Responsible automation

Responsible AI in public tenders: assist without inventing certainty

A methodology guide for assessing extraction, summarization, and prioritization while preserving source text, the confidence of each field, and the final decision in the hands of a person.

14 minReviewed: 1 September 2026Includes output template

Direct answer

This guide covers the controls of an AI system, not a specific product. AI can speed up classification, field extraction, summarization, and initial comparison, but every material output must point to evidence and show what still needs review. It should not claim by itself that a company meets solvency requirements or that an offer is ready.

Break down work by risk level

Low

Labeling, language detection, date normalization, and proposing search terms, with automatic validation and sampling.

Medium

Summarize scope, extract requirements, or compare with a profile—always with a snippet and the originating page.

High

Eligibility, contractual interpretation, price, signing, or a bid/no-bid decision require competent review.

Anatomy of an auditable output

Field	Example
Statement	“A similar reference is required”
Evidence	PCAP 11.2, page 18, linked excerpt
Confidence	High for extraction; low for similarity
Uncertainty	“Similar” is defined using the first three digits of the CPV
Action	Compare references and validate with the responsible person
Version	Document downloaded on 20/08/2026 09:12

If the system cannot locate evidence, the correct output is “not found” or “requires review”; do not fill the gap with a plausible-sounding inference.

Explainable matching

A useful affinity should be broken down: activity and CPV, territory, amount, experience, exclusions, deadline, and availability. Present positive and negative signals separately. Do not show a percentage with apparent precision if behind it there are only incomplete rules.

The explanation must allow a reviewer to change the decision without having to trust the model.

Personal data and minimization

A business profile can work with ranges, capabilities, and preferences without storing excessive information. If personal data are processed, you must define purpose, legal basis, retention, access, and processors. The AEPD reminds that accuracy and quality also cover inferences and results; a correct input datum does not guarantee a well-proportioned conclusion.

Transparency and editorial control

Since August 2026, transparency obligations under Article 50 of the AI Regulation apply in certain scenarios. The Commission distinguishes, among other matters, content generated without substantive human control or editorial review from that subjected to true human or editorial control. For a guide or summary, recording who reviewed, which sources were checked, and what could be changed is better than a decorative label.

Metrics that detect harm

Accuracy by field and by source—not just overall averages.
Rate of “not found” versus detected hallucinations.
False negatives: relevant opportunities the system discarded.
Human corrections and time until they are incorporated.
Drift due to format changes or a new document version.
Percentage of outputs with accessible evidence.

Official sources

NexoLicit editorial team

Technical and operational content; not a legal assessment of a specific system. Regulatory classification depends on use, context, and real responsibilities.

Reviewed by NexoLicit editorial review team. 1 September 2026