AIG-015 AI System Technical Documentation
Description
Technical documentation exists for each AI system before it is deployed and its current version corresponds to the deployed version of the system. The documentation covers the system architecture and its components, the intended purpose and the use cases in scope, performance measures and the limits of that performance, a summary of the training data, known failure modes and edge cases, the human oversight mechanisms in place, the hardware and compute the system requires and the maintenance and update schedule. The documentation is released to deployers, operators and competent authorities on request. Each release is recorded.
Rationale
Technical documentation is the artefact every AI governance audit opens with. A system that exists only in the heads of the engineers who built it cannot be assessed at all. Version correspondence fails most often: documentation written at first release and never updated describes a model that is no longer deployed. The public summary of training content that a general-purpose AI model provider publishes is a duty of that role and belongs to the GPAI provider bundle rather than here. AIG-010 holds the model registry, AIG-034 the instructions for use given to a deployer and AIG-013 the provenance record the training data summary draws on. A general-purpose model as such is documented under AIG-047; a model served inside the organisation's own product carries both controls, this one for the product and AIG-047 for the model.
Applicability (9 profiles)
The documentation is the provider's. The deployer's evidence is the released version, held against the inventory entry and matching the deployed version.
Art.11(1) fixes the artefact and its moment: the technical documentation is drawn up before the system is placed on the market or put into service, kept up to date and contains at minimum the Annex IV elements, so version correspondence is a condition of lawful placing and not a documentation habit. Art.72(3) puts the post-market monitoring plan inside that Annex IV documentation, with a Commission template due by 2 September 2027 (ADR-025), so the plan AIG-018 writes is versioned and produced with this documentation. Art.18(1) sets the retention at 10 years after the system is placed on the market or put into service, covering the technical documentation, the quality management system documentation, the documents and decisions of the notified body and the EU declaration of conformity. One thing here is a relief rather than a duty and is not a row of its own: the Art.11(1) second subparagraph as amended lets an SME, a start-up or a small mid-cap give the Annex IV elements in the Commission's simplified form, requires that form to be used where the option is taken and requires notified bodies to accept it (extract row EU-AI-Art.11.2).
The documentation is the provider's. The deployer's evidence is the released version, held against the inventory entry and matching the deployed version.
Framework Mappings (24)
| DSP-20 | Data Provenance and Transparency | informative |
| GRC-14 | Explainability Evaluation | informative |
| MDS-03 | Model Documentation | partial |
| MDS-04 | Model Documentation Requirements | full |
| MDS-05 | Model Documentation Validation | partial |
| EU-AI-Art.11.1 | Technical Documentation — Preparation and Maintenance | full |
| EU-AI-Art.11.2 | Technical Documentation — Simplified Format for SMEs and Small Mid-Caps | informative |
| EU-AI-Art.18.1 | Documentation Retention — 10-Year Retention Obligation | partial |
| EU-AI-Art.53.1 | GPAI Model Obligations — Technical Documentation | informative |
| COP-S-10 | Additional documentation and transparency | informative |
| COP-S-10.1 | Additional documentation | partial |
| COP-S-7.1 | Model description and behaviour | informative |
| COP-T-1 | Documentation | informative |
| COP-T-1.1 | Drawing up and keeping up-to-date model documentation | informative |
| COP-T-1.2 | Providing relevant information | informative |
| A.4.4 | Tooling resources | partial |
| A.4.5 | System and computing resources | full |
| A.6.2.7 | AI system technical documentation | full |
| SA-5 | System Documentation | informative |
| MG-3.2-002 | Pre-Trained Model Monitoring | MG-3.2-002 | informative |
| MG-3.2-003 | Pre-Trained Model Monitoring | MG-3.2-003 | partial |
| MS-2.5-002 | AI System Validity and Reliability | MS-2.5-002 | partial |
| MS-2.9-002 | AI Model Explainability and Validation | MS-2.9-002 | partial |
| MEASURE 2.8 | AI Transparency and Accountability Risks | partial |
Evidence (2)
AI system technical documentation for each production system, covering system architecture, intended purpose, performance metrics, known limitations, training data summary, failure modes, and human oversight mechanisms.
Example: Technical Documentation · AI Underwriting Assistant v3 (Confluence, version-controlled), covering architecture diagram, intended use cases, accuracy/recall metrics, 8 documented known limitations, training data summary, fail-safe behaviour description, and instructions for oversight personnel
Test: Request the technical documentation for a sample of production AI systems. Verify: (1) every section the control names is present, (2) the documentation is version-controlled and its current version corresponds to the deployed version in the model registry, (3) it is retrievable on request rather than held informally, (4) releases to deployers, operators or competent authorities are recorded with the date and the recipient.
Registry export pairing each production system's deployed version with the version of its technical documentation.
Example: Model registry export, 2026-08-31: 9 production entries, each carrying deployed_version and doc_version with the documentation URI
Test: Export the deployed version and the documentation version for every production AI system. Verify: (1) each production entry carries a documentation reference, (2) the documentation version recorded matches the version of the documentation held, (3) the documentation version is not older than the deployed version, (4) a system deployed in the period carries a documentation version dated on or before its deployment date, (5) entries with no documentation reference are reported rather than filtered out of the export.
Questions (3)
Does technical documentation exist for each AI system before it is deployed?
Technical documentation is the primary evidence artefact for AI governance audits and regulatory inspections. It must be version-controlled and the current version should correspond to the deployed model version.
Which of the following sections are included in your AI system technical documentation?
All eight sections are expected. Known failure modes and human oversight mechanisms are the two sections most often missing from documentation inherited from a general software template. They are also the two an enterprise buyer reads first.
Does the current version of each system's technical documentation correspond to the version of the system in production?
Compare the documentation version against the model or release version recorded in the registry for a sample of systems. Documentation that describes a version no longer deployed fails this question even where every required section is present.