AI Transparency & the EU AI Act

Tenderax uses artificial intelligence to score and analyse public tender notices. This page explains what the AI does, how it works, and how to challenge a result, in line with the transparency obligations of Article 50 of the EU AI Act.

Disclosure

Content you see in Tenderax — relevance scores, summaries, and answers to questions you ask about a tender document — is generated or assisted by an AI system. You are interacting with AI-generated output, not a human analyst, whenever you see a relevance score, an AI-written summary, or an answer from the "document grilling" feature.

Which model

Tenderax uses the Anthropic Claude API, specifically the Haiku-tier model, for tender relevance scoring and document analysis. We selected this tier because it delivers fast, cost-efficient scoring at the volume required to process tender notices across five jurisdictions daily, while remaining accurate enough for the decision-support task it performs.

What the AI does

• Relevance scoring: for each tender notice, the model compares the notice content against your company profile and produces a score from 0 to 10 indicating how well the opportunity matches your business.
• Document grilling: you can ask questions about a specific tender document (e.g. "what is the submission deadline?" or "does this require ISO 9001?") and the model answers based on the ingested notice content.

How the citation trail works

Every AI-generated claim in Tenderax is designed to be auditable, not taken on faith:

1. The model identifies which field or passage in the official notice supports each part of its score or answer;
2. Instead of letting the model retype or paraphrase that passage — which risks hallucination — Tenderax deterministically pulls the verbatim quote directly from the ingested official record and attaches it to the model's claim;
3. Each citation is presented as a deep link back to the exact location in the source notice, so you can verify it against the government portal yourself in one click.

This means citations in Tenderax cannot be fabricated by the model: the model chooses what to cite, but the quoted text itself is always sourced mechanically from the original record, not generated freeform.

Evaluation methodology

Scoring accuracy is measured against an internal evaluation set of tender notices with known-good relevance judgments, reviewed periodically as the model or prompts change. Because every citation links to a verifiable source passage, scoring quality is also human-checkable in production — you can always confirm whether the cited passage actually supports the score.

Known limitations

• Relevance scores are decision-support, not a guarantee of eligibility or fit — always read the full official notice before deciding to bid;
• The model can misjudge nuanced eligibility criteria (e.g. complex consortium or security-clearance requirements) that are not stated in plain text within the notice;
• Scoring quality depends on the completeness of your company profile — a vague or outdated profile will produce less accurate scores;
• Machine-translated or OCR-derived notice text (where a source jurisdiction supplies it) can occasionally reduce citation precision.

Data used for scoring

Scoring uses two inputs: the public tender notice content ingested from the five source jurisdictions, and your own company profile data. This data is sent to the Anthropic Claude API solely to generate your results. Under our commercial API terms with Anthropic, this data is not used to train Anthropic's models.

Contesting or overriding a score

AI relevance scores are suggestions, not final judgments. You can:

• Open the citation trail for any score to review exactly which source passages informed it;
• Manually mark a tender as relevant or not relevant regardless of the AI score — your manual judgment always takes precedence in your own view of the pipeline;
• Report a scoring error to [email protected] or via in-app feedback, which we review to improve the scoring prompts and evaluation set.