Investieren in Dubai als Deutscher: warum wir über die Holding gehen | Umair Zaffar
Umair Zaffar November 4, 2026 0 Comments

EU AI Act Article 25: The AI Value Chain, Implemented in Practice

EU AI Act Article 25 turns a bank that repurposes a purchased model into the provider, with every obligation in Article 16. We make the registry create a new record version for every change to weights, prompt policy, scope or intended purpose, and re-check the role on every version.

Most high-risk systems in banks and insurers are not built. They are bought and adapted. That is exactly where Article 25 of the AI Regulation starts. It sets out when a distributor, importer, deployer or other third party becomes a provider in its own right, what the original provider still owes after that, and what suppliers of components have to commit to in writing. To me it is the article that most often changes roles without anyone noticing. Nobody decides to become a provider. You fine-tune a model, change what it is used for, give it an internal name, and you are one.

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 25 is one of those high-risk obligations (Art. 113(c)). The Omnibus moved deadlines; it removed no obligation. The overview is in EU AI Act in practice.

What Article 25 requires

  • Art. 25(1)(a): Whoever puts its name or trademark on a high-risk system that has already been placed on the market or put into service is considered a provider and carries the obligations of Article 16. Contractual arrangements that allocate the obligations differently remain possible.
  • Art. 25(1)(b): Whoever makes a substantial modification to a high-risk system (Art. 3(23): not foreseen in the conformity assessment, affects compliance or the intended purpose) becomes a provider if the system remains high-risk.
  • Art. 25(1)(c): Whoever modifies the intended purpose of an AI system, explicitly including a general-purpose AI system, so that a system not previously classified as high-risk becomes high-risk under Article 6, becomes a provider. There is no contractual carve-out here.
  • Art. 25(2): Where one of these cases occurs, the initial provider is no longer the provider of that specific system. It must nevertheless cooperate closely with the new provider, make information available and give the reasonably expected technical access and other assistance, in particular for the conformity assessment. The consolidated text spells this out: technical documentation for assessing compliance with Article 16, known limitations and failure modes, targeted test access. The duty does not apply where the initial provider has clearly specified that its system is not to be changed into a high-risk system.
  • Art. 25(3): Where a high-risk system is a safety component of a product under Annex I Section A, the product manufacturer is considered the provider if the system is placed on the market with the product under its name or trademark, or put into service under its name or trademark afterwards. Not relevant for the bank in the example.
  • Art. 25(4): The provider of a high-risk system and any third party that supplies AI systems, tools, services, components or processes for it specify by written agreement which information, capabilities, technical access and other assistance, based on the state of the art, are needed so that the provider can fully comply. Exempt are third parties that make tools, services, processes or components publicly accessible under a free and open-source licence, with the exception of general-purpose AI models. The AI Office may develop voluntary model contractual terms, with a view to sector-specific requirements, and publishes them free of charge.
  • Art. 25(5): Paragraphs 2 and 3 are without prejudice to the protection of intellectual property, confidential business information and trade secrets under Union and national law.

The technical example

A bank buys a model for inbound-mail triage from a vendor. It classifies incoming documents, extracts income, employer and liabilities from payslips and bank statements, and routes the case to the responsible department. Intended purpose according to the vendor’s documentation: document classification and data extraction. No Annex III case, no high-risk system. The bank is a deployer under Article 26.

Then what happens in every second project happens. The credit team notices that the model reads the relevant fields well. It fine-tunes the model on its own applications with known outcomes, replaces the prompt policy with one that ends in a recommendation (fast track, manual review, decline recommended), names the system “Credit Pilot” in the advisor frontend and rolls it out to the branches. The vendor’s name disappears from the interface. Technically these are three parts: the vendor’s hosted base service, a fine-tuning artefact with the bank’s own weights, and a prompt policy that translates the output into a decision.

Now the three letters in the right order. Points (a) and (b) do not bite first, although the fine-tuning is a substantial modification, because both presuppose that the system was already high-risk. The decisive one is point (c): document triage became an evaluation of the creditworthiness of natural persons, Annex III point 5(b). Art. 6(3) does not help, because the system profiles people. The moment the new prompt policy goes live, the bank is the provider of a high-risk system and carries Article 16: the requirements of Art. 9 to 15, from risk management through data governance for its own training data to human oversight, plus quality management (Art. 17), conformity assessment (Art. 43), the declaration of conformity, CE marking and registration (Art. 49).

The vendor becomes two things at once: the former provider under paragraph 2, who has to hand over documentation, failure modes and test access, and the supplier of a component under paragraph 4, with whom the bank needs a written agreement. Both fail if its terms of use exclude high-risk use; then the cooperation duty under paragraph 2 falls away. Add the OCR service, the embedding model, the retrieval index and the hosting, each a paragraph 4 case. The open-source OCR library is exempt. An openly published general-purpose AI model as the base explicitly is not: its provider stays within paragraph 4 and in addition has its own information duties towards downstream providers under Art. 53(1)(b), unless Art. 53(2) exempts it. If there is nobody who can sign, the gap stays with the bank.

How we implement it

The core: the role is a field in the registry that is not set once at procurement but re-checked on every version. In the governance layer every model has a registry record, and without a valid record there is no route through the policy gateway. For Article 25 we added four things to the record.

  • The vendor’s intended purpose as a source, not a paraphrase. The record carries the wording from the vendor’s documentation with the document version, the use clause that says whether a change into a high-risk system is excluded, and next to that our own intended purpose. If the two diverge, you see it before anyone asks.
  • Four role-relevant fields, every change a new version. Model weights (as a hash of the artefact), prompt policy version, scope (data zones, permitted actions, tools) and intended purpose. If one of them changes, the registry automatically creates a new record version, and the gateway keeps routing to the old, approved one. “No registration, no route” here means: no role review, no route for the new version.
  • The role review is one page with three questions, one per point of paragraph 1: Does the system appear under our name towards customers or employees? Does the change go beyond what the vendor’s conformity assessment foresaw? Does the new intended purpose fall under Annex III or Annex I? Business owner, risk and legal sign. Result “provider”: the risk class jumps to high-risk, the record gets the list of Article 16 obligations with one owner each, and the high-risk gates stand before the route: the business unit’s evaluation set, regression, adversarial and bias tests, a rehearsed rollback, three-way sign-off, expiry.
  • The written agreement is an attachment to the record, per component. Each component, from base model to hosting, carries one of four states. “Agreement under paragraph 4 signed” with a link, “open source, exempt” with the licence, “GPAI model, Art. 53 documentation on file”, or “missing”. “Missing” on a high-risk record is a gate failure, and we write down how we close the gap ourselves, usually with our own tests and a narrower scope.

The evidence log stores the record version and the policy version per decision. An auditor therefore sees from which version onwards the bank was the provider and who signed that. The kill switch stays with the business owner, also after the role change.

Our contract template for paragraph 4 always asks for the same things before signature: the supplier’s intended purpose verbatim and whether high-risk use is excluded; technical documentation to the extent of Annex IV; known limitations and failure modes, in writing; training data and bias tests; version pinning and advance notice of every model change; test access, with confidence values where possible; status as a general-purpose AI model and the Art. 53 documentation; cooperation in conformity assessment, audits and incident reporting under Art. 73; and a list of what the supplier withholds with reference to paragraph 5. Without that list we do not sign.

Where the law leaves open whether a purely internal product name already counts as “own name” under point (a), we treat every renaming that customers or employees see as a trigger for review. That costs an hour. The discussion with the supervisor about whether “Credit Pilot” was a brand costs more.

Typical mistakes

  • Fine-tuning and prompt rebuilds pass as “configuration”, without a new version. The role is assessed once at procurement and the next time in the audit. Article 25 describes events, not states.
  • Terms of use with the clause “not intended for high-risk purposes” are accepted unread. The cooperation duty under paragraph 2 is gone before the project starts.
  • There is an agreement with the provider of the main model, but none for OCR, embedding, retrieval and hosting. Paragraph 4 does not know that distinction.
  • “Open source, so no obligations.” The exemption relieves the supplier, not the bank, and it does not apply to general-purpose AI models.

EU AI Act Article 25 checklist

  • Intended purpose and use clause of every vendor in the registry, verbatim and with document version.
  • Model weights, prompt policy, scope and intended purpose versioned; every change creates a new record version without a route.
  • Role review under Art. 25(1)(a) to (c) per version, three signatures, result in the record.
  • On a role change: risk class high-risk, Article 16 obligation list with an owner per obligation, provider gates before the route.
  • Cooperation of the initial provider under paragraph 2 secured by contract: documentation, failure modes, test access.
  • Written agreement under paragraph 4 for every component; open-source and GPAI status documented; “missing” as a gate failure.
  • List of information withheld under paragraph 5 with countermeasures; record version in the evidence log.

Open questions

  • Whether an internal product name is “own name or trademark” under point (a), although the system never reached the market under that name, is not settled.
  • Where the line to a substantial modification lies for a fine-tuning when the vendor’s conformity assessment foresaw continued learning under Art. 43(4) is not yet specified by the standards.
  • For openly published general-purpose AI models without a contracting party, the duty in paragraph 4 is in the text but nobody can fulfil it. How the downstream provider closes that gap is a research question, not only a legal one.

Frequently asked questions

Does this apply to us as a deployer if we only adjust the prompt?

As long as the adjustment stays within the vendor’s intended purpose, you remain a deployer and Article 26 is your yardstick. If the prompt turns a triage model into a decision aid for creditworthiness, that is point (c), whether or not a single weight changed. Version the prompt in both cases, otherwise you cannot prove the difference later. More on this under AI Governance.

What happens to the vendor’s declaration of conformity?

It covers the vendor’s system with the vendor’s intended purpose, not yours after the change. The conformity assessment under Art. 43 is then your task; the vendor has to cooperate under paragraph 2, unless it clearly excluded high-risk use.

Can we leave the provider role with the vendor by contract?

Only point (a) contains the reservation for contractual arrangements. Points (b) and (c) have none: whoever makes the change is the provider. A contract can allocate work and liability, not the role towards the supervisor. The other articles in the series: all articles at a glance.

Related: EU AI Act · All articles · AI Governance · AI Consultant

Leave Comment