Skip to content

AI Transparency

Fennibi uses AI in several places. This page describes exactly where, what works reliably, where the limits are, and how you can recognise AI-generated content. It is written for everyone working with Fennibi in your organisation — and supports you in meeting your own AI literacy obligation under Art. 4 of the AI Act.

Roles under the AI Act

The AI-supported functions of Fennibi are AI systems within the meaning of Regulation (EU) 2024/1689. The provider within the meaning of Art. 3(3) is Paul Moldmann; your organisation, as the using party, is a deployer within the meaning of Art. 3(4). Given the intended purpose laid down in the contract — automated documentation, quality assessment and analysis of Power BI Semantic Models for internal business purposes — Fennibi is neither a practice prohibited under Art. 5 nor a high-risk AI system under Annex III.

Where AI is used — and where it is not

A large part of Fennibi is deterministic: structure, dependency graphs, relationship diagrams, RLS evaluation and the changelog are produced without AI and are reproducible. AI is used only where metadata has to become language:

Business descriptions
Descriptions for tables, columns, measures and the model as a whole are generated from the metadata.
DAX Doctor
Improvement and refactoring suggestions for the measure architecture, fully automated and without manual review by a DAX expert.
Chat with your Data
Answering questions about your already documented models — based on metadata, not on raw data.
Live analysis
The AI system plans the analysis, generates read-only DAX queries itself, executes them in your own Fabric capacity and formulates a conclusion from the result rows.

Not AI-supported: the detection and masking of credentials, the billing logic, the version comparison (changelog) and the decision whether a run counts as a full refresh. And: Fennibi takes no automated decision about individuals within the meaning of Art. 22 GDPR.

Limits and typical failure modes

The following errors do occur in practice. Knowing them is the actual substance of AI literacy — not trusting that they will not occur.

Plausible but factually wrong
The most common and most dangerous failure mode of the live analysis: the query runs without error and the result is formally computed correctly — but the filter, period or reference figure was chosen incorrectly. The number then looks right and is not.
Invented business context
Descriptions can assert a business meaning that cannot be derived from the metadata. A measure called "margin" is then described the way a margin is usually calculated — not the way yours is.
Incomplete coverage
DAX Doctor does not find every optimisable spot, and the documentation does not capture every model structure. The absence of a finding is not evidence of correctness.
Truncated analyses
Every run has hard limits on steps, time and rows. When one of them is reached, the analysis answers in abbreviated form — this is intended and not a defect.
Varying results
The same question can be answered differently. AI outputs are not reproducible the way the deterministic parts of the documentation are.

How to verify a result

  • Use the evidence chain: every figure in the live analysis can be traced back to the DAX query it came from. Open the query and read which filters and periods were actually set.
  • Check against a known figure: compare the result with a number from a reviewed report before you rely on it.
  • Read descriptions against the model: an AI description is a proposal. Verify it before importing it into your Semantic Model via the TMDL file.
  • TMDL always on a copy first: the refactoring TMDL from DAX Doctor in particular changes your measure architecture — never apply it untested to a production model.

How to recognise AI-generated content

AI-generated outputs are marked — visibly at the interface and machine-readably in the artefact:

Chat and live analysis
A notice at the input field states that you are interacting with an AI system.
PDF export
The cover page and the footer of every page carry the notice; the document properties contain the marking in machine-readable form.
TMDL files
The delivered file itself is marked as AI-generated. That marking does not travel into your Semantic Model: once you have reviewed the descriptions and deliberately adopted them, they are your content and you carry editorial responsibility for it.
Fennibi API
Responses from the chat and analysis endpoints carry a field indicating AI generation.

The marking relates to what Fennibi delivers — in the delivered files and documents it must not be removed or obscured under the contract (§ 11a.7 of the Service Agreement). It ends where you adopt content into your Semantic Model after your own review: at that point it becomes your content (§ 11a.8). If you publish AI-generated content, you are subject to your own obligations under Art. 50 of the AI Act.

What Fennibi must not be used for

The Service Agreement excludes uses that would make Fennibi a high-risk AI system under Annex III of the AI Act — in particular the assessment of individual natural persons for employment decisions, creditworthiness, insurance pricing or access to essential services.

Expressly permitted is the evaluation of aggregated business metrics — including in regulated industries and including where the underlying data relate to individuals. The boundary lies at assessing individual persons for one of the purposes named above.

Models used

AI processing runs via OpenRouter, Inc. as a router to changing model providers. Zero-data-retention is enabled throughout, and the use of transmitted data for training AI models has been expressly objected to. Which model actually answers can change at any time — in particular through the automatic failover on unavailability. Details on data processing are set out in the privacy notice.