Provider, deployer, producer, user — the role you hold decides your obligations under the EU AI Act and what belongs in your ISO 42001 scope. The short answer to the most-asked version of this question: wrapping someone else's model into your product almost always makes it YOUR AI system.
Why it matters: role determination is the first scoping decision of any AIMS (clause 4.3) — get it wrong and you either audit things that aren't yours or, worse, exclude the things that are.
Plain-English explainer · EU AI Act + ISO/IEC 22989 terminology · Last reviewed July 2026
Under the EU AI Act (Article 3), a provider develops an AI system — or has one developed — and places it on the market under its own name. A deployer uses an AI system under its own authority. Integrating a third-party model into your product under your brand makes you the provider of that system. That single sentence resolves most of the “we just wrap an API” confusion. The ISO vocabulary (ISO/IEC 22989) slices the ecosystem more finely — AI providers, AI producers (developers/designers), AI customers and users, AI partners such as data suppliers and integrators, AI subjects (the people affected), and relevant authorities — and ISO 42001 expects your AIMS scope (clause 4.3) to reflect the roles you actually hold, per AI system. Most product companies hold more than one: provider of the features they ship, deployer of the AI tools their staff use internally. The roles carry different obligations, different Annex A control emphases, and different evidence you can — and cannot — inherit from your model vendor.
Key Takeaways
- Provider = develops (or has developed) an AI system and offers it under own name. Deployer = uses an AI system under own authority (EU AI Act, Article 3). ISO/IEC 22989 adds finer-grained roles: producer, customer/user, partner, subject.
- Wrapping a foundation-model API into your product makes you the provider of the resulting AI system — the vendor is a supplier governed by your A.10.3 controls, not a substitute for your own governance.
- Fine-tuning or rebranding shifts provider duties to you. Internal-only AI use makes you a deployer with a lighter (but non-zero) governance surface.
- Roles drive AIMS scope per system (clause 4.3): what you must evidence yourself vs what you inherit from suppliers is a role question before it is a control question.
- Most companies hold two roles at once — provider of what they ship, deployer of what they use — and the AIMS should say so explicitly.
Four Scenarios That Decide Most Cases
Work top to bottom. The first scenario that matches is usually your primary role. Hold multiple? Scope each AI system separately — the standard expects it.
“We wrap the OpenAI / Anthropic API in our product”
Integrating a third-party foundation model into a product you offer under your own name makes the resulting AI system yours — the model vendor is your supplier, not your shield. Your AIMS scopes the wrapper system: prompts, guardrails, data flows, monitoring, and supplier controls (A.10.3) over the model vendor. You inherit some evidence from the vendor (their certifications cover their controls, not your use of them).
“We fine-tune open or hosted models and ship the result”
Fine-tuning and placing the result in your product keeps you a provider, and pulls the data-for-AI controls (A.7: acquisition rights, quality, provenance, preparation) into the audit spotlight, along with verification and validation (A.6.2.4) of the tuned behavior. Under the EU AI Act, substantially modifying a system or rebranding it can likewise shift provider duties to you.
“We self-host open-weight models inside our platform”
Self-hosting adds the resource controls (A.4: compute, tooling, data, human competence) and operational monitoring (A.6.2.6) squarely to your side — there is no vendor SOC 2 to point at for the serving infrastructure. Scope the AIMS to the systems built on the models, not the weights themselves.
“We only use ChatGPT / Copilot internally”
Using AI under your own authority in professional activity makes you a deployer. Full ISO 42001 certification is rarely the first move here. An AI use policy, an approved-tools register, and data-handling rules usually are. Certification becomes worth it when customers start asking how you govern AI use — or when internal AI starts touching customer deliverables.
What Your Model Vendor's Certificate Does and Doesn't Cover
You can lean on the vendor for
— Their model development, training-data governance, and infrastructure controls — evidenced by their ISO 42001 / SOC 2 reports.
— Their model documentation and usage policies (inputs to your A.10.3 supplier assessment).
— Their uptime, security, and incident commitments — verified, not assumed.
You must evidence yourself
— Your use case: the impact assessment for your deployment context (A.5) — the vendor cannot know who your system affects.
— Your prompts, guardrails, retrieval data, and output handling (A.6 lifecycle, A.7 data).
— Your transparency to users and your human oversight (A.8, A.9).
— Your monitoring of the vendor itself — model changes, deprecations, incidents (A.10.3).
This split is the practical payoff of getting roles right: it converts “is the vendor certified?” into a concrete inheritance table your auditor — and your enterprise customers' AI questionnaires — can actually verify.
AI Roles — Common Questions
Provider, deployer, and the scope decisions that follow.
Does using the OpenAI or Anthropic API put us in scope of ISO 42001?
What is the difference between an AI provider and an AI deployer?
Can we be both provider and deployer at once?
Does fine-tuning a model change our role?
What are the ISO/IEC 22989 stakeholder roles?
If we are only a deployer, do we need ISO 42001 at all?
Related reading: the Learn hub, AI risk vs impact assessment, ISO 42001 vs the EU AI Act, the Annex A controls, and the ISO 42001 hub. Terms in the compliance glossary.
Related Reading
ISO 42001 Knowledge Hub
AIMS controls, EU AI Act mapping, risk assessment and guides.
Read moreEU AI Act Mapping
How ISO 42001 controls map onto EU AI Act obligations.
Read moreAI Governance & AIMS
How an AI Management System is structured: policy, accountability, principles.
Read moreISO 42001 Annex A Controls
The 38 AI-specific controls grouped by domain, mapped to clauses.
Read moreAI Risk Assessment
Risk identification, impact analysis and treatment for AI systems.
Read moreSOC 2 for AI Companies
Enterprise AI procurement, model/data security, and ISO 42001 pairing.
Read more