AI and decision transparency

1. System role

TenderHunter is a decision-support system for tender teams. It does not decide for an organisation, submit a bid, or represent eligibility to a contracting authority. A user always decides whether to proceed.

2. How a result is produced

1. Public-procedure documents are collected and versioned.
2. A language model helps extract requirements, values, and exact source references.
3. Schema and citation checks reject output without supported evidence or route it for review.
4. Deterministic rules compare requirements with the Company Passport. The language model does not choose the final label.
5. The interface displays BID, REVIEW, or IGNORE with requirement states, reasoning, sources, and gaps.

3. Meaning of labels

BID

Based on available current data, no unmet mandatory condition was detected. Human verification is still required.

REVIEW

At least one material item is incomplete, ambiguous, low-confidence, or needs legal or business judgement.

IGNORE

Available evidence indicates an unmet mandatory condition or qualification rule. The user should check the source before declining.

4. Data used by the system

The system uses procedure documents, public contracting-authority information, Company Passport data and evidence, and quality-administrator review labels. It does not use an individual's sensitive traits for profiling or assessment.

5. Limitations and known risks

• A document may be outdated, incomplete, incorrectly scanned, or amended after analysis.
• The model may miss a condition, assign a wrong value, or misinterpret legal language.
• The Passport may contain outdated declarations or evidence expiring before contract performance.
• Scores and labels simplify the real assessment; they do not cover the full strategy, price, or delivery risk.

For these reasons, missing information never counts as a pass, and low confidence is routed to REVIEW.

6. Required human oversight

Before deciding, the user should open the cited clauses, confirm the latest specification and amendment notices, check evidence expiry, resolve every REVIEW or UNKNOWN state, and consult a qualified adviser on legal questions. Platform administrators review selected low-confidence extractions, but this does not replace user verification.

7. No GDPR Article 22 decision

A recommendation concerns an organisation's business opportunity to participate in a tender. It does not itself produce a legal or similarly significant effect on an individual. It is not used for employment, credit, insurance, public benefits, or assessment of a person.

8. Quality and release

AI features for private data and public recommendations remain gated until documented quality and processing conditions are met. The minimum research gate includes 500 human labels, at least 90% precision, 70% ontology coverage, complete citation traceability, and zero reviewed false passes.

9. Questions and challenging a result

A user may report an incorrect requirement or result to their organisation administrator or roberto@brandsparkle.se, including the analysis identifier and the relevant source excerpt. A report does not change the procurement documents or pause their deadlines.

EU Artificial Intelligence Act on EUR-Lex