condition or next step that unblocks it, and flag every uncertainty for attorney review.
Example
Intake: "HR wants a tool that screens résumés and ranks candidates for recruiters."
Step 2: not a prohibited practice. Step 3: employment/recruitment is Annex III high-risk, and it affects access to livelihood. Route: assess-further → high-risk. MAP note: affected people = applicants (including protected classes); primary harm = discriminatory filtering; human-in-loop = recruiter reviews shortlist but relies on the ranking. Trigger aigov-aia-generation; require bias testing and a vendor conformity/role check; escalate the "is the recruiter review meaningful oversight?" question to attorney.
Pitfalls
- Skipping the prohibited screen. A limited-risk-looking chatbot that does emotion recognition on employees is a prohibited use — always run Article 5 first.
- "Human in the loop" as a free pass. Rubber-stamp review does not downgrade a high-risk decision; oversight must be able to override in practice.
- Triaging the vendor's marketing, not the actual use. Tier the concrete decision the system drives in your context, not the product category.
- Treating triage as the assessment. This is the fast sort; a "high-risk" or "prohibited?" result is the start of real work and attorney review, not the end.
Output format
AI USE-CASE TRIAGE (decision-support, not legal advice) — <case> — <date>
ONE-LINER: does <what> to <whom>, produces <decision/output>, human-in-loop=<yes/no/rubber-stamp>
TIER: prohibited | high-risk | limited-risk | minimal-risk
BASIS: <EU AI Act Art.5 / Annex III domain / transparency Art.50 / none>
NIST MAP: context=<> affected=<> primary-harm=<>
ROUTE: approve | approve-with-conditions | assess-further | prohibit
CONDITION / NEXT STEP: <the one thing that unblocks or the AIA trigger>
FULL AIA REQUIRED: yes/no -> aigov-aia-generation
ATTORNEY ESCALATION: <rights-affecting / prohibited-call / uncertain items>
UNVERIFIED: <assumptions to confirm>
Reference
Triage rests on the EU AI Act's risk-tier architecture, cross-checked with NIST AI RMF MAP. Tiers below are the current framework as commonly summarized; article numbers and Annex contents change, so confirm the operative text with counsel before relying on a tier.
EU AI Act risk tiers (the triage backbone)
| Tier |
What it covers |
Consequence |
| Prohibited (Art. 5) |
Unacceptable-risk practices — banned outright |
Do not deploy; escalate immediately |
| High-risk (Art. 6 + Annex III / safety components) |
Systems posing significant risk to health, safety, or fundamental rights |
Conformity assessment, risk mgmt, data governance, logging, human oversight, transparency, accuracy/robustness |
| Limited-risk (Art. 50) |
Interaction/content systems |
Transparency/disclosure obligations only |
| Minimal-risk |
Everything else |
No mandated obligations; voluntary codes |
Prohibited practices (Article 5) — a hit stops triage
- Subliminal, manipulative, or deceptive techniques that distort behavior and cause harm.
- Exploiting vulnerabilities (age, disability, socio-economic situation).
- Social scoring by public or private actors leading to detrimental/unjustified treatment.
- Untargeted scraping of facial images to build recognition databases.
- Emotion recognition in the workplace and education (narrow exceptions for safety/medical).
- Biometric categorization inferring sensitive attributes (race, political/religious/union status, sexual orientation).
- Real-time remote biometric identification in public spaces for law enforcement (narrow, authorized exceptions).
- Certain individual predictive policing based solely on profiling.
High-risk categories (Annex III domains)
| # |
Domain |
Typical use |
| 1 |
Biometrics |
Remote/post biometric ID, categorization, emotion recognition (where not prohibited) |
| 2 |
Critical infrastructure |
Safety of traffic, water, gas, electricity, digital infrastructure |
| 3 |
Education & vocational training |
Admissions, scoring, proctoring, assessment |
| 4 |
Employment & worker management |
Recruitment, screening, promotion, termination, task allocation |
| 5 |
Essential services |
Credit scoring, insurance risk/pricing, access to public benefits, emergency dispatch |
| 6 |
Law enforcement |
Risk assessment, evidence reliability, profiling |
| 7 |
Migration, asylum, border control |
Risk assessment, visa/application examination |
| 8 |
Justice & democratic processes |
Assisting judicial decisions, influencing elections |
Also high-risk: AI that is a safety component of, or is itself, a product under EU harmonization law requiring third-party conformity assessment.
Limited-risk transparency (Article 50)
Disclose when a person interacts with an AI system (chatbots), and label AI-generated or manipulated content (deepfakes, synthetic media). Emotion-recognition and biometric-categorization subjects must be informed. These are disclosure duties, not full-conformity obligations.
Triage decision rubric
Art. 5 prohibited practice? -> PROHIBIT (attorney)
else Annex III / safety component
OR rights/safety/livelihood-affecting decision? -> ASSESS-FURTHER (high-risk, AIA + attorney)
else interaction/synthetic content? -> LIMITED-RISK (transparency conditions)
else -> MINIMAL-RISK (approve / conditions)
Mapping to NIST AI RMF MAP
MAP is the RMF function that establishes context before you measure or manage. In triage, record: (1) intended purpose and deployment context, (2) affected individuals/groups and their vulnerability, (3) categorized benefits and potential harms, (4) whether human oversight can meaningfully intervene. A well-populated MAP is the evidence trail that justifies the assigned tier and feeds directly into a full aigov-aia-generation if the route is assess-further.
Escalation rule
High-stakes or rights-affecting decisions — anything touching employment, credit, benefits, health, safety, law enforcement, or a prohibited-practice question — escalate to a qualified attorney and route to a full algorithmic-impact assessment. Triage narrows scrutiny; it never substitutes for legal review on the cases that carry legal stakes.
1---2name: aigov-use-case-triage3description: Rapidly triage a proposed AI use case into a risk tier and a route (approve / conditions / assess further / prohibit) so teams know how much scrutiny it needs.4---56## When to use78Use this at first contact with a proposed or in-flight AI use case — an intake form, a new vendor pilot, an internal build request — when you need a fast, defensible sort before committing deeper effort. It answers one question: how much scrutiny does this deserve? The output routes the case to approve, approve-with-conditions, assess-further, or prohibit, and decides whether a full impact assessment is warranted.910**Not for:** the deep assessment itself once triage says "assess further" (use aigov-aia-generation), reviewing a vendor's contract and DPA (use aigov-vendor-ai-review), building the register of what you already run (use aigov-ai-inventory), mapping regulatory obligations to controls (use aigov-reg-gap-analysis), first-time scoping interviews (use aigov-cold-start-interview), or standing up policy (use aigov-policy-starter).1112This skill is decision-support, not legal advice. Triage is a screening judgment; any high-stakes or rights-affecting use case, any "prohibited?" call, and anything to be relied on or filed must be reviewed by a qualified attorney before you act on it.1314## Method15161. Capture the case in one line: what the system does, who/what it acts on, the decision or output it produces, and whether a human is meaningfully in the loop.172. Screen for prohibited practices first (EU AI Act Article 5 — social scoring, subliminal/manipulative techniques, exploitation of vulnerabilities, untargeted facial-image scraping, most real-time remote biometric ID in public, emotion recognition in workplace/education, certain predictive policing/biometric categorization). A hit stops triage.183. Screen for high-risk categories: Annex III domains plus safety-component/regulated-product uses. **Decision:** any match, or any decision that materially affects a person's rights, safety, access, or livelihood → route to assess-further and trigger aigov-aia-generation.194. If not high-risk, screen for limited-risk transparency duties (chatbots, AI-generated or manipulated content / deepfakes, emotion/biometric categorization outside prohibited contexts).205. Otherwise classify minimal-risk. **Decision:** minimal-risk with no sensitive data and no automated adverse decision → approve or approve-with-conditions.216. Map the case to NIST AI RMF MAP (context, purpose, affected people, benefits/harms) to record why the tier fits.227. Emit tier + route + rationale + the one condition or next step that unblocks it, and flag every uncertainty for attorney review.2324## Example2526Intake: "HR wants a tool that screens résumés and ranks candidates for recruiters."2728Step 2: not a prohibited practice. Step 3: employment/recruitment is Annex III high-risk, and it affects access to livelihood. **Route: assess-further → high-risk.** MAP note: affected people = applicants (including protected classes); primary harm = discriminatory filtering; human-in-loop = recruiter reviews shortlist but relies on the ranking. Trigger aigov-aia-generation; require bias testing and a vendor conformity/role check; escalate the "is the recruiter review meaningful oversight?" question to attorney.2930## Pitfalls3132- **Skipping the prohibited screen.** A limited-risk-looking chatbot that does emotion recognition on employees is a prohibited use — always run Article 5 first.33- **"Human in the loop" as a free pass.** Rubber-stamp review does not downgrade a high-risk decision; oversight must be able to override in practice.34- **Triaging the vendor's marketing, not the actual use.** Tier the concrete decision the system drives in your context, not the product category.35- **Treating triage as the assessment.** This is the fast sort; a "high-risk" or "prohibited?" result is the start of real work and attorney review, not the end.3637## Output format3839```40AI USE-CASE TRIAGE (decision-support, not legal advice) — <case> — <date>41ONE-LINER: does <what> to <whom>, produces <decision/output>, human-in-loop=<yes/no/rubber-stamp>4243TIER: prohibited | high-risk | limited-risk | minimal-risk44BASIS: <EU AI Act Art.5 / Annex III domain / transparency Art.50 / none>45NIST MAP: context=<> affected=<> primary-harm=<>4647ROUTE: approve | approve-with-conditions | assess-further | prohibit48CONDITION / NEXT STEP: <the one thing that unblocks or the AIA trigger>49FULL AIA REQUIRED: yes/no -> aigov-aia-generation5051ATTORNEY ESCALATION: <rights-affecting / prohibited-call / uncertain items>52UNVERIFIED: <assumptions to confirm>53```5455## Reference5657Triage rests on the EU AI Act's risk-tier architecture, cross-checked with NIST AI RMF MAP. Tiers below are the current framework as commonly summarized; article numbers and Annex contents change, so confirm the operative text with counsel before relying on a tier.5859### EU AI Act risk tiers (the triage backbone)6061| Tier | What it covers | Consequence |62|------|----------------|-------------|63| **Prohibited** (Art. 5) | Unacceptable-risk practices — banned outright | Do not deploy; escalate immediately |64| **High-risk** (Art. 6 + Annex III / safety components) | Systems posing significant risk to health, safety, or fundamental rights | Conformity assessment, risk mgmt, data governance, logging, human oversight, transparency, accuracy/robustness |65| **Limited-risk** (Art. 50) | Interaction/content systems | Transparency/disclosure obligations only |66| **Minimal-risk** | Everything else | No mandated obligations; voluntary codes |6768### Prohibited practices (Article 5) — a hit stops triage6970- Subliminal, manipulative, or deceptive techniques that distort behavior and cause harm.71- Exploiting vulnerabilities (age, disability, socio-economic situation).72- Social scoring by public or private actors leading to detrimental/unjustified treatment.73- Untargeted scraping of facial images to build recognition databases.74- Emotion recognition in the workplace and education (narrow exceptions for safety/medical).75- Biometric categorization inferring sensitive attributes (race, political/religious/union status, sexual orientation).76- Real-time remote biometric identification in public spaces for law enforcement (narrow, authorized exceptions).77- Certain individual predictive policing based solely on profiling.7879### High-risk categories (Annex III domains)8081| # | Domain | Typical use |82|---|--------|-------------|83| 1 | Biometrics | Remote/post biometric ID, categorization, emotion recognition (where not prohibited) |84| 2 | Critical infrastructure | Safety of traffic, water, gas, electricity, digital infrastructure |85| 3 | Education & vocational training | Admissions, scoring, proctoring, assessment |86| 4 | Employment & worker management | Recruitment, screening, promotion, termination, task allocation |87| 5 | Essential services | Credit scoring, insurance risk/pricing, access to public benefits, emergency dispatch |88| 6 | Law enforcement | Risk assessment, evidence reliability, profiling |89| 7 | Migration, asylum, border control | Risk assessment, visa/application examination |90| 8 | Justice & democratic processes | Assisting judicial decisions, influencing elections |9192Also high-risk: AI that is a safety component of, or is itself, a product under EU harmonization law requiring third-party conformity assessment.9394### Limited-risk transparency (Article 50)9596Disclose when a person interacts with an AI system (chatbots), and label AI-generated or manipulated content (deepfakes, synthetic media). Emotion-recognition and biometric-categorization subjects must be informed. These are disclosure duties, not full-conformity obligations.9798### Triage decision rubric99100```101Art. 5 prohibited practice? -> PROHIBIT (attorney)102 else Annex III / safety component103 OR rights/safety/livelihood-affecting decision? -> ASSESS-FURTHER (high-risk, AIA + attorney)104 else interaction/synthetic content? -> LIMITED-RISK (transparency conditions)105 else -> MINIMAL-RISK (approve / conditions)106```107108### Mapping to NIST AI RMF MAP109110MAP is the RMF function that establishes context before you measure or manage. In triage, record: (1) intended purpose and deployment context, (2) affected individuals/groups and their vulnerability, (3) categorized benefits and potential harms, (4) whether human oversight can meaningfully intervene. A well-populated MAP is the evidence trail that justifies the assigned tier and feeds directly into a full aigov-aia-generation if the route is assess-further.111112### Escalation rule113114High-stakes or rights-affecting decisions — anything touching employment, credit, benefits, health, safety, law enforcement, or a prohibited-practice question — escalate to a qualified attorney and route to a full algorithmic-impact assessment. Triage narrows scrutiny; it never substitutes for legal review on the cases that carry legal stakes.
Matrixx0070