EU AI Act Integrator: Proving AI Reliability upon Acceptance

The integration of artificial intelligence into industrial environments is redefining the standards of technical compliance. For system integrators, delivering a robotic cell is no longer limited to mechanical validation: it now requires proving algorithmic reliability through prediction, in real time, before the signing of the acceptance report. As the European Business Review points out, the preparatory obligations of the EU AI Act are already in force and engage the responsibility of deployers today. TrustalAI, as a reliability layer for industrial AI, supports you in your regulatory proofs required by the EU Regulation 2024/1689.

Machinery Directive and EU AI Act: Two Texts, a Single Responsibility

The system integrator finds themselves caught between two legal obligations that accumulate without replacing one another. The compliance of an industrial installation requires simultaneously meeting the physical safety imperatives of the Machinery Directive and the new algorithmic traceability constraints of the EU AI Act. As Tech Policy Press highlights, EU AI Act compliance is becoming the default path for any AI deployment in Europe.

To anchor this legal reality, one must refer to EU Regulation 2024/1689 (EU AI Act) and the Machinery Regulation 2023/1230. The integrator alone bears this double regulatory constraint towards their final client, regardless of the quality of the documentation provided by their AI module supplier.

What Machinery Regulation 2023/1230 Covers

The Machinery Directive imposes a strict obligation of result on the safety of the delivered machine as a whole. As soon as a robotic cell integrates an AI vision module whose probabilistic behavior cannot be proven reliable, the integrator assumes the full legal risk, including for perception errors of the AI module that they did not develop.

Let's take a concrete example: a robotic picking cell using industrial AI vision to guide the manipulator arm. If the AI module produces an undocumented detection error leading to a collision or a critical defect, it is the integrator who is held responsible before the final client, because they certified the entire system. As detailed by Robotics & Automation News, EU Regulation 2023/1230 precisely marks the transition from "freeze," one-time validation at delivery, to "continuous assurance": qualification does not stop at technical acceptance.

What the EU AI Act Adds That the Machinery Directive Did Not Cover

The Machinery Directive validates the mechanical and electrical safety of the installation. The EU AI Act imposes an additional layer: documenting the reliability of the AI component itself, prediction by prediction, in real time. Where the traditional directive requires the machine to stop in the event of hardware failure, the EU AI Act requires understanding why the model made a specific decision when faced with environmental variation, lighting changes, model drift, out-of-distribution parts, or tool wear.

Concretely, the EU AI Act requires for Annex III high-risk systems:

• Automatic logs per inference with confidence scores at the exact moment (Art. 12)

• A documented and activatable human oversight mechanism (Art. 14)

• Continuous risk management throughout the lifecycle (Art. 9)

• Comprehensive technical documentation including performance in actual deployment contexts (Art. 11 + Annex IV)

It is precisely this layer of algorithmic traceability that most integrators do not yet have in place, and that the Machinery Directive alone did not impose on them during technical acceptance.

The Integrator as "Deployer" under the EU AI Act

An integrator deploying a high-risk AI system is considered a deployer under EU Regulation 2024/1689. They bear their own regulatory obligations, independent of those of the AI provider. They cannot transfer this responsibility by contract.

These specific obligations for the deployer are listed in Articles 16, 17, and 26 of the Regulation. Article 26 notably requires the deployer to keep the automatic logs generated by the system for an appropriate period, an obligation that Datenschutz-Notizen details as the core of the compliance infrastructure for high-risk AI systems. The deployer status is attached to the act of deployment, not to the commercial contract.

What the Integrator Must Prove upon Acceptance

Before accepting a robotic cell integrating AI, the integrator must prove: (1) that the AI system has been evaluated in its actual deployment context (Art. 9), (2) that each AI decision is traced with its confidence level at the exact moment of inference (Art. 12), (3) that a documented human oversight mechanism is activatable if confidence falls below a defined threshold (Art. 14), and (4) that this technical documentation is maintained continuously, not just produced at delivery.

Traceability by Prediction: The Most Overlooked Obligation

Article 12 imposes automatic logging of events to identify risky situations. Unlike post-mortem monitoring that notes errors once the action is executed, reliability by prediction logs the confidence score of each inference in real time before the mechanical decision is made. As The Recursive puts it: "If you can't reconstruct a decision, you can't ship." The EU AI Act literally requires the ability to replay each AI decision with its exact context.

TrustalAI automatically generates the timestamped logs per individual prediction required by this article. Each inference is traced with its exact confidence level before the mechanical decision is made, in less than 100ms in the Cloud and 20ms at the Edge, thereby providing the proof of auditability required during the Site Acceptance Test (SAT).

Human Oversight: Proven, Not Just Declared

Article 14 requires that high-risk systems be designed to allow for effective human oversight. It is not enough to declare that an operator is monitoring the line. As confirmed by DevDiscourse, 100% automated AI decisions without proof of measurable reliability face explicit legal barriers under the EU AI Act.

TrustalAI produces the documented and activatable escalation thresholds required by Art. 14. By quantifying model uncertainty at each prediction, the system mathematically determines the precise moment when the AI model is no longer able to make a safe decision, triggering a transfer of control to the human operator based on objective confidence metrics.

Continuous Risk Management: Not Just an Acceptance Report

Article 9 states that the risk management system must consist of a continuous iterative process throughout the entire lifecycle. The classic mistake made by integrators: producing a compliance report at delivery and then considering the matter closed. The EU AI Act imposes post-deployment monitoring, which is precisely highlighted by the shift from "freeze" to "continuous assurance" described by Robotics & Automation News.

TrustalAI ensures reliability by prediction continuously, without modifying the existing model. By acting as a black-box compatible overlay, the solution detects model drift, perception anomalies, and out-of-distribution situations throughout the lifecycle of the robotic cell.

The Contractual Risk That Integrators Do Not Anticipate

The legal risk associated with AI deployment is systematically underestimated, as integrators mistakenly believe that liability lies exclusively with the AI module provider. In practice, this lack of awareness translates into critical situations during production launch. This is not a theoretical debate on European law, but a direct threat to project profitability, billing milestone validation, and industrial operator safety.

When the Client Refuses Acceptance

At the Site Acceptance Test (SAT), the client requires proof that the embedded AI module in the cell is reliable, not just that the machine is mechanically safe. Without prediction reliability logs and documented human oversight thresholds, the integrator cannot respond to this legitimate demand. This scenario of refusing technical acceptance is becoming increasingly common as industrial clients become aware of their own EU AI Act obligations as end-users.

The immediate consequences:

• Blocking of acceptance and freezing of the final payment milestone (often 20-30% of the total amount)

• Application of contractual delay penalties

• Severely damaged client relationships

• Potential triggering of the performance guarantee

Without a solution capable of providing real-time prediction traceability, resolving this can take months of back-and-forth with the provider of the algorithmic component.

Compliance Cannot Be Transferred by Contract

Even if the AI provider certifies their module and provides comprehensive technical documentation, the integrator deploying it within a cell carries their own deployer obligations (Articles 16, 17, and 26 of EU Regulation 2024/1689). A subcontracting clause or an obligation of result does not transfer regulatory obligations: they remain attached to the act of deployment and commissioning in the client's final environment.

Compliance is achieved through the rigorous documentation of the system's behavior in its actual deployment environment, and not through a simple commercial agreement.

What TrustalAI Produces for the Acceptance File

To meet the requirements of EU Regulation 2024/1689, TrustalAI automatically generates three documentary deliverables, essential for validating the technical acceptance of a robotic cell integrating AI.

Sector benchmarks show a 40% reduction in perception incidents on industrial robotic systems (VKTR, 2024). The approach is entirely plug-and-play on the existing AI module: it is installed without modifying the model and without changing operational processes. The first documentary elements — Art. 12 logs, Art. 13 metrics, and Art. 14 thresholds, based on your actual production data, are available within 2 weeks.

Conclusion: Secure Your AI Projects, Prove Your Reliability

To validate the technical acceptance of an AI-powered robotic cell, the integrator must prove risk assessment in the actual deployment context (Art. 9), traceability by prediction (Art. 12), documented activation of human oversight (Art. 14), and the continuous maintenance of this documentation throughout the lifecycle. The Machinery Directive 2023/1230 and EU Regulation 2024/1689 accumulate strictly: one does not replace the other, imposing a double mechanical and algorithmic validation.

You deliver a machine that knows when it does not know, and you can prove it to your client.

The EU AI Act obligations for AI integrated into machines will apply on August 2, 2028 (Annex I, date set by the Digital Omnibus). Until then, building the technical acceptance file for high-risk AI systems remains a priority for any integrator active in industrial robotics and AI vision projects.

FAQ: Your Questions on the EU AI Act and Integrator Liability

Does the Machinery Directive cover the AI components of a robotic cell?

Not entirely. The Machinery Regulation 2023/1230 covers the mechanical safety of the delivered machine as a whole. It does not provide for documentation of the reliability of the AI component itself: its prediction logs, its confidence metrics, or its activatable human oversight mechanism. Historically, regulation on industrial machinery was designed for deterministic systems. Integrating industrial AI vision models introduces probabilistic behavior that the Machinery Directive cannot audit millisecond by millisecond. This is precisely the additional layer imposed by the EU AI Act for high-risk AI systems (Annex III, EU Regulation 2024/1689).

Can an integrator transfer their EU AI Act liability to their AI provider?

No. An integrator deploying a high-risk AI system carries their own deployer documentary obligations (Articles 16, 17, and 26 of EU Regulation 2024/1689), regardless of the contract signed with the AI provider. Article 26 notably requires the deployer to keep automatic logs generated by the system for an appropriate duration. If an accident occurs on the production line, market surveillance authorities will demand proof that the system was correctly supervised in its actual deployment environment. A financial compensation clause does not void this regulatory obligation. Consult a legal expert for the analysis of specific contractual risks.

What happens if the client refuses acceptance due to a lack of AI reliability proof?

The client can legitimately block acceptance if the integrator cannot produce the algorithmic reliability documentation required by the EU AI Act. Without prediction reliability logs (Art. 12) and documented human oversight thresholds (Art. 14), the integrator is unable to prove that the embedded AI module is compliant during the Site Acceptance Test (SAT). Consequences: freezing of the final payment milestone, contractual delay penalties, and triggering of the performance guarantee. Consult a specialized legal advisor to analyze the risks specific to each contractual situation.