Many DACH engineering leaders assume that if an external team develops their AI system, the external team carries the EU AI Act obligations. Under Regulation (EU) 2024/1689 that assumption is usually wrong. The provider is the organisation that develops an AI system or has it developed and places it on the market or puts it into service under its own name or trademark (Article 3(3)). If your company commissions the work and ships the result under your brand, you are, as a rule, the provider. The external team is a supplier of services and components. The practical consequence: you need to contract for the evidence that the provider role requires, not just for working code.
Where German companies stand
For most DACH companies the question is no longer theoretical. Bitkom's September 2026 survey of 603 German companies found that 57 percent now use AI, up from 36 percent in 2025 and 20 percent in 2024. The same survey shows the AI Act moving onto the agenda: 37 percent expect to be affected as users of AI (23 percent a year earlier), 34 percent are still checking, and companies that expect to be affected anticipate an average of 2.1 high-risk AI systems. Among companies that are affected or still checking, 89 percent expect a high implementation effort. Official statistics are more conservative but point the same way: according to Destatis, 26 percent of German companies with at least ten employees used AI technologies in 2025. Where such systems are built or extended by external engineers, the provider question decides who carries the evidence burden.
The four roles that matter in practice
The AI Act distinguishes several operator roles. For a scaleup or Mittelstand company working with an external engineering team, four are relevant.
- Provider (Article 3(3)): develops or has developed the system and places it on the market or puts it into service under its own name. Carries the bulk of obligations for high-risk systems, including the quality management system (Article 17), technical documentation (Article 11 and Annex IV) and conformity assessment.
- Deployer (Article 3(4)): uses an AI system under its authority in a professional context. Obligations are narrower, for example use according to instructions, human oversight and log retention for high-risk systems (Article 26).
- Third-party supplier: supplies tools, services, components or processes that are integrated into a high-risk AI system. Article 25(4) requires the provider and such a supplier to specify, by written agreement, the information, capabilities, technical access and other assistance needed for the provider to meet its obligations.
- Deployer that becomes a provider (Article 25(1)): a deployer or other third party is treated as provider of a high-risk system if it puts its name or trademark on it, makes a substantial modification to it, or changes the intended purpose of a system so that it becomes high-risk.
The last case is the one most teams miss. A company that licenses a third-party model, wraps it in its own product and markets it for an Annex III use case can move from deployer to provider without noticing.
The timeline after the Digital Omnibus
The Digital Omnibus on AI, now published and in force, shifted the high-risk deadlines. Obligations for Annex III high-risk systems apply from 2 December 2027 (previously 2 August 2026). For high-risk systems covered by Annex I sectoral legislation (Article 6(1), for example machinery or medical devices), the date is 2 August 2028. Systems already on the market before the relevant date are, in general, only caught if they are substantially modified afterwards. Prohibited practices under Article 5 and the obligations for general-purpose AI models already apply.
The delay is not a reason to postpone engineering work. Annex IV documentation is very hard to reconstruct after the fact. Data lineage, design decisions, evaluation results and versioned training configurations must be captured while the system is being built. If an external team builds it, that capture has to be part of their scope.
Mapping provider obligations to deliverables
The table below maps the main high-risk requirements to what an external engineering team can concretely deliver. The provider remains accountable for every line. The supplier produces the evidence.
| AI Act requirement | Deliverable from the external team | Acceptance evidence |
|---|---|---|
| Risk management (Art. 9) | Technical risk register entries for model failure modes, test cases per identified risk | Register version linked to release tag |
| Data and data governance (Art. 10) | Dataset datasheets, lineage records, bias examination reports, split definitions | Reproducible data pipeline run with hash of each dataset version |
| Technical documentation (Art. 11, Annex IV) | Architecture description, training methodology, evaluation protocol, compute and tooling description | Annex IV section checklist completed per release |
| Record-keeping (Art. 12) | Logging design and implementation for inference events | Log schema and retention test |
| Transparency to deployers (Art. 13) | Draft instructions for use: capabilities, limitations, performance metrics | Reviewed by provider product owner |
| Human oversight (Art. 14) | Override and stop mechanisms, confidence display, escalation paths | Functional test in staging |
| Accuracy, robustness, cybersecurity (Art. 15) | Evaluation suite, adversarial and drift tests, model artefact signing | Evaluation report with thresholds agreed in advance |
A clause pattern for the Leistungsbeschreibung
Under a Werkvertrag (BGB §631), acceptance (Abnahme, BGB §640) is tied to defined results. That structure fits the AI Act well, because the provider needs defined artefacts. A workable pattern for each sprint or release:
The contractor delivers, for each release, (a) the model artefact with version identifier and signature, (b) the training and evaluation configuration sufficient to reproduce the artefact, (c) updated Annex IV documentation sections 1 to 3 as defined in the attached template, and (d) an evaluation report against the acceptance thresholds in Appendix B. Acceptance requires all four items. The client remains provider within the meaning of Article 3(3) Regulation (EU) 2024/1689 and decides on placing on the market.
This wording does two things. It makes evidence a condition of acceptance rather than a favour, and it states the role allocation explicitly, which supports the written agreement Article 25(4) expects for high-risk systems.
Fine-tuning and general-purpose models
Many teams build on a general-purpose AI model rather than training from scratch. Fine-tuning or otherwise modifying such a model can, depending on the extent of the modification, make you the provider of a modified general-purpose model. The Commission's guidelines on general-purpose AI providers use an indicative threshold based on the compute used for the modification relative to the original model (one third of the original training compute). Most product fine-tunes stay well below that threshold, but the assessment should be documented, and the external team should record the compute used for each fine-tuning run so the figure exists when someone asks.
Checklist before you sign
- Classify the intended purpose: Annex III, Annex I product, Article 50 transparency case, or minimal risk. Write the result down.
- Confirm who will place the system on the market under whose name. That party is the provider.
- Check whether any licensed third-party model would make you a provider under Article 25(1).
- Attach an Annex IV documentation template to the contract and make its sections part of acceptance.
- Agree evaluation thresholds before training starts, not after results are known.
- Require reproducibility: pinned dependencies, dataset hashes, seeds, and infrastructure as code.
- Log fine-tuning compute per run.
- Clarify data protection roles separately: the AI Act role and the GDPR role (controller or processor) are different questions.
Where an external pod fits
An external team is well placed to produce the engineering evidence the provider needs, provided the evidence is specified up front. In practice, one senior role should own the documentation and evaluation structure while engineers deliver the pipeline and model work, and both should be scoped as deliverables in the Leistungsbeschreibung. Details on how we handle data and access are on our security page.
This article is general information and not legal advice. For a binding assessment of your role and obligations under the AI Act, consult a qualified lawyer.
The Omnibus has bought time, not removed the work: the evidence a provider needs is cheapest to produce while the system is being built. In a niogod™ pod, the Senior Consultant owns the Annex IV documentation and evaluation structure as a defined Werkvertrag deliverable, accepted sprint by sprint alongside the code. If a high-risk system sits on your roadmap before December 2027, it is a conversation worth having early.
Talk to niogod™
Planning engineering capacity in DACH? Book a 30-minute consultation with the niogod™ team.
Book consultation