EU AI Act Article 26: Deployer Obligations, Implemented in Practice
EU AI Act Article 26 shifts responsibility for a high-risk system from the manufacturer to the company that uses it: use it as instructed, staff the oversight, check the inputs, monitor, keep the logs, inform the people affected. We turn each of these duties into a control in the governance layer, with the registry record as the anchor.
Most of the companies I work with do not build high-risk systems. They buy them. That makes them deployers under Art. 3(4), and Article 26 of the AI Regulation is the article that hits them directly. It is less well known than the provider obligations in Articles 9 to 15, and it closes the gap between “the provider has a declaration of conformity” and “we know what the system does in our shop”. It is addressed to the deployer, in paragraph 7 as employer, in paragraph 8 as public authority, in paragraph 10 as law enforcement authority.
The EU AI Act has been in force since 1 August 2024. Prohibitions and AI literacy have applied since February 2025, the obligations for GPAI models since August 2025, the transparency obligations under Article 50 since August 2026. After the postponement by the amending Regulation (EU) 2026/1744 (Digital Omnibus), the obligations for high-risk systems under Annex III apply from 2 December 2027, for Annex I products from 2 August 2028. Article 26 is one of those high-risk obligations. The Omnibus moved deadlines; it removed no obligation. The overview is in EU AI Act in practice.
What Article 26 requires
Art. 26(1): The deployer takes appropriate technical and organisational measures to use the system in accordance with the instructions for use that accompany it. The instructions are the provider’s document under Article 13.
Art. 26(2): Human oversight is assigned to natural persons who have the necessary competence, training and authority, as well as the necessary support.
Art. 26(3): Other obligations under Union or national law are unaffected, as is the deployer’s freedom to organise its own resources for the oversight measures the provider has indicated.
Art. 26(4): To the extent the deployer exercises control over the input data, it ensures that the data is relevant and sufficiently representative in view of the intended purpose.
Art. 26(5): The deployer monitors operation on the basis of the instructions for use and, where relevant, informs the provider in accordance with Art. 72. If use as instructed points to a risk within the meaning of Art. 79(1), it informs the provider or distributor and the market surveillance authority without undue delay and suspends use. A serious incident it reports first to the provider, then to the importer or distributor and the authorities; if the provider cannot be reached, Art. 73 applies mutatis mutandis. For financial institutions the monitoring obligation is deemed fulfilled by complying with the governance rules of Union financial services law.
Art. 26(6): Automatically generated logs are kept, to the extent they are under the deployer’s control, for a period appropriate to the intended purpose, of at least six months, unless Union or national law, in particular data protection law, provides otherwise. Financial institutions keep them as part of their supervisory documentation.
Art. 26(7): Before putting the system into service or using it at the workplace, employers inform workers’ representatives and the affected workers, following the applicable rules on informing workers.
Art. 26(8): Public authorities and Union bodies comply with the registration obligations under Art. 49. If the system is not registered in the EU database under Art. 71, they must not use it and inform the provider or distributor.
Art. 26(9): Where applicable, the deployer uses the information under Art. 13 to carry out its data protection impact assessment under Art. 35 GDPR or Art. 27 of Directive (EU) 2016/680.
Art. 26(10): Special rules for post-remote biometric identification in law enforcement, including prior authorisation and annual reports. Not relevant for the insurer.
Art. 26(11): Deployers of Annex III systems that make decisions about natural persons, or assist in making them, inform those persons that they are subject to the use of the system, without prejudice to Article 50. Art. 86 builds on this: the right to a clear and meaningful explanation of the role of the system and the main elements of the decision.
Art. 26(12): The deployer cooperates with the competent authorities in any action they take in relation to the system.
The technical example
A health insurer buys a claims-triage model from a vendor. Every incoming medical invoice is sorted into one of three classes: reimburse automatically, send to a case handler, or flag as suspicious and send to claims review. The vendor supplies the model, the instructions for use and the declaration of conformity, the insurer the data, the people and the process; it has classified the system as high-risk. The boundary of Annex III point 5(c), which names risk assessment and pricing in health and life insurance, is debatable; classification is the subject of Article 6, not of this post.
The data flow: the invoice arrives in the claims system and goes to the policy gateway, which pseudonymises, so insured-person ID instead of name, date of birth and address, checks data zone and scope and calls the vendor model. The model returns class, confidence and the features that carried the result; the gateway writes the evidence record, maps the ID back and delivers to the claims system. “Reimburse automatically” runs with sampling, “suspicious” always lands with a human.
Where the obligation bites: the insurer did not train the model and does not see the training data. It still has to know whether its invoices match what the model was built for (paragraph 4), whether it drifts (paragraph 5) and what it logs (paragraph 6). And it has to tell case handlers and insured persons that the system is there (paragraphs 7 and 11). The vendor’s declaration of conformity answers none of these questions.
How we implement it
At SIFAMO we treat Article 26 not as a documentation duty but as a set of controls in the governance layer. The anchor is the registry record, and it belongs to the insurer, not to the vendor.
Instructions for use as a registry record (paragraph 1). The vendor’s instructions are not filed, they are imported: intended purpose, permitted inputs, known limitations, accuracy figures, indicated oversight measures and logging functions become fields in the record. From those fields we derive the scope the gateway checks on every call. An invoice from a tariff the instructions do not name gets no route. The instructions become configuration, not a PDF on a shared drive. The business owner is the head of claims, and she signs the record.
Oversight with names (paragraph 2). The oversight field names roles, not departments: who reviews suspicious cases, who draws samples, who may stop the system. Each role carries a training record with the date and the version of the instructions. Authority means the reviewer overrides the model’s class without asking anyone and sees the invoice before the model’s score. Support means review time is budgeted, and seeded error cases measure whether reviewers catch them. Paragraph 3 leaves the organisation to us, so we tier by decision type.
Input checks at the gateway (paragraph 4). The insurer controls the input data, so paragraph 4 applies. The gateway checks every invoice against the input schema from the instructions: invoice type, country, currency, language, tariff type, completeness. Whatever does not fit goes to a case handler without a model. Representativeness only shows in the population, so monitoring compares, weekly, the distribution of incoming invoices with the evaluation set the business owns and with the population the vendor describes in the instructions. If it drifts, say after a new tariff launches, the business owner gets an alert before the model quietly degrades.
Monitoring and reporting (paragraph 5). Monitoring reads the evidence log and the gateway. The registry record holds the vendor’s incident contact and the format in which it expects feedback for its post-market monitoring under Art. 72. Two thresholds are defined. “Risk under Art. 79(1)”: the business owner sets the stop flag, informs the vendor and the market surveillance authority, and invoices continue manually. “Serious incident” under Art. 3(49): the vendor first, then the authority, and if the vendor cannot be reached we report ourselves under Art. 73. As a financial institution, the insurer’s monitoring obligation is deemed fulfilled through the governance rules of Solvency II. We build the same thing anyway, because those rules demand monitoring and the reporting duties remain.
Logs for at least six months (paragraph 6). The logs the vendor model generates itself under Article 12 we have delivered to us by contract, otherwise they are not under our control. The gateway’s evidence log is ours: per invoice the input hash, model and instructions version, policy version, output, signals, reviewer action. Six months is a floor. We set the retention period per registry record, here the lifetime of the claim plus the objection period, and justify it there. Because the logs contain only IDs, the period does not collide with GDPR storage limitation.
Informing workers (paragraph 7). The case handlers are subject to the system: it sorts their work, and the log records who overrode which class. Informing them before go-live is a gate: without the date and the document of the information given to the works council and the team there is no registration. Because of co-determination we store reviewer actions under a role ID and restrict per-person analysis through access rights.
Public bodies (paragraph 8) and DPIA (paragraph 9). For a private insurer paragraph 8 does not apply; whether a statutory health insurer counts as a public authority is a question I would put to legal. Technically it is one field: the record of a public body requires the identifier from the EU database under Art. 71, and without it there is no route. The data protection impact assessment refers to the registry record and the imported instructions, describes pseudonymisation, data zones and oversight tiers as its measures, has a version and is reopened on a model change.
Informing and explaining to affected persons (paragraph 11 and Art. 86). The acknowledgement letter to the insured person states that an AI system pre-sorts the invoice and that humans decide. For Art. 86 the insurer has to be able to explain the role of the system and the main elements of the decision. We generate that from the evidence record: which class the model proposed, which rule or policy applied, whether and how a human decided. We do not hand out the model’s own reasoning text as the explanation, because we cannot verify that it is true.
The auditor sees the registry record with imported instructions, roles, training status, retention period and stop authorisations, per invoice the evidence record, the drift reports, every report with a timestamp and the information given to workers. Cooperation under paragraph 12 is then an export, not a project. I am an engineer, not a lawyer. But a deployer who can show this holds Article 26 in its own hands, not its vendor’s.
Typical mistakes
“The vendor has the declaration of conformity.” It covers Articles 9 to 15, not Article 26. What the system does inside your process is something the vendor cannot explain.
Six months as a deletion deadline. The minimum becomes the maximum, and when the objection arrives two years later the evidence is gone.
Works council after go-live. The conversation is held late, and the first conflict is about performance monitoring, not about AI.
Quietly becoming a provider. Whoever retrains the model, changes its intended purpose or offers it under its own name becomes a provider under Art. 25(1), with the obligations of Article 16.
EU AI Act Article 26 checklist
The vendor’s instructions for use as structured fields in the registry record; scope derived from them and enforced by the gateway.
Oversight roles with names, training records and override authority; seeded error cases measure effectiveness.
Input schema at the gateway; distribution compared with the evaluation set and the vendor’s population, with an alert to the business owner.
The vendor’s incident contact, the reporting path to the market surveillance authority and the stop flag defined and rehearsed once.
Retention period per record, at least six months, justified and agreed with data protection; vendor logs secured by contract.
Information to workers’ representatives and workers before go-live documented and anchored as a gate.
Notice to affected persons built into the process; explanation under Art. 86 can be generated from the evidence log.
Open questions
Whether claims handling in health insurance falls under Annex III point 5(c), which names risk assessment and pricing, is unsettled. We treat such systems as high-risk because Article 26 costs less than the dispute afterwards.
“Relevant and sufficiently representative” in paragraph 4 asks the deployer for a comparison with a training population it does not see. What the vendor has to supply for that under Art. 13 is not yet spelled out by the standards.
How detailed an explanation under Art. 86 has to be, and whether it may consist of model outputs nobody can verify, is open.
Frequently asked questions
Does this apply to us as a deployer if the vendor has declared conformity?
Yes. The declaration of conformity covers the provider’s obligations; Article 26 addresses you as deployer regardless: paragraphs 1 to 6, 9, 11 and 12 for every deployer of a high-risk system, paragraph 7 for employers, paragraph 8 for public bodies, paragraph 10 for law enforcement. More on this under AI governance.
Are six months enough for the logs?
Six months is the floor in paragraph 6, not the target. The period has to be appropriate to the purpose, and for a benefits decision an insured person can challenge years later, it is longer. Justify it in the registry record and agree it with data protection, to which paragraph 6 gives precedence.
Do we have to tell our insured persons that an AI sorts their invoice?
Yes, if the system is an Annex III system and makes decisions about persons or assists in making them. Paragraph 11 requires the notice, and Art. 86 additionally gives the person the right to an explanation of the role of the system and the main elements of the decision. The other articles of the series: all articles at a glance.
EU AI Act Article 26: Deployer Obligations, Implemented in Practice
EU AI Act Article 26 shifts responsibility for a high-risk system from the manufacturer to the company that uses it: use it as instructed, staff the oversight, check the inputs, monitor, keep the logs, inform the people affected. We turn each of these duties into a control in the governance layer, with the registry record as the anchor.
Most of the companies I work with do not build high-risk systems. They buy them. That makes them deployers under Art. 3(4), and Article 26 of the AI Regulation is the article that hits them directly. It is less well known than the provider obligations in Articles 9 to 15, and it closes the gap between “the provider has a declaration of conformity” and “we know what the system does in our shop”. It is addressed to the deployer, in paragraph 7 as employer, in paragraph 8 as public authority, in paragraph 10 as law enforcement authority.
The EU AI Act has been in force since 1 August 2024. Prohibitions and AI literacy have applied since February 2025, the obligations for GPAI models since August 2025, the transparency obligations under Article 50 since August 2026. After the postponement by the amending Regulation (EU) 2026/1744 (Digital Omnibus), the obligations for high-risk systems under Annex III apply from 2 December 2027, for Annex I products from 2 August 2028. Article 26 is one of those high-risk obligations. The Omnibus moved deadlines; it removed no obligation. The overview is in EU AI Act in practice.
What Article 26 requires
The technical example
A health insurer buys a claims-triage model from a vendor. Every incoming medical invoice is sorted into one of three classes: reimburse automatically, send to a case handler, or flag as suspicious and send to claims review. The vendor supplies the model, the instructions for use and the declaration of conformity, the insurer the data, the people and the process; it has classified the system as high-risk. The boundary of Annex III point 5(c), which names risk assessment and pricing in health and life insurance, is debatable; classification is the subject of Article 6, not of this post.
The data flow: the invoice arrives in the claims system and goes to the policy gateway, which pseudonymises, so insured-person ID instead of name, date of birth and address, checks data zone and scope and calls the vendor model. The model returns class, confidence and the features that carried the result; the gateway writes the evidence record, maps the ID back and delivers to the claims system. “Reimburse automatically” runs with sampling, “suspicious” always lands with a human.
Where the obligation bites: the insurer did not train the model and does not see the training data. It still has to know whether its invoices match what the model was built for (paragraph 4), whether it drifts (paragraph 5) and what it logs (paragraph 6). And it has to tell case handlers and insured persons that the system is there (paragraphs 7 and 11). The vendor’s declaration of conformity answers none of these questions.
How we implement it
At SIFAMO we treat Article 26 not as a documentation duty but as a set of controls in the governance layer. The anchor is the registry record, and it belongs to the insurer, not to the vendor.
Instructions for use as a registry record (paragraph 1). The vendor’s instructions are not filed, they are imported: intended purpose, permitted inputs, known limitations, accuracy figures, indicated oversight measures and logging functions become fields in the record. From those fields we derive the scope the gateway checks on every call. An invoice from a tariff the instructions do not name gets no route. The instructions become configuration, not a PDF on a shared drive. The business owner is the head of claims, and she signs the record.
Oversight with names (paragraph 2). The oversight field names roles, not departments: who reviews suspicious cases, who draws samples, who may stop the system. Each role carries a training record with the date and the version of the instructions. Authority means the reviewer overrides the model’s class without asking anyone and sees the invoice before the model’s score. Support means review time is budgeted, and seeded error cases measure whether reviewers catch them. Paragraph 3 leaves the organisation to us, so we tier by decision type.
Input checks at the gateway (paragraph 4). The insurer controls the input data, so paragraph 4 applies. The gateway checks every invoice against the input schema from the instructions: invoice type, country, currency, language, tariff type, completeness. Whatever does not fit goes to a case handler without a model. Representativeness only shows in the population, so monitoring compares, weekly, the distribution of incoming invoices with the evaluation set the business owns and with the population the vendor describes in the instructions. If it drifts, say after a new tariff launches, the business owner gets an alert before the model quietly degrades.
Monitoring and reporting (paragraph 5). Monitoring reads the evidence log and the gateway. The registry record holds the vendor’s incident contact and the format in which it expects feedback for its post-market monitoring under Art. 72. Two thresholds are defined. “Risk under Art. 79(1)”: the business owner sets the stop flag, informs the vendor and the market surveillance authority, and invoices continue manually. “Serious incident” under Art. 3(49): the vendor first, then the authority, and if the vendor cannot be reached we report ourselves under Art. 73. As a financial institution, the insurer’s monitoring obligation is deemed fulfilled through the governance rules of Solvency II. We build the same thing anyway, because those rules demand monitoring and the reporting duties remain.
Logs for at least six months (paragraph 6). The logs the vendor model generates itself under Article 12 we have delivered to us by contract, otherwise they are not under our control. The gateway’s evidence log is ours: per invoice the input hash, model and instructions version, policy version, output, signals, reviewer action. Six months is a floor. We set the retention period per registry record, here the lifetime of the claim plus the objection period, and justify it there. Because the logs contain only IDs, the period does not collide with GDPR storage limitation.
Informing workers (paragraph 7). The case handlers are subject to the system: it sorts their work, and the log records who overrode which class. Informing them before go-live is a gate: without the date and the document of the information given to the works council and the team there is no registration. Because of co-determination we store reviewer actions under a role ID and restrict per-person analysis through access rights.
Public bodies (paragraph 8) and DPIA (paragraph 9). For a private insurer paragraph 8 does not apply; whether a statutory health insurer counts as a public authority is a question I would put to legal. Technically it is one field: the record of a public body requires the identifier from the EU database under Art. 71, and without it there is no route. The data protection impact assessment refers to the registry record and the imported instructions, describes pseudonymisation, data zones and oversight tiers as its measures, has a version and is reopened on a model change.
Informing and explaining to affected persons (paragraph 11 and Art. 86). The acknowledgement letter to the insured person states that an AI system pre-sorts the invoice and that humans decide. For Art. 86 the insurer has to be able to explain the role of the system and the main elements of the decision. We generate that from the evidence record: which class the model proposed, which rule or policy applied, whether and how a human decided. We do not hand out the model’s own reasoning text as the explanation, because we cannot verify that it is true.
The auditor sees the registry record with imported instructions, roles, training status, retention period and stop authorisations, per invoice the evidence record, the drift reports, every report with a timestamp and the information given to workers. Cooperation under paragraph 12 is then an export, not a project. I am an engineer, not a lawyer. But a deployer who can show this holds Article 26 in its own hands, not its vendor’s.
Typical mistakes
EU AI Act Article 26 checklist
Open questions
Frequently asked questions
Does this apply to us as a deployer if the vendor has declared conformity?
Yes. The declaration of conformity covers the provider’s obligations; Article 26 addresses you as deployer regardless: paragraphs 1 to 6, 9, 11 and 12 for every deployer of a high-risk system, paragraph 7 for employers, paragraph 8 for public bodies, paragraph 10 for law enforcement. More on this under AI governance.
Are six months enough for the logs?
Six months is the floor in paragraph 6, not the target. The period has to be appropriate to the purpose, and for a benefits decision an insured person can challenge years later, it is longer. Justify it in the registry record and agree it with data protection, to which paragraph 6 gives precedence.
Do we have to tell our insured persons that an AI sorts their invoice?
Yes, if the system is an Annex III system and makes decisions about persons or assists in making them. Paragraph 11 requires the notice, and Art. 86 additionally gives the person the right to an explanation of the role of the system and the main elements of the decision. The other articles of the series: all articles at a glance.
Related: EU AI Act · All articles · AI Governance · AI Consultant
Recent Posts
Recent Comments