EU AI Act Article 13: Transparency for Deployers, Implemented in Practice
EU AI Act Article 13 requires a high-risk system to be understandable for the deployer and to ship with instructions for use that state its limits honestly. At SIFAMO we implement those instructions as a machine-readable model card that lives in the registry, ships with every version and is imported directly by the deployer’s registry.
Article 13 connects the two ends of the value chain. The provider knows what its system can do and where it fails. The deployer has to work with it, assign people to oversee it, keep the logs and, in the public sector, write a fundamental rights impact assessment. All of that depends on knowledge only the provider has. Article 13 therefore formally addresses the provider, but the beneficiary is the deployer. Any public authority, bank or insurer buying a high-risk system should know this article as well as the manufacturer does. I describe the wider framework in EU AI Act in practice.
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 amending Regulation (EU) 2026/1744 (Digital Omnibus), from 2 December 2027, and for Annex I products from 2 August 2028. Article 13 belongs to the requirements for high-risk systems in Chapter III Section 2 and follows those dates.
What Article 13 requires
Three paragraphs, one logic: the system has to be understandable by design, it comes with instructions, and the instructions have a minimum content.
Art. 13(1): High-risk systems must be designed and developed so that their operation is sufficiently transparent for deployers to interpret the output and use it appropriately. The yardstick is the set of provider and deployer obligations in Section 3 (Art. 16 to 27). This is a design obligation, not a documentation obligation.
Art. 13(2): The system is accompanied by instructions for use, digital or otherwise. They must be concise, complete, correct and clear, and relevant, accessible and comprehensible to the deployer.
Art. 13(3)(a): Identity and contact details of the provider and, where applicable, of its authorised representative.
Art. 13(3)(b): Characteristics, capabilities and limitations of performance: the intended purpose (i); accuracy with its metrics, robustness and cybersecurity under Art. 15 against which the system was tested and which can be expected, with the circumstances that affect that level (ii); known or foreseeable circumstances under intended use or reasonably foreseeable misuse that may lead to risks under Art. 9(2) (iii); capabilities to explain the output (iv); performance for specific groups of persons (v); specifications for the input data and information on training, validation and testing data (vi); information for interpreting the output (vii).
Art. 13(3)(c): Changes to the system and its performance that the provider pre-determined at the initial conformity assessment.
Art. 13(3)(d): The human oversight measures under Art. 14, including technical measures that help the deployer interpret the output.
Art. 13(3)(e): The computational and hardware resources needed, the expected lifetime, maintenance and care including the frequency of software updates.
Art. 13(3)(f): Where relevant, the mechanisms that allow the deployer to collect, store and interpret the logs under Art. 12.
The counterparts are in Art. 26: deployers use the system in accordance with the instructions (paragraph 1), monitor its operation on their basis (paragraph 5) and need them for the fundamental rights impact assessment under Art. 27. Art. 11 with Annex IV requires the same content in more depth in the technical documentation. I therefore read Article 13 as the part of the technical documentation that has to leave the manufacturer’s building.
The technical example
A software vendor delivers a document-classification model to several public authorities. It sorts the incoming mail of a benefits administration into seven classes, among them application, supporting document, objection and other, and proposes the responsible unit. Other lands in a holding tray. A recognised objection starts deadlines running.
The data flow is short. A scan arrives in the digital mailroom, OCR extracts the text, the text and a few metadata fields go to the model in pseudonymised form. Back come a class, a confidence score, the three most likely alternatives and the text passages that support the class. The case application routes automatically above a threshold and to a clerk below it.
Whether this is high-risk is for the provider to decide and justify. Annex III point 5(a) covers systems used by public authorities to evaluate eligibility for public assistance benefits. Pure sorting is a preparatory task, for which Art. 6(3) provides an exemption. We nevertheless treat the system as high-risk, because the class decides whether an application is handled as an application and whether a deadline runs. That affects people, even without an eligibility assessment.
Article 13 bites in four places. The authorities have different inboxes, a lot of handwriting here, clean portal uploads there; the deployer has to know where the model is weak. It has to set the routing threshold, which only works with metrics per evaluation set. It has to organise oversight: who sees the uncertain cases, with what support, what happens to the correction. And it has to know how long the version lives and what to do when a form is changed and the input distribution shifts.
How we implement it
Our answer to Article 13 is the model card as a registry record. Not a PDF in the contract annex, but a structured, versioned document that belongs to the model version the way the hash belongs to the artefact. The provider maintains it in its registry, signs it at release on the business side and the technical side, and it ships with every version. Simplified, with example values:
{
"model_id": "DOC-CLS-2026-07", "version": "3.2.0", "card_hash": "sha256:9f3a...",
"provider": {"name": "Vendor Ltd", "contact": "ai-compliance@vendor.example"},
"intended_purpose": "Classification of incoming documents in benefits administration into seven classes plus a routing proposal. No eligibility assessment.",
"out_of_scope": ["Eligibility assessment", "Fraud detection", "Mail outside benefits administration"],
"known_limitations": ["Handwriting: markedly lower accuracy", "Several documents in one scan count as one", "Evaluated for German only"],
"metrics": [
{"eval_set": "provider-reference-2026Q2", "n": 4800, "macro_f1": 0.94, "recall_objection": 0.97},
{"eval_set": "handwriting-subset-2026Q2", "n": 350, "macro_f1": 0.71},
{"eval_set": "deployer-owned", "n": null, "macro_f1": null, "note": "to be filled by the deployer before release"}
],
"input_requirements": {"formats": ["PDF", "TIFF", "JPEG"], "min_dpi": 200, "ocr_text": "required", "languages": ["de"]},
"output": {"class": "one of seven", "confidence": "0 to 1, calibrated on provider-reference-2026Q2", "alternatives": "top 3", "evidence_spans": "supporting text passages"},
"predetermined_changes": {"routing_threshold": [0.60, 0.95], "class_set": "fixed", "fine_tuning": "not foreseen"},
"human_oversight": {"mandatory_review": ["confidence below threshold", "class objection"], "reviewer_ui": "show alternatives and text passages", "override": "reviewer changes class, signal goes to provider"},
"lifetime": {"released": "2026-03-02", "expected_lifetime_months": 18, "re_evaluation_due": "2026-12-01", "update_cadence": "quarterly"},
"logging": {"schema": "evidence-log-v2", "fields": ["input_hash", "model_version", "card_hash", "class", "confidence", "reviewer_action"]}
}
Every field has an address in the law: provider is point (a), intended purpose through output cover point (b) from (i) to (vii), predetermined changes is point (c), human oversight point (d), lifetime point (e), logging point (f). We built the schema along the paragraph so that an auditor finds the mapping without us.
Three decisions in there are our interpretation. Art. 13(2) leaves the format open; we chose machine-readable and render the human-readable version from it, because instructions no gateway can read are instructions nobody follows. Concise and complete at the same time only works in layers; the card is the concise layer and links to the technical documentation under Annex IV. And we do not accept metrics without an evaluation set; every figure carries the name and size of its set, and the card has an empty slot for the deployer’s own set.
The real leverage is on the deployer side. The authority’s registry has an import for model cards. It verifies the signature against the provider’s key, creates the record for this model version and marks the fields the authority has to fill itself: business owner, result on its own evaluation set built from real mail, chosen threshold, oversight tiers, data zone and expiry date. The expiry date is the earlier of two dates, the provider’s re-evaluation date from the card and the authority’s own deadline from its risk assessment. As long as a mandatory field is empty or the sign-off by business, risk and IT is missing, there is no route in the gateway. The principle from our governance layer applies here too: no registration, no route.
Then the instructions become configuration. The gateway enforces the input requirements: a fax with too low a resolution is rejected and logged as out of spec instead of being silently misclassified. The threshold has to lie within the predetermined range, otherwise the registry rejects the record. The mandatory reviews become the escalation rule: objections and uncertain cases go to the review queue, the reviewer sees alternatives and text passages, her correction lands in the evidence log and goes back to the provider as a signal. Per decision, the evidence log stores the hash of the card that was in force. Months later an auditor sees not only what the model did, but under which instructions.
And if the authority modifies the system? Then Art. 25 applies. Under Art. 25(1) a deployer becomes a provider itself if it puts its name on the system (point (a)), substantially modifies it (point (b)) or changes the intended purpose so that it becomes high-risk (point (c)). The original provider then has to cooperate and supply information under Art. 25(2). In the example: moving the threshold within the range is predetermined, the authority stays a deployer. Fine-tuning on its own documents or adding a class for suspected fraud is something else. Our registry treats every change outside predetermined changes as a fork: the record is branched, the role switches to provider, the fields from Art. 16 and Annex IV become mandatory, and the route exists only after the full gate. Nobody should become a provider by accident.
Typical mistakes
The instructions are a PDF in the contract annex. Version 1.0 of the instructions, version 3 of the model, and nobody notices because the two are not linked anywhere.
Metrics without a test set. “Accuracy above 95 percent” says nothing if it does not say on which documents.
Limitations are missing because sales does not like them. The weakness on handwriting is written down nowhere until the first objection sits in the holding tray and the deadline has passed.
The deployer does not import the instructions. Threshold, input requirements and review rules are in the document and nowhere in the system. The instructions are then fulfilled and ineffective.
Nobody defined what the deployer may change. The authority’s IT fine-tunes, and nobody notices the role change under Art. 25.
EU AI Act Article 13 checklist
The instructions for use are versioned and bound by hash to exactly one model version.
They contain every item of Art. 13(3)(a) to (f), and each item has a field.
Every metric names its evaluation set and size; there are metrics for subgroups and for weak inputs.
Known limitations and reasonably foreseeable misuse are named concretely, not as boilerplate.
Predetermined changes are defined as a range; anything outside triggers the role change under Art. 25.
The oversight measures under Art. 14 are described as rules a deployer can carry over into its system.
The deployer’s registry imports the instructions, enforces the mandatory fields and stores per decision which version was in force.
Open questions
Format. As long as no harmonised standard under Art. 40 casts the minimum content into a schema, every provider defines its own. A deployer with ten providers has ten imports.
Calibration. A confidence score calibrated on the provider’s data can be systematically off on the mail of one particular authority. How much recalibration a deployer may do before it slides into Art. 25 is open.
The edge of the exemption. Where does the preparatory task under Art. 6(3) end? For document sorting with deadline effects we choose the stricter reading. Where the neighbouring articles pick up is shown in all articles at a glance.
Frequently asked questions
Does Article 13 also apply to us as a deployer?
Formally it addresses the provider. But Art. 26(1) obliges you to use the system in accordance with the instructions for use, and Art. 26(5) to monitor its operation on their basis. Without usable instructions you cannot demonstrate your own compliance. If you do not get them, do not go to production. If you modify the system, you can become a provider yourself under Art. 25.
Is a model card in the style of the large model vendors enough?
Not on its own. It is a good start for point (b). Art. 13(3) additionally requires oversight measures, lifetime and maintenance, predetermined changes and the mechanisms for the logs. And the metrics have to belong to the shipped version, not to the model family. That is the difference between marketing and AI governance.
What happens on a model update?
New version, new card, new import. The deployer’s registry accepts the version only once its own evaluation has run on the new state and the sign-off has been renewed. Predetermined updates go through a lighter gate with a regression test, everything else through the full one. The old version stays registered as the rollback until the new one has passed.
EU AI Act Article 13: Transparency for Deployers, Implemented in Practice
EU AI Act Article 13 requires a high-risk system to be understandable for the deployer and to ship with instructions for use that state its limits honestly. At SIFAMO we implement those instructions as a machine-readable model card that lives in the registry, ships with every version and is imported directly by the deployer’s registry.
Article 13 connects the two ends of the value chain. The provider knows what its system can do and where it fails. The deployer has to work with it, assign people to oversee it, keep the logs and, in the public sector, write a fundamental rights impact assessment. All of that depends on knowledge only the provider has. Article 13 therefore formally addresses the provider, but the beneficiary is the deployer. Any public authority, bank or insurer buying a high-risk system should know this article as well as the manufacturer does. I describe the wider framework in EU AI Act in practice.
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 amending Regulation (EU) 2026/1744 (Digital Omnibus), from 2 December 2027, and for Annex I products from 2 August 2028. Article 13 belongs to the requirements for high-risk systems in Chapter III Section 2 and follows those dates.
What Article 13 requires
Three paragraphs, one logic: the system has to be understandable by design, it comes with instructions, and the instructions have a minimum content.
The counterparts are in Art. 26: deployers use the system in accordance with the instructions (paragraph 1), monitor its operation on their basis (paragraph 5) and need them for the fundamental rights impact assessment under Art. 27. Art. 11 with Annex IV requires the same content in more depth in the technical documentation. I therefore read Article 13 as the part of the technical documentation that has to leave the manufacturer’s building.
The technical example
A software vendor delivers a document-classification model to several public authorities. It sorts the incoming mail of a benefits administration into seven classes, among them application, supporting document, objection and other, and proposes the responsible unit. Other lands in a holding tray. A recognised objection starts deadlines running.
The data flow is short. A scan arrives in the digital mailroom, OCR extracts the text, the text and a few metadata fields go to the model in pseudonymised form. Back come a class, a confidence score, the three most likely alternatives and the text passages that support the class. The case application routes automatically above a threshold and to a clerk below it.
Whether this is high-risk is for the provider to decide and justify. Annex III point 5(a) covers systems used by public authorities to evaluate eligibility for public assistance benefits. Pure sorting is a preparatory task, for which Art. 6(3) provides an exemption. We nevertheless treat the system as high-risk, because the class decides whether an application is handled as an application and whether a deadline runs. That affects people, even without an eligibility assessment.
Article 13 bites in four places. The authorities have different inboxes, a lot of handwriting here, clean portal uploads there; the deployer has to know where the model is weak. It has to set the routing threshold, which only works with metrics per evaluation set. It has to organise oversight: who sees the uncertain cases, with what support, what happens to the correction. And it has to know how long the version lives and what to do when a form is changed and the input distribution shifts.
How we implement it
Our answer to Article 13 is the model card as a registry record. Not a PDF in the contract annex, but a structured, versioned document that belongs to the model version the way the hash belongs to the artefact. The provider maintains it in its registry, signs it at release on the business side and the technical side, and it ships with every version. Simplified, with example values:
Every field has an address in the law: provider is point (a), intended purpose through output cover point (b) from (i) to (vii), predetermined changes is point (c), human oversight point (d), lifetime point (e), logging point (f). We built the schema along the paragraph so that an auditor finds the mapping without us.
Three decisions in there are our interpretation. Art. 13(2) leaves the format open; we chose machine-readable and render the human-readable version from it, because instructions no gateway can read are instructions nobody follows. Concise and complete at the same time only works in layers; the card is the concise layer and links to the technical documentation under Annex IV. And we do not accept metrics without an evaluation set; every figure carries the name and size of its set, and the card has an empty slot for the deployer’s own set.
The real leverage is on the deployer side. The authority’s registry has an import for model cards. It verifies the signature against the provider’s key, creates the record for this model version and marks the fields the authority has to fill itself: business owner, result on its own evaluation set built from real mail, chosen threshold, oversight tiers, data zone and expiry date. The expiry date is the earlier of two dates, the provider’s re-evaluation date from the card and the authority’s own deadline from its risk assessment. As long as a mandatory field is empty or the sign-off by business, risk and IT is missing, there is no route in the gateway. The principle from our governance layer applies here too: no registration, no route.
Then the instructions become configuration. The gateway enforces the input requirements: a fax with too low a resolution is rejected and logged as out of spec instead of being silently misclassified. The threshold has to lie within the predetermined range, otherwise the registry rejects the record. The mandatory reviews become the escalation rule: objections and uncertain cases go to the review queue, the reviewer sees alternatives and text passages, her correction lands in the evidence log and goes back to the provider as a signal. Per decision, the evidence log stores the hash of the card that was in force. Months later an auditor sees not only what the model did, but under which instructions.
And if the authority modifies the system? Then Art. 25 applies. Under Art. 25(1) a deployer becomes a provider itself if it puts its name on the system (point (a)), substantially modifies it (point (b)) or changes the intended purpose so that it becomes high-risk (point (c)). The original provider then has to cooperate and supply information under Art. 25(2). In the example: moving the threshold within the range is predetermined, the authority stays a deployer. Fine-tuning on its own documents or adding a class for suspected fraud is something else. Our registry treats every change outside predetermined changes as a fork: the record is branched, the role switches to provider, the fields from Art. 16 and Annex IV become mandatory, and the route exists only after the full gate. Nobody should become a provider by accident.
Typical mistakes
EU AI Act Article 13 checklist
Open questions
Frequently asked questions
Does Article 13 also apply to us as a deployer?
Formally it addresses the provider. But Art. 26(1) obliges you to use the system in accordance with the instructions for use, and Art. 26(5) to monitor its operation on their basis. Without usable instructions you cannot demonstrate your own compliance. If you do not get them, do not go to production. If you modify the system, you can become a provider yourself under Art. 25.
Is a model card in the style of the large model vendors enough?
Not on its own. It is a good start for point (b). Art. 13(3) additionally requires oversight measures, lifetime and maintenance, predetermined changes and the mechanisms for the logs. And the metrics have to belong to the shipped version, not to the model family. That is the difference between marketing and AI governance.
What happens on a model update?
New version, new card, new import. The deployer’s registry accepts the version only once its own evaluation has run on the new state and the sign-off has been renewed. Predetermined updates go through a lighter gate with a regression test, everything else through the full one. The old version stays registered as the rollback until the new one has passed.
Related: EU AI Act · All articles · AI Governance · AI Consultant
Recent Posts
Recent Comments