EU AI Act provider vs deployer for ERP AI integrations

Understand when using, commissioning or changing ERP-connected AI can create provider responsibilities, and which facts determine your EU AI Act role.

DigiventiLast reviewed · 7 min read
On this page

Buying access to an AI model does not settle who is responsible for the ERP-connected system built around it. The business might be using a supplier’s finished feature, commissioning its own assistant, or assembling a new application from a model, business data and workflow connections. Those situations need different legal analysis.

The useful starting point is to identify the system being supplied or put into use, its intended purpose, and the organisation under whose name it enters service. Whether the assistant sits inside the ERP interface or connects from outside helps describe the arrangement. It does not, by itself, determine the EU AI Act role.

Provider and deployer describe different activities

The AI Act’s definitions distinguish a provider that 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, from a deployer that uses an AI system under its authority. The provider definition covers paid and free supply.

This makes commissioning important. Paying a development company to write the software does not automatically make that company the provider of the resulting system. Examine who has the system developed and under whose name it is supplied or first used. Putting an AI system into service expressly includes first use for an organisation’s own purposes in the Union, so internal use is not an automatic escape from provider status.

Keep the model and the application separate, too. The Act recognises a downstream system provider that integrates an AI model supplied by another entity. A business can therefore provide an ERP-connected AI system without having developed the underlying model. Integrating a third-party model does not, on that fact alone, establish that the business is a general-purpose AI model provider.

The practical question is not simply “Who sold us the AI?” It is “What did each party develop or supply, and which system are we now putting into service?”

Architecture helps locate the system boundary

The ERP-agent webinar contrasts AI embedded within ERP software with external AI connected through interfaces such as APIs and MCP servers. These are useful architectural patterns for tracing the integration chain, rather than a basis for assigning legal status.

Consider an assistant asked to process an invoice. The enterprise-transformation discussion contrasts probabilistic agents with deterministic workflow execution: instead of improvising invoice processing, an agent can invoke an already-built workflow. That distinction makes the boundary tangible. Interpreting the request, selecting a tool and executing the accounting workflow can be separate parts of the arrangement.

For role allocation, draw those parts and identify their suppliers. Does the ERP vendor supply the complete assistant and its documented use? Does an integrator supply a separate application? Or does your business commission the orchestration, retrieval and interface as its own system? An external agent could be a finished supplier product; an embedded interface could front a system commissioned by your business. Location is not the deciding fact.

The investment tradeoffs between embedded ERP AI and external agents remain a separate question. A preferred architecture still needs its own role analysis.

The high-risk value-chain rules are more specific

Do not collapse the general provider definition into the special rules for existing high-risk systems. Article 25’s value-chain provisions identify circumstances in which a distributor, importer, deployer or other third party is treated as the provider of a high-risk AI system and becomes subject to the provider obligations under Article 16.

  • Putting its name or trademark on an existing high-risk system, without prejudice to contractual arrangements allocating the obligations otherwise.
  • Substantially modifying an existing high-risk system so that it remains high-risk.
  • Changing the intended purpose of an existing, previously non-high-risk system so that it becomes high-risk.

These are not a general rule that every ERP integration transfers provider responsibility. Nor should every technical change be called a substantial modification. The Commission-hosted implementation text describes the legal test by reference to changes not foreseen or planned in the initial conformity assessment that affect compliance with the high-risk requirements, or modify the intended purpose for which the system was assessed.

Creating a new system from components is a different route. The same implementation text distinguishes assembling and supplying a new AI system from modifying another provider’s existing system. Its explanation assigns the integrator responsibility as provider of the new system as a whole, with component-level compliance information assisting that responsibility rather than replacing it.

For a change review, first establish whether you are configuring an existing system, changing that system, or creating a new one. Then examine the relevant definition or high-risk trigger. Starting with “we changed some code” skips the distinction that matters.

Examples of ERP-connected role allocation

Illustrative example: using a vendor feature as supplied

A business enables a supplier-branded assistant to summarise purchasing records, within the purpose and conditions documented by the supplier. The starting point is use of the supplier’s system under the business’s authority, which points towards the deployer definition.

Confirm what the supplier actually supplies. Check whether the business has merely configured the feature or commissioned an additional system around it, whether the branding identifies the supplier, and whether the intended use remains within the documented purpose. An ERP subscription alone does not answer those questions.

Illustrative example: commissioning an internally branded assistant

A business commissions an assistant under its own name. A contractor connects a third-party model to purchasing records and an established invoice workflow, and employees use the resulting application internally.

Here, “we bought the model” and “the contractor built it” are incomplete answers. Having an AI system developed and putting it into service under the business’s own name can meet the general provider definition, including for own use. Examine the commissioned deliverable, the first intended use and the identity under which the complete system enters service. If the arrangement assembles components into a new system, analyse that new system rather than assuming it is a substantial modification of the model supplier’s product.

Illustrative example: repurposing an assistant for recruitment

A business takes a third-party general-purpose AI system not classified as high-risk and integrates it into an internal process to sift job applications. The Commission-hosted implementation text’s recruitment example treats this changed purpose as creating a high-risk system and making the deployer its provider under Article 25(1)(c).

That is a specific recruitment use, not a classification of all HR assistance or ERP AI. Examine whether the assistant is actually sifting applications, what influence its outputs have, and how the resulting system is classified. Merely attaching an “HR assistant” label does not describe the decision being supported.

Across these examples, intended purpose is more than an internal project title. The statutory definition includes the specific context and conditions described in instructions, promotional or sales materials, statements and technical documentation. Compare those records with the use you intend to introduce.

Make responsibility explicit before commissioning

Bring the system description, development scope, branding, intended purpose and proposed changes into the same review. Ask each supplier to identify what it provides and what documentation supports that boundary. Resolve the statutory role from the facts, then align the delivery agreement and handover with that conclusion.

Where the high-risk value-chain triggers apply, the consequence is provider obligations under Article 16, not merely a different project label. Before approving that scope, establish how the responsible organisation will obtain the information and technical support needed for the complete system.

Keep legal role allocation separate from operational authority. The distinction between drafting, approving and executing Odoo actions remains useful: permission to prepare content should not imply permission to commit a change. But a human approval gate does not itself answer who developed or supplied the system. Likewise, plan role-specific AI literacy alongside the system work, rather than treating employee training as a substitute for organisational role analysis.

If your ERP AI project involves commissioned development or a changed business purpose, discuss the system boundary and supporting technical work with Digiventi’s EU AI Act consultants and AI specialists before fixing the delivery scope.

Need help with data and AI?

We provide IT consultants for projects and ongoing technical work. Tell us what you need help with.

Contact us
Back to all articles