Create AI reports automatically: What is permitted under the EU AI Act and GDPR

Regulation (EU) 2024/1689 on artificial intelligence, or EU AI Act for short, classifies AI systems according to risk and prescribes a conformity assessment procedure in accordance with Annex VI for high-risk applications. AI-supported reports in the compliance function are at this threshold: they support the decision of the compliance officer, but do not replace it. Anyone who uses a language model to create data protection impact assessments, NIS 2 risk analyses or supply chain reports must also observe Art. 22 GDPR (prohibition of automated individual decisions with significant effects) and Section 130 OWiG (supervisory obligations).

This article explains which parts of a compliance report can be automated in a legally secure manner, at which point the human officer reservation applies, and which audit requirements apply Art. 12 EU AI Act (logging) and Section 9 BSIG implementation result and what a production-ready workflow looks like. You will also receive an assessment of commercially available tools, a checklist for the data protection impact assessment of the AI ​​system itself and a model of how CIVAC interlinks the automated assessment workflow with the workspace and the officer appointment. At the end of the article, we will show you how you can set up an audit-proof automation path within 2 working days without going beyond the regulatory framework. The requirements apply equally to law firms, medium-sized industrial companies and corporate compliance departments, with varying depths of documentation.

A compliance report documents the legal assessment of a matter and ends with a recommendation. Classically, it includes five building blocks: presentation of the facts, relevant standards, subsumption, risk assessment and recommendation for action. In practice, officers spend 60 to 80% of their time on the first three components because this involves research effort that requires little depth of legal assessment. According to a survey by the professional association, a representative law firm with ten medium-sized clients and 35 annual reports spends an average of 8 hours per report, including 5 hours for research and the initial draft.

Language models make a measurable contribution in precisely these building blocks. The presentation of facts can be generated from structured inputs (documents, forms, email threads). The relevant standards are researched from a curated corpus (GDPR, BDSG, NIS-2 Implementation Act, IT-SiG, LkSG, ISO/IEC 27001:2022, BAFA handouts). The subsumption in an initial draft follows the structure of relevant Senate decisions and information from the supervisory authorities.

The risk assessment and in particular the recommendation for action remain outside the area of ​​automation because they require an assessment that takes into account professional experience, knowledge of the law firm and the client situation. The officer reservation applies here. CIVAC technically reflects this separation in the workspace: The AI ​​draft is created in a marked preliminary stage, the release is carried out by the appointed data protection or compliance officer with a signature. The appointment certificate, signed, filed, verifiable. The preliminary stage can be clearly distinguished from the released report by colour and in the metadata record, so that an examination can immediately recognise the level of maturity of each document. A version history shows how the draft was modified from the AI ​​system to officer approval, including all comments, source corrections and additions from professional practice. When audited by the supervisory authority, the version history is used as evidence of substantial human intervention, which is regularly missing in pure Word or PDF workflows and leads to complaints.

The EU AI Act has been in force since August 2024, with staggered transition periods. High-risk Annex III systems will be fully compliant from August 2026. Compliance reports that prepare decisions in the human resources area, in the lending process or in critical infrastructure fall under Annex III if the report is directly incorporated into a person-relevant decision.

If, on the other hand, the AI ​​system creates reports that an officer independently checks and approves without the AI ​​recommendation being automatically adopted, the classification as a high-risk system must be viewed differently according to the current interpretation of the European Commission and the supervisory authorities. What matters is the actual effect of the AI ​​output on the final decision. A pure research and design function with clear human approval is generally not considered a high-risk system, but must be operated as a general-purpose AI system with comprehensible documentation.

Regardless of the high-risk classification, Articles 13 and 14 of the EU AI Act require transparency, human supervision and documentation. In concrete terms, this means: Every inference is logged (input, model, version, output, timestamp), the officer release is signed and linked to the design, the model is kept in an inventory and assigned a risk class. The EU AI Act obligations from August 2026 are already mapped in the CIVAC Workspace, so that the introduction of AI-supported reports does not require separate tooling. A reclassification over time can be documented with just a few clicks, for example if the authorities tighten their interpretation. The model is maintained in the workspace inventory with the provider, region of the inference endpoint and intended use, supplemented by the associated risk information in accordance with Annex IV of the EU AI Act and the annual review by the officer.

Art. 22 GDPR prohibits decisions that are based solely on automated processing and have legal effects on the data subject or similarly significantly affect them. This fact is relevant in the compliance function if the AI ​​report leads directly to a decision about an employee, applicant or client, for example in whistleblower procedures, anti-money laundering assessments or loan approvals.

The way out lies in substantial human intervention. Substantial does not mean simply nodding off an AI recommendation, but rather an independent assessment with a documented possibility of deviation. If the officer checks the content of the AI ​​draft, adapts it if necessary, rejects it or supplements it, and makes the final decision under his own responsibility, Art. 22 GDPR is not violated. This appreciation must be verifiably documented, for example through a release note with reference to the checked sources, changed passages and the justification for the final recommendation.

CIVAC implements this officer reservation in the workspace using two technical mechanisms. First, each AI-generated draft is marked with a watermark and a version number, which remains in the approved final report. Second: The release requires a textual justification, a deviation field and the digital signature of the appointed officer. Others run compliance like a filing cabinet. We run it like software. The reviewer sees not only the end result, but the entire development path from AI design to officer approval, including all intermediate versions and comments. This transparency regularly has the effect of reducing fines in the supervisory authority's procedure because the substantial human intervention is not only alleged, but is technically proven. An additional line of protection lies in the four-eyes principle for particularly risky reports, for example if the result directly influences a personnel decision or a loan approval.

Art. 12 EU AI Act requires logging for high-risk systems that includes at least: period of use, reference database, input data, identification of the natural persons who check the results. For non-high-risk systems, the Data Protection Conference recommends comparable logging in its guidance on AI from May 2024, as long as input data is personal.

The audit trail of an AI-supported compliance report documents seven data points per application in the CIVAC implementation: input prompt with timestamp, model and version used, model provider with region of inference endpoint, output in full text, verified standard sources with version status, Officer release with signature and justification, final report as a signed PDF with hash value.

The retention period is based on the intended use. Compliance reports on GDPR conformity are regularly stored for three years (limitation period for administrative offenses according to Section 31 OWiG), and for income tax and accounting documents ten years according to Section 147 AO. According to Article 12 Paragraph 2 of the EU AI Act, inference logs must be stored for at least six months, and in high-risk applications at least twelve months, unless other laws provide for longer periods. CIVAC stores this data in an ISO/IEC 27001:2022 certified data centre with EU data residency and, upon request of the supervisory authority, exports it in a machine-readable format (JSON, XML) with hash checksum. The auditor calls, the evidence is ready., including the model version with which the report was created. If the inference is reproduced in the audit, the same prompt can be re-run with the same model version so that the regulator can check reproducibility. Databases of previous inferences are encrypted and only accessible to the appointed officer and internal audit; every access is logged.

Not every report benefits equally from automation. Experience has shown that four categories are particularly effective. First: data protection impact assessments in accordance with Art. 35 GDPR for standard procedures such as cloud migration, new HR tools or employee surveys. The structure is standardised (description, necessity, risks, measures), the relevant sources can be curated (GDPR, BDSG, short papers from the data protection conference).

Second: NIS-2 risk analyses for medium-sized institutions. The structure follows Section 30 of the NIS 2 Implementation Act and can be systematized with reference to BSI IT-Grundschutz modules. Third: LkSG risk analyses for direct suppliers with standardised questionnaires and industry risk profiles. Fourth: ISO/IEC 27001:2022 ISMS statements of applicability with reference to the 93 controls.

Reports with a high level of firm or case specificity, such as antitrust market definitions or special accounting law issues, are less suitable. Here the appreciation portion predominates; the AI ​​only provides initial research. Individual employment law reports with a decision to terminate should also continue to be produced using the classic procedure because the evidence and the balancing of interests undermine any attempt at standardization. CIVAC provides 490 audit templates for the above-mentioned standard reports in the workspace, each with a curated source corpus and predefined officer release path. The templates are adjusted quarterly to reflect changes in the law and new information from regulatory authorities, without the law firm or company having to update it themselves. In practice, this eliminates most of the research and structuring effort, so that the officer can concentrate his time on the assessment and recommending action. Experience has shown that with 35 annual reports, the processing effort is reduced from 8 hours to 3 hours per report without the quality decreasing.

The selection of the language model determines confidentiality, data residency and auditability. Three criteria are crucial. First: data location. Inferences with personal or business-critical data belong in EU data centres. Microsoft Azure OpenAI with EU data residency, Mistral with French hosting, Aleph Alpha with German data centres or open source models in their own hosting environment meet this requirement.

Second: order processing agreement and subprocessor list. Every model provider is a processor according to Art. 28 GDPR. The AV contract must name the subprocessors, regulate the training exclusion right (no use of the inputs for model training) and contain the standard contractual clauses 2021/914 if components are located in third countries. Third: model versioning and reproducibility. The AI ​​system must run in a fixed version so that identical inputs produce identical outputs or at least the range is documented.

CIVAC operates the AI ​​opinion workflow in the workspace with EU data residency and an inference endpoint certified according to ISO/IEC 27001:2022. The right to exclude training is anchored in the AV contract, the subprocessors are identified in the directory. Licence the workspace for your internal representatives, or have our representatives order it. Both paths use the same inference endpoint with the same audit logs. An additional configuration for particularly sensitive mandates, for example with pseudonymization of the input data before inference, can be activated in the workspace using a switch and reduces the data protection classification of the input by one level. An on-premise version with locally hosted models is available for authorities and users in the KRITIS sector, which works without a connection to external inference endpoints and therefore also meets the BSI minimum standard for cloud use in the federal administration.

Before the productive use of an AI-supported expert opinion workflow, a data protection impact assessment of the AI ​​system must be carried out in accordance with Art. 35 GDPR because there is a systematic and extensive evaluation of personal data using new technology. The Data Protection Conference confirmed this obligation in its list of processing-specific risks from April 2024.

The impact assessment comprises six building blocks. First: Systematic description of the processing (which inputs, which model, which outputs, which officer workflow). Second: assessment of necessity and proportionality. Third: assess the risks to the rights and freedoms of those affected. Fourth: Planned remedial measures (pseudonymization, access controls, audit trail, officer reservation). Fifth: Consultation of the DPO. Sixth: If the risk remains high, consult the supervisory authority in accordance with Art. 36 GDPR.

CIVAC provides the DPIA as a template in the workspace and transfers the assessments to new model versions on a quarterly basis as soon as the providers change their inference infrastructure. Audit-proof, documented, Section 35-proof. The template meets the requirements of the LfDI Baden-Württemberg, the BfDI and the Data Protection Conference for the structured DPIA. In addition, a risk matrix is ​​maintained in the workspace, which assesses the AI ​​risks with the probability of occurrence and extent of damage and links them to the remedial measures, so that in the course of an audit, each assessment is linked to the corresponding measure. If the assessment becomes more stringent over time, for example because a new model is to be used without EU inference, the workspace automatically triggers a new DPIA and blocks the model until it is approved by the DSB. Consultation with the supervisory authority in accordance with Art. 36 GDPR is required if the DPIA shows a high residual risk and no sufficient remedial measures can be found.

AI language models do a lot when it comes to creating compliance reports, but not everything. Three limitations are particularly important to note. First: hallucinations when quoting norms. If the model invents paragraphs that do not exist or cites Senate decisions with incorrect reference numbers, only the trained eye will notice it. The solution lies in a curated source corpus (Retrieval-Augmented Generation, RAG) that only contains approved standard texts and information from regulatory authorities.

Secondly: topicality. Language models have a level of knowledge at the training date. Changes to the law, new information from supervisory authorities and Senate decisions based on current knowledge will not be taken into account unless submitted via RAG. The officer checks the currency of each standard cited in the report. Third: partnership context. The assessment of a matter differs depending on the industry, risk appetite, client or employee history and strategic situation. AI systems are not aware of these factors.

CIVAC addresses these limits through a combination of RAG with a daily updated source corpus, an explicit officer release and an overview in the workspace that highlights unusual or unsubstantiated statements. If standards are cited that are not included in the source corpus, the system highlights the passage and requires an officer confirmation. This mechanic does not prevent every hallucination, but it measurably reduces the likelihood of them occurring. In internal tests, the number of unsubstantiated standard citations fell from an average of 4.2 per report without RAG to 0.3 per report with RAG and officer reservation. The remaining quota will be absorbed in the officer release step. Training for officers in recognising typical hallucination patterns is offered once a year and documented in the training certificate. Experience has shown that the recognition rate increases from 65% to over 90% after the first training, which further reduces the remaining rate.

CIVAC is a compliance platform and officer-as-a-service based in Germany. We map the AI ​​assessment workflow in the CIVAC Workspace: input, curated source corpus, officer reservation, audit trail, EU data residency according to ISO/IEC 27001:2022. The 490 audit templates for GDPR, NIS 2, LkSG and ISMS reports are ready for use in the workspace, with quarterly updates on legal changes and supervisory authority information.

Licence the workspace for your internal representatives, or have our representatives appointed. Both paths lead to the same file: appointment certificate, documented reporting line, automated draft with human approval, inference log with model version, signed PDF with hash value. The auditor calls, the evidence is ready. Separate licensing for the language model is not required because the inference endpoint is included in the workspace.

If you want to productively implement AI-powered compliance assessments, a clearly documented path between automation and officer reservation is critical. You will receive an initial consultation on workflow architecture, model selection, DPIA creation and the audit trail free of charge. Write to info@civac.de or use the contact form on civac.de. We will contact you within one working day and arrange a 30-minute appointment to take stock. If necessary, we will involve your appointed data protection or compliance officer in the initial meeting so that the AI ​​workflow is approved from the start. An on-premise version with locally hosted models is available for KRITIS operators and corporate compliance departments, which works without a connection to external inference endpoints. Turn reading into an assignment.