acttrace

EU AI Act compliance for non-financial SaaS. Two jobs: classify an AI feature's risk category, and draft an Article 50 transparency notice. Backed by the local ActTrace MCP server — a deterministic rules engine, offline, no API key.

When fired

"Is my chatbot EU AI Act compliant?" / "Classify this AI feature" / "What risk category is our CV-screening tool?" / "Write an Article 50 disclosure for our support assistant" / "Do we need an AI transparency notice?"

Tools

• acttrace_classify — deterministic risk classification. Returns one of

prohibited, possible_high_risk, limited_risk_transparency, minimal_risk, unknown, out_of_scope_financial_services, plus confidence, rationale, the obligations list, and source references.

• acttrace_generate_transparency_notice — an Article 50 notice: body text,

suggested placement, caveats, by notice_type and tone.

Process

1. Gather what the user actually knows about the AI feature: what it does,

whether it is user-facing, the model provider, whether a human reviews the output before it acts, and the domain (HR, education, support, etc.).

1. Call acttrace_classify with those facts. Do not invent missing facts —

omit them; the engine reports what is missing and returns unknown if it cannot classify.

1. Present the risk category, the rationale, and the obligations list verbatim.
2. If the category is possible_high_risk, prohibited, needs_legal_review

or out_of_scope_financial_services, surface the escalation_warning prominently.

1. If the feature is user-facing and limited_risk_transparency, offer to

generate a notice with acttrace_generate_transparency_notice.

Hard rules

• This is not legal advice and does not certify compliance. Always show the

disclaimer the tools return.

• ActTrace covers non-financial SaaS only. If the use case is financial

(banking, trading, portfolio/investment, credit scoring), the classifier returns out_of_scope_financial_services — relay that; do not improvise a financial-services compliance opinion.

• Do not override the classifier's verdict. If it returns unknown, ask the

user for the missing facts rather than guessing.

• The EU AI Act's Article 50 transparency obligations apply from August 2026.