Legal
AI Transparency
Last updated: May 2026
What this page covers
Veln.ai uses large language models to draft documents, score founder applications, and surface signals. This page explains which models we use, what they see, how a human stays in the loop, and how we align with the EU AI Act. Every AI-generated artifact in Veln.ai is a draft for human review, not a final decision.
Models and providers
Document drafting and scorecard rationale use OpenAI GPT-5 (primary) and GPT-5-mini (utility calls), invoked under OpenAI enterprise terms with zero data retention and no training on customer data. Semantic search uses OpenAI text-embedding-3-large, with Voyage voyage-3 as a fallback when the primary embedding service is unavailable. Web-grounded market research uses Perplexity, and only runs when an evaluator or paying founder explicitly enables it. We do not use any generative AI provider that retains customer prompts for model training.
Grounding and refusal to fabricate
Every claim in a Veln.ai-generated document, scorecard, or chat reply is grounded in approved founder truths or cited source documents. If a fact is unknown, the system writes "TBD" or "no link available" rather than inventing a number, a competitor, or a URL. URLs that appear in generated content are HEAD-verified before they reach the user. Founders see the per-area rationale behind every score and can edit any generated artifact before it leaves the platform.
Human-in-the-loop and EU AI Act
Veln.ai does not make a final funding decision. The evaluator (incubator, accelerator, angel, fund, bank, grant body) reviews every scorecard and sets the verdict. Founders can challenge, edit, or reject any AI-generated draft. Our scoring is positioned as decision support, not automated decision-making under Article 22 GDPR. We align with the EU AI Act risk classification by limiting the system to assistive use, providing transparency about model identity, and keeping operator override available at every step.
Questions
For AI-specific questions — model identity in a particular surface, how a score was reached, or a data-protection impact assessment — contact ai-questions@veln.ai.