ISHIGHRISK AI
Roles

Provider, deployer, and the Article 25 flip

Three acts turn a deployer into a provider under Article 25 and hand you the full Article 16 obligation set. How role determination actually works.

Reviewed 22 July 2026Regulation (EU) 2024/1689, as amended by the Digital Omnibus
In short

Six roles exist under Regulation (EU) 2024/1689 and a single entity routinely holds several at once. Article 25 turns a deployer, importer or distributor into a provider on three triggers: putting your name or trademark on a high-risk system, making a substantial modification under Article 3(23), or repurposing any system, including a general-purpose one, into a high-risk use. The flip hands you the full Article 16 obligation set rather than a subset, and it is enforceable from 2 December 2027 for standalone Annex III systems and 2 August 2028 for AI embedded in Annex I products.

Your role is not a label you select. It is a legal consequence of what you do with a system, and it can change without anyone sending you a notice. A company that buys a high-risk AI system, puts its own logo on the interface and rolls it out to staff has not made a branding decision. It has made itself the provider of that system, with everything Article 16 attaches to the word.

The six AI Act roles, and why you hold more than one

Regulation (EU) 2024/1689 distributes obligations across six operator roles. Each is defined by conduct, so the determination is made per system and not per company.

  • Provider - develops an AI system, or has one developed, and places it on the market or puts it into service under its own name or trademark. This role carries almost every substantive obligation in the high-risk regime.
  • Deployer - uses an AI system under its own authority in a professional capacity. The duty set is real but materially lighter, and it is built around using the system the way the provider assessed it.
  • Importer - established in the Union and places on the Union market a system carrying the name or trademark of a person established outside the Union.
  • Distributor - anyone else in the supply chain who makes a system available on the Union market, other than the provider or the importer.
  • Product manufacturer - places a product on the market together with a high-risk AI safety component under its own name or trademark, and by doing so takes the provider seat. Art 25(3)
  • Authorised representative - appointed in writing by a provider established outside the Union to act for it within the Union.

These roles stack. Picture a mid-sized insurer. It builds an internal pricing model for health cover, which is an Annex III point 5 use case, so it is the provider and the deployer of that model at once. It buys a CV screening tool from a US vendor for its own hiring, so it is the deployer of that one, and it would additionally be the importer only if it made the vendor-branded system available to others on the Union market rather than using it internally. It resells an unmodified, vendor-branded claims-triage tool to a subsidiary, so it is the distributor of a third. Three systems, one legal entity, four separate role assignments, and every duty attaches to the specific system it arose from. If you want the per-system answer for your own estate, the free AI Act triage classifier runs role determination as its final stage.

Provider vs deployer: what each role actually owes

The reason role determination matters is the size of the gap. Nearly the whole compliance burden of the high-risk regime sits on one side of it.

DutyProviderDeployer
Core high-risk requirementsArticles 9 to 15 in full: risk management, data governance, technical documentation, logging, instructions for use, human oversight, accuracy, robustness, cybersecurityNone directly. You operate inside the design and the instructions for use
Technical documentationAnnex IV, nine blocks, drawn up before market placement, kept current, retained around ten yearsRecords of your own use only
Conformity assessmentArticle 43, via Annex VI internal control or the Annex VII notified-body routeNone
Declaration and markingEU declaration of conformity under Article 47, plus CE markingNone
EU databaseRegistration under Article 49, and registration even where you self-assess out of high-risk under Article 6(3)Not the provider registration
Article 6(3) derogationMust document the assessment before market placement, per Article 6(4)Cannot claim it for a system you did not place on the market
Article 50 transparency50(1) chatbot disclosure, and 50(2) synthetic-content marking from 2 December 2026 for systems placed on the market before 2 August 2026, from placement for later ones50(3) emotion and biometric notice, 50(4) deepfake and public-interest text
Penalty exposureThe higher of 15,000,000 EUR or 3% of total worldwide annual turnover. SMEs and start-ups pay the lower of the two under Article 99(6)The higher of 15,000,000 EUR or 3% of total worldwide annual turnover. SMEs and start-ups pay the lower of the two under Article 99(6)

Two rows deserve attention. The penalty tier is the same for both roles, so the flip does not raise your exposure ceiling, it raises the number of ways you can breach and reach it. And Article 50 runs on its own logic: the disclosure duties split provider and deployer differently from the high-risk regime, which is why a company can be a pure deployer with no Article 16 duties at all and still owe the Article 50 transparency disclosures from 2 August 2026.

2 Aug 2026Article 50 transparency applies and AI Office enforcement begins. Deployer duties under 50(3) and 50(4) bite here, whatever your high-risk status.
2 Dec 2026Article 50(2) machine-readable marking applies to generative systems placed on the market before 2 August 2026. Systems placed on the market from 2 August 2026 comply from placement.
2 Dec 2027High-risk obligations apply to standalone Annex III systems. If the flip has caught you by then, the Article 16 set is enforceable against you from this date.
2 Aug 2028High-risk obligations apply to AI embedded in Annex I products. This is the date that governs Article 25(3) product manufacturers.

The Article 25 flip: three acts that make you a provider

Article 25 is the provision that catches organisations who never intended to build anything. A deployer, importer or distributor is treated as a provider of a high-risk system, and inherits the provider obligations, on any one of three triggers.

TriggerWhat it looks likeWhat it takes to fire
Name or trademark on a high-risk system already on the marketYou licence a high-risk tool and present it to staff or customers under your own brandBranding alone. No code change, no retraining, no new feature is required
Substantial modification that leaves the system high-riskYou retrain on your own data, or extend the system past what its assessment coveredA post-market change not foreseen in the initial conformity assessment that affects compliance or changes intended purpose Art 3(23)
Modified intended purpose that makes a system high-riskYou point a general-purpose assistant at CV shortlisting, credit decisions or access to public benefitsApplies to any system, including a general-purpose AI system, and applies even where the system was not high-risk before you touched it

The flip is all or nothing. Inheriting provider status hands you the full Article 16 obligation set, not the parts that seem proportionate to what you changed. Rebranding a system you did not build still obliges you to hold the Articles 9 to 15 requirements, produce Annex IV technical documentation, run conformity assessment under Article 43, issue the EU declaration of conformity and register the system. There is no reduced tier for the accidental provider, and no authority sends you a notice when the trigger fires.

The corollary matters just as much. Once you are the provider of the modified or rebranded system, the vendor conformity assessment describes the vendor system, not yours. You cannot rely on paperwork that was drawn up for a different configuration and a different intended purpose, and you will need the underlying technical information to build your own. Article 25(2) obliges the original provider to cooperate and to make available the information reasonably required to meet the Article 16 obligations, but the practical time to pin down what that means, and in what format and on what timescale, is before signature.

Substantial modification under Article 3(23), and where the line sits

Article 3(23) is the definition that decides trigger two. A substantial modification is a change made to an AI system after it has been placed on the market or put into service, which was not foreseen or planned in the initial conformity assessment, and which affects compliance with the high-risk requirements or changes the intended purpose.

Read it as two limbs, both of which have to be satisfied.

  • Not foreseen in the initial assessment. A provider that anticipated periodic retraining, documented it, and covered it in its conformity assessment has already absorbed that change. Updates inside that envelope do not flip anyone. Changes the provider never contemplated fall outside it.
  • Affects compliance or changes intended purpose. Cosmetic and integration work generally does not. Changing what the system decides, what data it decides on, or how a human can intervene generally does.

Trigger two only applies where the system stays high-risk after the change. If your modification takes a high-risk system out of scope entirely, you are in a different analysis, and one you should document with the same care as the Article 6(3) derogation assessment. If your modification takes a system that was not high-risk into a high-risk use, you are in trigger three, not trigger two.

Be honest about how thin the guidance is here. The Commission missed its February 2026 statutory deadline for the Article 6 high-risk classification guidelines, so authoritative worked examples are scarce across the whole classification question, and modification sits directly downstream of it. The conservative reading is the defensible one: if you cannot point to a line in the vendor conformity assessment that anticipated your change, assume the change was not foreseen and record why you concluded it did or did not affect compliance.

Article 25(3): when the product manufacturer is the provider

Article 25(3) handles the embedded case. Where a high-risk AI system is a safety component of a product and is placed on the market together with that product under the name or trademark of the product manufacturer, the manufacturer is the provider of the AI system. The component supplier does not carry it for you.

This is the Annex I route into the high-risk regime, where the AI is a safety component of, or is itself, a product covered by Annex I harmonisation legislation requiring third-party conformity assessment. Two practical points follow. The Digital Omnibus gave AI embedded in Machinery Regulation products a targeted carve-out from the AI Act direct high-risk rules, layering the AI-specific requirements in through the Machinery Regulation instead, so a machinery manufacturer reads a different rulebook for the same component. Medical devices and toys remain fully in scope. And the applicable date for embedded Annex I systems is 2 August 2028, not the 2 December 2027 date that governs standalone Annex III systems.

Fine-tuning a GPAI model can make you the provider of a new model

Two different flips live in the same neighbourhood and they are easy to confuse. Article 25(1) operates on AI systems: repurposing a general-purpose AI system into a high-risk use makes you the provider of a high-risk system. The GPAI rules operate on models: fine-tuning or substantially modifying a general-purpose AI model can make you the provider of a new model, which pulls you into the Article 53 baseline.

That baseline is Annex XI technical documentation, Annex XII information for downstream providers, a copyright policy and a public summary of training data. If the resulting model crosses into systemic risk, presumed above 10^25 FLOPs of training compute under Article 51, Article 55 adds model evaluation and adversarial testing, systemic-risk assessment and mitigation, serious-incident reporting to the AI Office, cybersecurity protection and Commission notification within two weeks. The full GPAI obligation set is set out separately, including the Article 3(63) generality test and the indicative 10^23 FLOPs figure that marks a model as general-purpose in the first place.

The one-third-compute threshold is guidance, not law. The July 2025 Commission guidance offers an indicative threshold of about one third of the original training compute for when a fine-tune creates a new model provider. It is an interpretive figure, it is not written into Regulation (EU) 2024/1689, and it is not a safe harbour. Fine-tunes below it can still change a model behaviour enough to matter, and downstream providers frequently cannot measure the original compute at all. Document the reasoning you actually used rather than pointing at a number you cannot verify.

Four ways companies become an accidental provider

The same four mistakes recur. None of them looks like a compliance decision at the moment it is made, which is precisely the problem.

1. Assuming vendor compliance covers you

A CE marking and a declaration of conformity attach to the vendor system as assessed. They are not a transfer of liability and they say nothing about your deployer duties, which you owe in your own right whatever the vendor has done. A procurement team that files the compliance pack and considers the matter closed has documented the vendor position, not its own. Worse, the pack becomes actively misleading the moment any Article 25 trigger fires, because it then describes a system you are no longer running.

2. Rebranding a purchased system

This is the cheapest way to become a provider and the one teams walk into blind. Putting your name or trademark on a high-risk system already on the market is trigger one on its own. It does not require you to change a single line of code. White-labelling, reselling under your own brand, or presenting a licensed tool to customers as your product all reach it. If your brand team is applying a logo to a third-party high-risk system, that is a compliance event and it should route through the same review as a code change.

3. Repurposing a general tool into a high-risk use

Trigger three applies to any system, including a general-purpose AI system, and it does not care that the system was harmless in its prior use. Picture a general-purpose assistant licensed for drafting internal memos. A team quietly starts using it to rank job applicants against a role description. The tool has not changed at all, but the intended purpose now sits in Annex III point 4, and the organisation that made that decision is the provider of a high-risk system. See the recruitment and HR guide for how quickly the employment use case is reached. This mistake is usually invisible to central compliance because nothing was procured and nothing was announced.

4. Procurement contracts that never allocate the triggers

AI vendor contracts commonly carry a general warranty of AI Act compliance without allocating the Article 25 triggers. That warranty is close to worthless if a trigger fires on your side. Article 25(1)(a) does preserve contractual arrangements that allocate the obligations otherwise, so the branding trigger is one you can and should address in the contract. Triggers two and three follow the conduct, and a clause cannot undo a substantial modification or a change of intended purpose that you actually made. A contract that actually handles this does four things: it states who may apply branding and on what terms; it defines which modifications the provider has anticipated in its conformity assessment and which fall outside it; it records the assessed intended purpose in writing so a departure is detectable; and it obliges the provider to hand over the information you would need to stand up as provider if the flip did land, on notice, both ways.

The practical defence against all four is the same. Determine the role per system, write down the determination with its date and the facts it rested on, and re-run it whenever the branding, the model or the use changes. That record is what you would put in front of a market surveillance authority, and it is the difference between a documented role determination and a rebranding decision nobody can account for. The penalty tier that sits behind it is the higher of 15,000,000 EUR or 3% of total worldwide annual turnover. The primary text is worth reading directly at EUR-Lex, alongside the European Commission AI policy pages.

Check your own system

The free classifier walks the same tests in order and tells you which of them your system actually trips, with the article each answer rests on.

Run the triage →

Frequently asked questions

What is the difference between a provider and a deployer under the EU AI Act?

A provider develops an AI system or a general-purpose AI model, or has one developed, and places it on the market or puts it into service under its own name or trademark, whether for payment or free of charge. A deployer uses a system under its own authority in a professional capacity. The gap is large. The provider carries the Articles 9 to 15 requirements, the Annex IV technical documentation, conformity assessment under Article 43, the EU declaration of conformity under Article 47 and registration under Article 49. The deployer carries a much shorter set built around using the system as instructed. Under Article 50 the split partly reverses: 50(1) and 50(2) are provider duties, while 50(3) and 50(4) are deployer duties.

When does a deployer become a provider under Article 25?

On any one of three triggers. First, you put your name or trademark on a high-risk system already placed on the market. Second, you make a substantial modification to a high-risk system and it remains high-risk. Third, you modify the intended purpose of any system, including a general-purpose AI system, so that it becomes high-risk. One trigger is enough, and no authority notifies you when it fires. The consequence is identical in each case: you inherit the full Article 16 obligation set for that system, not a reduced version of it, and the vendor conformity assessment no longer describes what you are actually running.

What counts as a substantial modification under the AI Act?

Article 3(23) defines it as a change made to an AI system after it has been placed on the market or put into service, which was not foreseen or planned in the initial conformity assessment, and which affects compliance with the high-risk requirements or changes the intended purpose. Two limbs have to be read together. The change must fall outside what the original assessment anticipated, and it must touch either compliance or intended purpose. Routine updates the provider planned for and covered do not flip you. Retraining on your own data, extending the system into decisions it was never assessed for, or working around its human-oversight design usually will.

Does my vendor AI Act compliance cover me as a deployer?

No. A vendor CE marking and EU declaration of conformity attach to the vendor system as it was assessed, not to your use of it. You still owe the deployer duties in your own right, and the moment you trip an Article 25 trigger the vendor assessment stops describing the thing you are running. This is one of the most common role mistakes in procurement: a compliance pack is treated as a transfer of liability, which it is not. Article 99 sets the exposure for the high-risk regime and for Article 50 at 15,000,000 EUR or 3% of total worldwide annual turnover, whichever is higher for companies. The penalty provisions have applied since 2 August 2025.

Does fine-tuning a general-purpose AI model make me a provider?

It can. Fine-tuning or substantially modifying a general-purpose AI model can make you the provider of a new model, which brings the Article 53 baseline with it: Annex XI technical documentation, Annex XII information for downstream providers, a copyright policy and a public summary of training data. The Commission guidance of July 2025 offers an indicative threshold at roughly one third of the original training compute. Treat that as guidance rather than black-letter law, because it is an interpretive figure and not a number in the regulation. If your fine-tune approaches it, or if you cannot measure the original compute at all, document your reasoning as though you were the provider.

Can one company be both a provider and a deployer at the same time?

Yes, and most organisations of any size already are. The roles are defined by activity rather than by corporate identity, so one legal entity can be the provider of the system it built, the deployer of three systems it bought, the importer of one from a non-EU vendor and the distributor of another it resells to a subsidiary. Each role attaches to a specific system and each carries its own duties, and they stack rather than cancel each other out. Run the determination per system rather than per company, and record the answer with the date and the facts it rested on.

What should a procurement contract say about the Article 25 flip?

It should name the three triggers and allocate them explicitly. Set out who may apply branding to the system and on what terms, define which modifications the provider has anticipated in its conformity assessment and which fall outside it, and require the provider to state the assessed intended purpose in writing so that a departure from it is detectable. Add a duty on the provider to hand over the information you would need if the flip did land, plus a notification obligation running both ways whenever either side changes the system or the use. A bare clause declaring the customer to be a deployer does not help against triggers two and three, because there the role follows the conduct. For the branding trigger, Article 25(1)(a) operates without prejudice to contractual arrangements that allocate the obligations otherwise, so the allocation has to be drafted rather than asserted.

This page is triage guidance, not legal advice. It reflects Regulation (EU) 2024/1689 as amended by the Digital Omnibus, reviewed 22 July 2026, when the Omnibus was adopted and signed but awaiting Official Journal publication. Final classification for ambiguous cases needs qualified counsel.