Skip to content
← Back to blog
AI ActArticle

AI Act provider vs deployer: map the role before building

Buying a model API does not settle your role in the finished AI system. Map the product, its intended use and the organisations involved before agreeing the implementation scope.

Author

Syntalith

Published Updated 5 min read

Buying a model API does not settle your role in the finished AI system. Before commissioning an application, identify the system being discussed, whose name it carries and how it will be used. Those facts give your legal adviser a useful basis for assessing provider and deployer roles under the EU AI Act.

This guide concerns preparation for that review. It does not classify your system, determine its risk category or give a deployment deadline. The references below were checked on 10 October 2026.

What provider and deployer mean

Under Article 3, a provider develops, or commissions development of, an AI system or general-purpose AI model and markets it, or puts the system into service, under its own name or trademark. Payment is not required. A deployer uses an AI system under its authority, excluding personal, non-professional use. Putting into service can include own use. These definitions are in the consolidated AI Act text.

Article 25 addresses provider responsibilities for high-risk systems when another party rebrands a system, substantially modifies it while it remains high-risk, or changes its purpose so that it becomes high-risk. For rebranding, the provision also recognises contractual arrangements that allocate the obligations differently. These conditions need case-specific assessment; an ordinary configuration change is not automatically a transfer of responsibilities. See Article 25 and its surrounding provisions.

Start with a diagram of the actual product

Consider a fictional company commissioning an assistant that drafts answers from its service manuals. A model vendor supplies the API. An implementation company builds retrieval, permissions and the review screen. The customer decides who uses the application and which manuals it may read.

Calling the API vendor “the provider” leaves several questions unanswered. Is the review about the underlying model, the vendor's application or the assembled assistant? Who commissions the finished system? Under whose name does it enter use? What do the instructions and sales materials say it is for?

Draw those components separately. Record the organisation responsible for each one and attach the relevant agreement or product documentation. Mark unknown answers explicitly. These proposed preparation steps give the legal adviser material to assess.

RecordMaterial to collect
System being reviewedApplication name, version and included components
Intended useUsers, tasks, outputs and excluded uses
Development arrangementWho builds or commissions each component
Name and presentationBranding, instructions and sales description
Operating arrangementWho controls access, inputs and use of outputs
Planned changesNew purposes, recipients, integrations or functionality

Revisit the assessment when the job changes

In the fictional example, the first release prepares answers for staff to review. Later, the customer proposes turning the same application into a separate product for other businesses. The team should record the new distribution arrangement and obtain an updated role assessment before treating the first assessment as sufficient.

A second change could involve the purpose of the output. A tool that retrieves documents might be proposed for evaluating people or making consequential recommendations. That requires a fresh description of what the system will do, who is affected and how the output is used. This example does not determine whether a particular system is high-risk.

For an engineering handover, keep the old and proposed scopes side by side. Include changes to input sources, permissions and review steps. A note saying “same model, new feature” leaves the adviser to reconstruct the application from incomplete information.

Separate the role assessment from the operating plan

Once the role assessment is available, the implementation team still needs named owners for acceptance tests, access changes and incident handling. A legal role label does not tell an engineer which person can approve a new connector or where to send a failed test.

The operating plan should refer to the assessed version and purpose. If either changes, assign someone to check whether the assessment and implementation requirements need updating. Our guide to responsibility for AI agent decisions covers the practical handover between process owners, reviewers and implementation partners.

Data-processing arrangements also need their own examination. The GDPR and DPA review guide addresses the agreement and data flow. Do not use a provider/deployer label as a replacement for that work.

What to ask Syntalith to prepare

Syntalith can scope the technical material for a role review: the application boundary, component list, data flow and proposed changes. The implementation proposal can identify which tests and records are included. Legal classification and advice should be assigned separately; this preparation is not compliance certification.

For the fictional service-manual assistant, a proposed first deliverable could be a short system map with an unresolved-questions list. The map would separate the purchased model API, the retrieval application and the customer’s operating process. It would identify which organisation commissions each part and whose name appears in the user interface and instructions. The questions list could flag that a later plan to sell the assistant has no agreed product owner yet. Your legal adviser would then assess those facts.

Before agreeing that work, we would check which supplier documents already exist and who in your team can confirm the intended use. The quote should say whether preparing the map ends with a handover or also includes updating it after the adviser’s questions. That distinction affects the work to commission; it does not imply a legal opinion is included.

For a first discussion, bring a description of one application and the change you want to make. We can compare the technical work needed with the information already available from your supplier. Service pricing is on the pricing page; use the contact form to describe the scope without sending confidential documents.

Working time: a calculation scenario for 2026

Scope: assembling a system-role dossier for legal review. This hypothetical scenario calculates team effort in the system review.

AssumptionValue
Tasks in the period24
Complete manual handling per task60 min
Preparation with the tool15 min
Employee review of the result25 min
Average allowance for clarification and corrections per task5 min

Manual handling requires 24 hours for this scenario.

Tool-assisted handling, including employee review and corrections, requires 18 hours.

The team recovers 6 hours for other work in the period.

Under these assumptions, implementation by Syntalith could reduce effort for this task by 25%.

Recovered-time formula:

24 × (60 − (15 + 25 + 5)) / 60 = 6 h

The relative reduction is recovered time divided by manual handling time. Before buying, measure complete handling on your own cases and include implementation, licence and maintenance costs. The appropriate employee retains the decision and approval.

Project context: primary reference. The table contains this scenario's planning assumptions.

Source and calculations checked: 21 October 2026.

Further implementation decisions

OpenAI Select Partner

Syntalith is an OpenAI Select Partner in the OpenAI Partner Network.

We help your team respond to customers faster and find information in company documents. We choose and set up the right AI tools, then teach your team how to use them.

How to introduce ChatGPT and OpenAI at work

Next step

Is there a process that takes your team hours every week?

Walk an engineer through it in 30 minutes. Within 2 business days you get, in writing, what it costs, what to do first and the starting price.

Process review
€0written summary within 2 business days

Times are shown in your own time zone. We work with clients across time zones.

See where to start

What our OpenAI and Anthropic partnerships mean for you.

OpenAI makes ChatGPT and Anthropic makes Claude. We belong to the partnership programs of both companies, so we work with their tools every day and learn from the creators themselves.

OpenAI Select PartnerClaude Partner Network
01We will select a tool for the task
We implement both ChatGPT and Claude. We check both on your documents, and you choose the one that makes fewer corrections and saves more time.
02Knowledge straight from the creators
We learn from official OpenAI and Anthropic courses and materials. Your team learns about the features and settings that work today.
03Licenses without our margin
When we roll out ChatGPT, Claude or Copilot, you buy the licences directly from OpenAI, Anthropic or Microsoft and they invoice you. We add nothing to the licence.
04Certified engineers
On the team, we have an AWS certification in machine learning and an Anthropic certification in systems architecture on Claude.

Participation in programs does not mean that OpenAI or Anthropic endorses our services.