AI transparency notice
What the AI in LeadRadar does, what it never does, how it is classified under the EU AI Act, and how a person stays in control.
Classification
LeadRadar prioritises companies for B2B sales follow-up. This is not a listed high-risk use under Annex III of the EU AI Act. We treat it as a limited-risk system with transparency obligations: AI involvement is labelled in the interface, and drafts carry an AI notice. The system must not be repurposed to evaluate individuals for recruitment, credit, insurance or similar purposes; that would be a separate project with its own assessment.
Three layers, three jobs
| Layer | Does | Role |
|---|---|---|
| Frontier LLM (structured output) | Reads text; returns typed facts with the exact quote, event date, polarity and uncertainties; later writes explanations and drafts from approved facts. | Interpretation and communication |
| Structured classifier | ICP filter, service match, evidence-quality rubric, routing proposals, second check of the extractor. | Bounded decisions |
| Deterministic, versioned code | Quote validation, identity resolution, deduplication, the score P = clamp(0.35·F + 0.65·R − N), gates, idempotent CRM write. | Calculation and action |
What the AI never does
- Compute the score. Changing the generative model does not change P when the validated facts are identical.
- Send anything. A person approves every draft and every CRM record.
- Infer what a text does not say. If the fact is not explicit, the answer is null and the item is marked unknown.
- Profile persons. Evidence is about companies; buying roles are recommended as roles.
- Take instructions from documents. Source pages get no tool permissions; hostile text is part of the acceptance suite.
Human oversight
Every Decision Case shows the contributions, the evidence with sources and the model version used. Reviewers approve, edit or reject with a reason; rejections feed recalibration as reviewed versions, never automatic changes. Contradictions between extractor and classifier route to human review.
Known limits
Buying-stage and 45–90-day window estimates are heuristics labelled as such until calibrated on outcomes. Model confidence is never presented as a probability of purchase. Precision is measured in the pilot (target Precision@20 ≥ 80%). Model, prompt and extractor versions are stored with every result.
Documentation
AI system card, model card, evaluation protocol and human-oversight procedure are available to customers on request.
Items highlighted like this are filled in per customer or before launch. We describe controls we operate; we do not present GDPR compliance, AI Act conformity or ISO certification as achieved results until verified.