AI Transparency

AZTender is an analytics platform for public procurement data. Some of its features use artificial intelligence. This page explains which ones, what models power them, how AI-produced values are marked, and what role humans play — in plain language.

Where AZTender uses AI

• CPV classification inference. Many Azerbaijani source records are published without a CPV product code. A large language model infers the most likely code from the tender title. Every inferred code is stored with a marker identifying the model that produced it, so inferred values are never confused with codes published by the source.
• Machine translation. Titles and descriptions of non-Azerbaijani corpora are machine-translated into English for display. Translations are stored separately from the originals; the original text is always retained and available.
• Insights (bid-fit analysis). When a user submits a tender text and a company brief, a large language model produces a fit assessment, gap list, and draft structure. The output is decision support for the user — it is not a decision.
• Editorial content. Articles in our News section are prepared with AI assistance and reviewed editorially before publication. They are labeled accordingly.
• Risk scoring and collusion signals are rule-based and statistical computations over the corpus — not machine learning. They surface patterns worth examining; they do not assert wrongdoing. Our working principle is «questions, not findings».

Models and providers

Language-model features run on hosted models from Azure OpenAI (GPT family) and Anthropic (Claude family), accessed through their APIs in configurations that do not use our inputs to train provider models. We do not train models on customer data.

Record-keeping

AI-inferred values carry provenance markers in the underlying records (which model produced the value, and that it was inferred rather than published by the source). Bulk data operations are run with pre-change snapshots and are reversible. Source-data errors we detect (for example, implausible contract amounts on the official portal) are flagged and excluded from analytics transparently — never silently rewritten.

Human oversight

AZTender outputs are analytical inputs for professional users — procurement teams, auditors, analysts. No AZTender feature makes automated decisions that produce legal effects concerning natural persons. Audit-side signals are starting points for human review, and the platform's own accuracy limits are communicated to audit users.

EU AI Act posture

Our AI features are decision-support analytics for organizations working with public procurement records of legal entities. They are not intended or marketed for any use listed in Annex III of Regulation (EU) 2024/1689 (high-risk AI systems), and AZTender does not process personal data in its commercial datasets (see Data Sources & Privacy). In line with Article 50 transparency obligations, AI-assisted published content is labeled, and this page discloses where users interact with AI-powered features. We maintain internal records of model use and data operations consistent with the record-keeping practices described above.

Limitations

Inferred values can be wrong: a CPV code inferred from a title is a best estimate, machine translations may lose nuance, and LLM-produced assessments can contain errors. Numbers from official sources are shown as published, including source-side errors, which we flag rather than correct. If you find an AI-produced value you believe is wrong, write to consulting@elshanmusayev.com — we review and correct provenance-marked values without touching source data.

Operator: EM Consulting (owner: Elshan Musayev), Baden-Baden, Germany. Last updated: 18 August 2026.