EU AI Act Article 8: Compliance with the Requirements, Implemented in Practice
EU AI Act Article 8 requires a high-risk system to meet the requirements of Articles 9 to 15, measured against intended purpose and state of the art, and at SIFAMO we answer it with a conformity matrix per model that the governance layer generates from the registry.
Article 8 is the shortest article in Chapter III, Section 2, and the one most often skipped. It says that the seven articles that follow must be met together, and it sets the yardstick. It addresses the provider, which for a bank that builds its own credit model means the bank. For me as an engineer, Article 8 is the auditor’s first question: for this one model, show me where the evidence for each requirement lives. The basics on deadlines and roles are in my EU AI Act guide.
For context: the EU AI Act has been in force since 1 August 2024. The prohibitions and AI literacy obligations have applied since February 2025, the obligations for GPAI models since August 2025, the transparency obligations under Article 50 since August 2026. The obligations for high-risk systems under Annex III apply, after the postponement by amending Regulation (EU) 2026/1744 (Digital Omnibus), from 2 December 2027, and for Annex I products from 2 August 2028. Article 8 belongs to these high-risk obligations. The Omnibus moved deadlines; it removed no obligation.
What EU AI Act Article 8 requires
Article 8 has two paragraphs and is the hinge between the requirements in Section 2 and the provider obligations in Section 3:
Art. 8(1), first sentence: High-risk systems must comply with the requirements of this Section, meaning Articles 9 to 15. The yardstick is the intended purpose and the “generally acknowledged state of the art on AI and AI-related technologies”.
Art. 8(1), second sentence: The risk management system under Article 9 must be taken into account when ensuring compliance. The risk register and the evidence for Articles 10 to 15 belong together.
Art. 8(2), first sentence: Where a product contains a high-risk system and also falls under the Union harmonisation legislation in Annex I, Section A, such as machinery or medical devices, the provider is responsible for the product fully meeting all applicable requirements of that legislation.
Art. 8(2), second sentence: To avoid duplication, the provider may integrate the testing, reporting, information and documentation for Section 2 into the procedures and documentation that the Annex I product legislation already requires.
Article 8 has no penalty of its own; it is enforced through Art. 16(a) and the conformity assessment under Article 43.
The technical example
A bank runs a pre-check model in retail lending. It does not decide on the loan; it proposes a route: fast lane through the rules engine, standard review by a case handler, or specialist review. Annex III point 5(b) covers the evaluation of creditworthiness, and as in the Article 9 post we choose the stricter reading: high-risk. The bank’s platform team runs the governance layer for every model, but Article 8 asks about a single one: for this model, in this version, with this intended purpose, are Articles 9 to 15 met?
The application platform calls the gateway with a use-case ID; the gateway resolves the registry record, checks scope and data zone, pseudonymises, calls the model and writes the evidence log. Article 8 bites in two places. First, the intended purpose: a routing model proves accuracy as agreement between its route and the judgement of experienced reviewers, not as prediction of a default. Second, the state of the art: it is a moving target, and harmonised standards are still missing for most requirements.
For contrast, a case the bank does not have: a machinery manufacturer whose AI component acts as a safety component falls under Annex I, Section A. For it, Art. 8(2) applies: the AI requirements run inside the machinery procedure, with a single set of documentation for both acts under Art. 11(2) and tests run once. For the bank, paragraph 2 does not apply; it stays with the internal control under Annex VI.
How we implement it
At SIFAMO we answer Article 8 with an artefact we call the conformity matrix. It is not a document somebody maintains. It is a build output of the governance layer: one row per requirement article, generated from the registry record, the gate results and the dossier. Simplified:
Each row answers four questions: which artefact evidences the requirement, which gate produced it, who owns it, where it lives. The pipeline fills them from the registry, the gate logs, the dossier manifest and the gateway configuration. An empty cell breaks the build. “Not applicable” works only with a reason and the risk function’s signature. For the second sentence of Art. 8(1), every hazard in the risk register points to the row that evidences its measure, for instance proxy discrimination via postcode to Article 10.
Pinning the state of the art
The text does not say who determines the state of the art. We decided: each declaration carries a dated baseline that the risk function reviews. Each row records what it was measured against, in a fixed order of preference: a harmonised standard under Article 40 where one exists, because it triggers the presumption of conformity; otherwise a common specification under Article 41; otherwise, today almost always, a documented own solution under Annex IV point 7: which method, checked against which draft standards, supervisory publications and published methods, why sufficient for this intended purpose. We read “generally acknowledged” as what a competent peer would expect at the date of the baseline. The auditor then argues about a date and a justification, not about a moving target. At the expiry date the baseline is reviewed again.
From the matrix to Annex VI and Article 47
Annex VI point 3 requires the technical documentation to show that Section 2 is met. For us that means: all seven rows carry evidence with a hash, the baseline has been reviewed. Only then does the pipeline generate the Article 47 declaration per Annex V; it names the matrix version and hash, and its point 7 takes the list of standards from the baseline. Three people sign: business owner, risk function, IT. The declaration ID is stored in the registry record, and the gateway routes no version without a valid declaration. Conformity is a runtime condition, not a filing note.
Re-run on every change
Every change to model, prompt, feature list or retrieval index changes the version fields in the registry and triggers the gates. The matrix is then rebuilt and shown as a diff against the previous version. If the change lies within the change framework the dossier describes in advance under Annex IV point 2(f), for example retraining on the same feature list within declared bands, the declaration stays valid. If it is a substantial modification under Art. 3(23) and Art. 43(4), such as a new feature or a new customer group, there is a new declaration or no route. The expiry date forces the full run regardless of changes.
For the machinery manufacturer we generate the same seven rows, but as a chapter in the technical file under machinery law. That is exactly what the second sentence of Art. 8(2) permits. What the auditor sees: seven evidenced rows, the dated baseline with a justification per row, the signed declaration with the matrix hash, and the history of rebuilds. Each row is described in its own post of this series: Article 9 (risk management system), Article 10 (data and data governance), Article 11 (technical documentation under Annex IV), Article 12 (record-keeping), Article 13 (transparency for deployers), Article 14 (human oversight) and Article 15 (robustness and cybersecurity). More on the principle under AI Governance.
Typical mistakes
Seven folders instead of one matrix. Every article has its storage, but nobody can say for a specific version whether all seven fit together.
State of the art without a date. “Reflects the state of the art” appears in the dossier, but not as of which date and measured against what.
Intended purpose written by marketing. “Supports the credit decision” is not an intended purpose, and without a sharp purpose accuracy and oversight cannot be evidenced.
EU AI Act Article 8 checklist
Every high-risk model has a conformity matrix with one row per Article 9 to 15, generated from the registry.
Every row names the artefact, the gate that produced it, the owner and the storage location with hash; empty cells break the build.
The intended purpose in the registry is sharp enough to derive the accuracy metrics and the oversight tiers from it.
Each row documents what it was measured against: harmonised standard, common specification or own solution under Annex IV point 7, with a dated baseline.
The declaration under Article 47 references the matrix version and hash, and the gateway routes no version without a valid declaration.
Every change rebuilds the matrix; substantial modifications and the expiry date force a new declaration.
Open questions
What does “generally acknowledged” mean without standards? As long as harmonised standards are missing for most requirements, every provider sets its own baseline.
Two procedures, one body. For Annex I products the product legislation’s notified body also examines the AI requirements. How it judges the state of the art for AI is open.
Is Article 8 an obligation of its own? It states a yardstick; the penalty hangs on Articles 16 and 99. Whether an authority can demand the matrix itself, supervisory practice will show.
Frequently asked questions
Does this apply to us as a deployer too?
Article 8 addresses the provider. As a deployer under Art. 26 you ask for the declaration of conformity and the instructions for use and you run the oversight, but you do not keep a matrix of your own. As soon as you build the system yourself or substantially modify it, you are the provider under Art. 25, and the matrix is yours.
Do we have to apply harmonised standards?
No. Standards are voluntary and under Article 40 only trigger the presumption that the requirement is met. If you do not apply them, you describe the solution chosen instead under Annex IV point 7. Today that is the normal case.
Is our model vendor’s certification enough?
No. The requirements apply to your system in your intended purpose: pre-processing, model, rules engine, oversight. Vendor documents are input for individual rows, marked as third-party statements.
EU AI Act Article 8: Compliance with the Requirements, Implemented in Practice
EU AI Act Article 8 requires a high-risk system to meet the requirements of Articles 9 to 15, measured against intended purpose and state of the art, and at SIFAMO we answer it with a conformity matrix per model that the governance layer generates from the registry.
Article 8 is the shortest article in Chapter III, Section 2, and the one most often skipped. It says that the seven articles that follow must be met together, and it sets the yardstick. It addresses the provider, which for a bank that builds its own credit model means the bank. For me as an engineer, Article 8 is the auditor’s first question: for this one model, show me where the evidence for each requirement lives. The basics on deadlines and roles are in my EU AI Act guide.
For context: the EU AI Act has been in force since 1 August 2024. The prohibitions and AI literacy obligations have applied since February 2025, the obligations for GPAI models since August 2025, the transparency obligations under Article 50 since August 2026. The obligations for high-risk systems under Annex III apply, after the postponement by amending Regulation (EU) 2026/1744 (Digital Omnibus), from 2 December 2027, and for Annex I products from 2 August 2028. Article 8 belongs to these high-risk obligations. The Omnibus moved deadlines; it removed no obligation.
What EU AI Act Article 8 requires
Article 8 has two paragraphs and is the hinge between the requirements in Section 2 and the provider obligations in Section 3:
Article 8 has no penalty of its own; it is enforced through Art. 16(a) and the conformity assessment under Article 43.
The technical example
A bank runs a pre-check model in retail lending. It does not decide on the loan; it proposes a route: fast lane through the rules engine, standard review by a case handler, or specialist review. Annex III point 5(b) covers the evaluation of creditworthiness, and as in the Article 9 post we choose the stricter reading: high-risk. The bank’s platform team runs the governance layer for every model, but Article 8 asks about a single one: for this model, in this version, with this intended purpose, are Articles 9 to 15 met?
The application platform calls the gateway with a use-case ID; the gateway resolves the registry record, checks scope and data zone, pseudonymises, calls the model and writes the evidence log. Article 8 bites in two places. First, the intended purpose: a routing model proves accuracy as agreement between its route and the judgement of experienced reviewers, not as prediction of a default. Second, the state of the art: it is a moving target, and harmonised standards are still missing for most requirements.
For contrast, a case the bank does not have: a machinery manufacturer whose AI component acts as a safety component falls under Annex I, Section A. For it, Art. 8(2) applies: the AI requirements run inside the machinery procedure, with a single set of documentation for both acts under Art. 11(2) and tests run once. For the bank, paragraph 2 does not apply; it stays with the internal control under Annex VI.
How we implement it
At SIFAMO we answer Article 8 with an artefact we call the conformity matrix. It is not a document somebody maintains. It is a build output of the governance layer: one row per requirement article, generated from the registry record, the gate results and the dossier. Simplified:
One row, four fields, no empty cell
Each row answers four questions: which artefact evidences the requirement, which gate produced it, who owns it, where it lives. The pipeline fills them from the registry, the gate logs, the dossier manifest and the gateway configuration. An empty cell breaks the build. “Not applicable” works only with a reason and the risk function’s signature. For the second sentence of Art. 8(1), every hazard in the risk register points to the row that evidences its measure, for instance proxy discrimination via postcode to Article 10.
Pinning the state of the art
The text does not say who determines the state of the art. We decided: each declaration carries a dated baseline that the risk function reviews. Each row records what it was measured against, in a fixed order of preference: a harmonised standard under Article 40 where one exists, because it triggers the presumption of conformity; otherwise a common specification under Article 41; otherwise, today almost always, a documented own solution under Annex IV point 7: which method, checked against which draft standards, supervisory publications and published methods, why sufficient for this intended purpose. We read “generally acknowledged” as what a competent peer would expect at the date of the baseline. The auditor then argues about a date and a justification, not about a moving target. At the expiry date the baseline is reviewed again.
From the matrix to Annex VI and Article 47
Annex VI point 3 requires the technical documentation to show that Section 2 is met. For us that means: all seven rows carry evidence with a hash, the baseline has been reviewed. Only then does the pipeline generate the Article 47 declaration per Annex V; it names the matrix version and hash, and its point 7 takes the list of standards from the baseline. Three people sign: business owner, risk function, IT. The declaration ID is stored in the registry record, and the gateway routes no version without a valid declaration. Conformity is a runtime condition, not a filing note.
Re-run on every change
Every change to model, prompt, feature list or retrieval index changes the version fields in the registry and triggers the gates. The matrix is then rebuilt and shown as a diff against the previous version. If the change lies within the change framework the dossier describes in advance under Annex IV point 2(f), for example retraining on the same feature list within declared bands, the declaration stays valid. If it is a substantial modification under Art. 3(23) and Art. 43(4), such as a new feature or a new customer group, there is a new declaration or no route. The expiry date forces the full run regardless of changes.
For the machinery manufacturer we generate the same seven rows, but as a chapter in the technical file under machinery law. That is exactly what the second sentence of Art. 8(2) permits. What the auditor sees: seven evidenced rows, the dated baseline with a justification per row, the signed declaration with the matrix hash, and the history of rebuilds. Each row is described in its own post of this series: Article 9 (risk management system), Article 10 (data and data governance), Article 11 (technical documentation under Annex IV), Article 12 (record-keeping), Article 13 (transparency for deployers), Article 14 (human oversight) and Article 15 (robustness and cybersecurity). More on the principle under AI Governance.
Typical mistakes
EU AI Act Article 8 checklist
Open questions
Frequently asked questions
Does this apply to us as a deployer too?
Article 8 addresses the provider. As a deployer under Art. 26 you ask for the declaration of conformity and the instructions for use and you run the oversight, but you do not keep a matrix of your own. As soon as you build the system yourself or substantially modify it, you are the provider under Art. 25, and the matrix is yours.
Do we have to apply harmonised standards?
No. Standards are voluntary and under Article 40 only trigger the presumption that the requirement is met. If you do not apply them, you describe the solution chosen instead under Annex IV point 7. Today that is the normal case.
Is our model vendor’s certification enough?
No. The requirements apply to your system in your intended purpose: pre-processing, model, rules engine, oversight. Vendor documents are input for individual rows, marked as third-party statements.
This post is part of the series on the EU AI Act article by article, all articles at a glance.
Related: EU AI Act · All articles · AI Governance · AI Consultant
Recent Posts
Recent Comments