EU AI Act Article 43: Conformity Assessment, Implemented in Practice
EU AI Act Article 43 requires a conformity assessment for every high-risk AI system before go-live, for most Annex III systems as internal control under Annex VI. Our answer: a pipeline that runs at every release. The quality gates produce the evidence, the documentation is built, the declaration of conformity under Article 47 is a signed registry record, and the CE marking under Article 48 hangs off it.
Article 43 is the article where many people picture a vehicle inspection for AI and wait for an external testing body. For most high-risk systems in Annex III, none comes. The provider assesses itself, declares the result and is liable for it. The addressee is the provider, via Art. 16(f), (g) and (h). I cover Article 43 together with Articles 47 and 48, because in practice the three are one process.
The EU AI Act has been in force since 1 August 2024. The prohibitions and the AI literacy duty 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 the amending Regulation (EU) 2026/1744 (Digital Omnibus), from 2 December 2027, and for Annex I products from 2 August 2028. Article 43 belongs to these high-risk obligations. The Omnibus moved deadlines; it removed no obligation.
What EU AI Act Article 43 requires
Six paragraphs, plus Annexes VI and VII and the follow-on Articles 47 and 48.
Art. 43(1): Biometric systems under Annex III point 1. The provider chooses between internal control under Annex VI and the procedure with a notified body under Annex VII, but only if it fully applies harmonised standards under Article 40 or common specifications under Article 41. Where they are missing or applied only in part, Annex VII is mandatory; for law enforcement and migration, the market surveillance authority takes that role.
Art. 43(2): All other Annex III systems, points 2 to 8, including creditworthiness (point 5(b)): internal control under Annex VI, no notified body.
Art. 43(3): Annex I products. The assessment runs in the procedure of the relevant product law and checks the requirements of Chapter III Section 2 alongside. The Omnibus adds a deadline: the notified bodies under those acts apply for designation by 28 January 2028.
Art. 43(4): Every substantial modification triggers a new assessment. Exception: changes to continuously learning systems that the provider pre-determined and described in the technical documentation. Substantial, under Art. 3(23), is an unforeseen change that affects compliance or alters the intended purpose.
Art. 43(5) and (6): The Commission may amend Annexes VI and VII and may extend the notified-body procedure to points 2 to 8 if internal control does not adequately prevent risks.
Annex VI, internal control: Three checks by the provider itself. The quality management system complies with Article 17 (point 2), the technical documentation shows that Section 2 is met (point 3), the development process and the post-market monitoring under Article 72 are consistent with the documentation (point 4).
Art. 47, EU declaration of conformity: One written, machine-readable, signed declaration per system, kept available to the authorities for ten years (para. 1), stating that Section 2 is met, with the content of Annex V and in a language the authorities understand (para. 2), as a single declaration where several acts apply (para. 3). With it the provider assumes responsibility and keeps it up to date (para. 4). Annex V requires eight items, from identification of the system via sole responsibility, data protection statement and standards applied to place, date and signature.
Art. 48, CE marking: Visible, legible, indelible (para. 3), for digitally provided systems digital through the interface or a machine-readable code (para. 2), followed by the identification number of the notified body where there was one (para. 4).
On the standards question. Article 40 gives harmonised standards a presumption of conformity, Article 41 lets the Commission adopt common specifications as a fallback. According to the Commission, the first harmonised standard was published in October 2026, the others are still in progress, and there are no common specifications. Most providers therefore assess today against Articles 9 to 15 themselves and document under Annex IV point 7 which solution they chose in place of a standard.
The technical example
A bank builds its retail credit pre-check itself: a model trained on historical applications and repayment histories returns, per application, a score, a risk class and the driving factors. A human makes the decision. That is Annex III point 5(b). Because the bank develops the system itself and puts it into service for its own use, it is provider under Art. 3(3) and deployer at the same time, and under Art. 43(2) internal control applies.
The data flow: the application journey calls the policy gateway with a use-case ID, never the model. The gateway resolves the ID in the registry, checks purpose and data zone, pseudonymises the customer data and passes on only released features. Below the confidence threshold, or where model and rules engine disagree, the case goes to full review. Every recommendation lands in the evidence log, every deviation by the case handler flows back as a signal.
For comparison: the same bank wants to recognise people in branch camera footage against a reference database of known fraud cases. That is remote biometric identification under Annex III point 1(a), if data protection law allows it at all. Because harmonised standards for biometrics are missing, Art. 43(1) leads straight to Annex VII with a notified body. Video identification at onboarding does not fall here: it only verifies a claimed identity, and point 1(a) excludes it.
Where does Article 43 bite? Before the first go-live and at every substantial modification. Retraining on the new quarter is routine; a new feature, a new customer group such as the self-employed or a switch of model family is a different system. Without a rule that separates the two, every retraining is formally a new assessment, or none is.
How we implement it
At SIFAMO the conformity assessment is not a meeting with a projector but a pipeline run in the governance layer. The three checks in Annex VI are three stages with a machine-verifiable result.
Annex VI as three stages
Stage one, QMS: the processes of the governance layer are the quality management system under Article 17, versioned in the repository, and the pipeline checks which version the system was built against. Stage two, documentation: the dossier build under Article 11 runs, and on top of it the conformity matrix, one row per requirement in Articles 9 to 15 with evidence, standard applied or own solution, reviewer and result. An empty row breaks the run. Stage three, process against documentation: the pipeline compares deployment manifests and monitoring configuration with the dossier, from the pinned versions to the plan under Article 72. If production deviates, there is no declaration.
The quality gates produce the evidence
The matrix invents nothing, it points. The evaluation on the business-owned test set evidences Article 15, the bias run with varied protected attributes Art. 10(2)(f), the adversarial tests Art. 15(5), the rehearsed rollback Article 9, the oversight configuration Article 14, the evidence log Article 12, each with a dated, signed report. The three-way sign-off by business, risk and IT closes the gates. It is not yet the declaration.
The declaration of conformity is a signed registry record
The declaration under Article 47 is, in our setup, a record in the registry with the eight items from Annex V, electronically signed. The AI Act does not say who signs. We chose a member of the management board designated by the provider, because the declaration binds the company, not a team. Simplified:
The record sits write-once next to the dossier, for ten years under Art. 18(1), every version is kept, and the authority reads it in the reading portal. The registration under Article 49 points to the same declaration ID.
CE marking and route
For a system provided purely digitally, the CE marking under Art. 48(2) is digital: in the case handler interface and at a machine-readable endpoint that returns the declaration ID and the dossier hash, with no identification number. And the gateway reads the registry record on every call: a version without a valid declaration gets no route.
Substantial modification: the trigger
For Art. 43(4) the dossier carries, under Annex IV point 2(f), a change envelope: for example quarterly retraining on the same feature list and architecture, as long as accuracy and bias metrics stay within declared bands. A change inside the envelope runs through the gates and attaches its reports to the existing declaration. Anything outside it, new features, new customer group, changed purpose, risk classifies as substantial, and the pipeline starts again at stage one with a new declaration version. When in doubt, substantial, because a second assessment costs us hours, not months.
When a notified body joins
For the biometric system, the body receives in the reading portal dossier, matrix and QMS documentation, on request access to training data under Annex VII point 4.3, and to the model only as a last resort under point 4.5. Certificate and body go into item 7 of the declaration, the identification number behind the CE marking, and the audits under point 5 become dates in the registry. More on the principle under AI governance.
Typical mistakes
Waiting for the auditor. For Annex III points 2 to 8, none comes.
The declaration first. A signed PDF without matrix and evidence is responsibility without cover.
No change envelope. Without pre-determined changes, every retraining is formally substantial, and in practice nothing gets reassessed.
CE marking as a free pass. Whoever retrains a purchased system or extends its purpose becomes a provider under Art. 25(1).
Article 43 checklist
Is it determined and justified for every high-risk system which procedure applies: Annex VI, Annex VII or product law?
Is there a conformity matrix per requirement in Articles 9 to 15 with evidence, standard or own solution, reviewer and result?
Does a signed declaration with all eight Annex V items exist for the running version, machine-readable and kept for ten years?
Is the CE marking affixed, for digital systems accessible through the interface or a machine-readable code?
Does the dossier state which changes are pre-determined, and who classifies a change as substantial?
Does the gateway refuse a route to versions without a valid declaration, and is the system registered under Article 49?
Open questions
Presumption without standards. Until the standards cover the requirements, every provider assesses against the legal text, and two authorities can judge the same evidence differently.
Where does retraining end? Art. 3(23) and Art. 43(4) give a frame, not a threshold. How wide a change envelope may be, only supervisory practice will show.
Paragraph 6. If the Commission extends the notified-body procedure to points 2 to 8, credit scoring needs a body too, and their capacity is open today.
Frequently asked questions
Does this apply to us as a deployer too?
Articles 43, 47 and 48 address the provider. As a bank buying a scoring system, you check under Article 26 that declaration and CE marking are in place. If you build it yourself, retrain it or change its purpose, you are the provider yourself under Art. 3(3) or Article 25.
Do we need a notified body?
For Annex III points 2 to 8, no, internal control under Annex VI applies. For biometrics under point 1, in practice yes, as long as no harmonised standards are fully applicable. For Annex I products, product law decides.
Do we have to reassess after every retraining?
Only if it is a substantial modification. What you pre-determined in the dossier and what stays within the declared bands is not one. That is why you write the change envelope before the first go-live.
EU AI Act Article 43: Conformity Assessment, Implemented in Practice
EU AI Act Article 43 requires a conformity assessment for every high-risk AI system before go-live, for most Annex III systems as internal control under Annex VI. Our answer: a pipeline that runs at every release. The quality gates produce the evidence, the documentation is built, the declaration of conformity under Article 47 is a signed registry record, and the CE marking under Article 48 hangs off it.
Article 43 is the article where many people picture a vehicle inspection for AI and wait for an external testing body. For most high-risk systems in Annex III, none comes. The provider assesses itself, declares the result and is liable for it. The addressee is the provider, via Art. 16(f), (g) and (h). I cover Article 43 together with Articles 47 and 48, because in practice the three are one process.
The EU AI Act has been in force since 1 August 2024. The prohibitions and the AI literacy duty 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 the amending Regulation (EU) 2026/1744 (Digital Omnibus), from 2 December 2027, and for Annex I products from 2 August 2028. Article 43 belongs to these high-risk obligations. The Omnibus moved deadlines; it removed no obligation.
What EU AI Act Article 43 requires
Six paragraphs, plus Annexes VI and VII and the follow-on Articles 47 and 48.
On the standards question. Article 40 gives harmonised standards a presumption of conformity, Article 41 lets the Commission adopt common specifications as a fallback. According to the Commission, the first harmonised standard was published in October 2026, the others are still in progress, and there are no common specifications. Most providers therefore assess today against Articles 9 to 15 themselves and document under Annex IV point 7 which solution they chose in place of a standard.
The technical example
A bank builds its retail credit pre-check itself: a model trained on historical applications and repayment histories returns, per application, a score, a risk class and the driving factors. A human makes the decision. That is Annex III point 5(b). Because the bank develops the system itself and puts it into service for its own use, it is provider under Art. 3(3) and deployer at the same time, and under Art. 43(2) internal control applies.
The data flow: the application journey calls the policy gateway with a use-case ID, never the model. The gateway resolves the ID in the registry, checks purpose and data zone, pseudonymises the customer data and passes on only released features. Below the confidence threshold, or where model and rules engine disagree, the case goes to full review. Every recommendation lands in the evidence log, every deviation by the case handler flows back as a signal.
For comparison: the same bank wants to recognise people in branch camera footage against a reference database of known fraud cases. That is remote biometric identification under Annex III point 1(a), if data protection law allows it at all. Because harmonised standards for biometrics are missing, Art. 43(1) leads straight to Annex VII with a notified body. Video identification at onboarding does not fall here: it only verifies a claimed identity, and point 1(a) excludes it.
Where does Article 43 bite? Before the first go-live and at every substantial modification. Retraining on the new quarter is routine; a new feature, a new customer group such as the self-employed or a switch of model family is a different system. Without a rule that separates the two, every retraining is formally a new assessment, or none is.
How we implement it
At SIFAMO the conformity assessment is not a meeting with a projector but a pipeline run in the governance layer. The three checks in Annex VI are three stages with a machine-verifiable result.
Annex VI as three stages
Stage one, QMS: the processes of the governance layer are the quality management system under Article 17, versioned in the repository, and the pipeline checks which version the system was built against. Stage two, documentation: the dossier build under Article 11 runs, and on top of it the conformity matrix, one row per requirement in Articles 9 to 15 with evidence, standard applied or own solution, reviewer and result. An empty row breaks the run. Stage three, process against documentation: the pipeline compares deployment manifests and monitoring configuration with the dossier, from the pinned versions to the plan under Article 72. If production deviates, there is no declaration.
The quality gates produce the evidence
The matrix invents nothing, it points. The evaluation on the business-owned test set evidences Article 15, the bias run with varied protected attributes Art. 10(2)(f), the adversarial tests Art. 15(5), the rehearsed rollback Article 9, the oversight configuration Article 14, the evidence log Article 12, each with a dated, signed report. The three-way sign-off by business, risk and IT closes the gates. It is not yet the declaration.
The declaration of conformity is a signed registry record
The declaration under Article 47 is, in our setup, a record in the registry with the eight items from Annex V, electronically signed. The AI Act does not say who signs. We chose a member of the management board designated by the provider, because the declaration binds the company, not a team. Simplified:
The record sits write-once next to the dossier, for ten years under Art. 18(1), every version is kept, and the authority reads it in the reading portal. The registration under Article 49 points to the same declaration ID.
CE marking and route
For a system provided purely digitally, the CE marking under Art. 48(2) is digital: in the case handler interface and at a machine-readable endpoint that returns the declaration ID and the dossier hash, with no identification number. And the gateway reads the registry record on every call: a version without a valid declaration gets no route.
Substantial modification: the trigger
For Art. 43(4) the dossier carries, under Annex IV point 2(f), a change envelope: for example quarterly retraining on the same feature list and architecture, as long as accuracy and bias metrics stay within declared bands. A change inside the envelope runs through the gates and attaches its reports to the existing declaration. Anything outside it, new features, new customer group, changed purpose, risk classifies as substantial, and the pipeline starts again at stage one with a new declaration version. When in doubt, substantial, because a second assessment costs us hours, not months.
When a notified body joins
For the biometric system, the body receives in the reading portal dossier, matrix and QMS documentation, on request access to training data under Annex VII point 4.3, and to the model only as a last resort under point 4.5. Certificate and body go into item 7 of the declaration, the identification number behind the CE marking, and the audits under point 5 become dates in the registry. More on the principle under AI governance.
Typical mistakes
Article 43 checklist
Open questions
Frequently asked questions
Does this apply to us as a deployer too?
Articles 43, 47 and 48 address the provider. As a bank buying a scoring system, you check under Article 26 that declaration and CE marking are in place. If you build it yourself, retrain it or change its purpose, you are the provider yourself under Art. 3(3) or Article 25.
Do we need a notified body?
For Annex III points 2 to 8, no, internal control under Annex VI applies. For biometrics under point 1, in practice yes, as long as no harmonised standards are fully applicable. For Annex I products, product law decides.
Do we have to reassess after every retraining?
Only if it is a substantial modification. What you pre-determined in the dossier and what stays within the declared bands is not one. That is why you write the change envelope before the first go-live.
This post belongs to my series EU AI Act in practice. Here you find all articles at a glance.
Related: EU AI Act · All articles · AI Governance · AI Consultant
Recent Posts
Recent Comments