Responsible AI & EU framework

1. Begin with the intended use

Before inference service activation, the parties identify the application, intended users, data categories, model components and operating territory. A conversational assistant, an image-production tool and an automated decision system can present different duties and risks.

We provide managed inference through authenticated APIs and agreed application integration services. Perchy retains server administration; customers do not receive server access or model-training or fine-tuning functionality. Depending on the actual activity, a party may also have duties as an AI-system provider, deployer, importer, distributor or general-purpose AI model provider. Contract labels alone do not determine legal roles.

2. European Union AI Act

The EU AI Act can apply to organisations outside the EU when the relevant conditions are met, including certain services placed on the EU market or systems whose outputs are used in the EU. Managed hosting or open weights do not create a blanket exemption.

Responsibilities depend on the actor, intended purpose and risk classification. We assess our own role for the agreed service. Customers assess their application and deployment duties, including prohibited practices, appropriate oversight and any applicable high-risk requirements. Each party remains responsible for obligations imposed on it by law.

Implementation dates and supporting guidance can change. Consult the European Commission’s current regulatory-framework guidance and applicable law for the operative requirements, rather than relying on an old timetable or a marketing badge.

3. Transparency and generated content

Applications may need to tell users that they are interacting with AI or that content has been generated or manipulated. Some uses require particular marking, disclosure, provenance or human-review measures. The product integration must implement the duties that apply; a website policy alone is insufficient.

Customers must review generated text, images and video for accuracy, rights and suitability before publishing or relying on them. We do not guarantee factual correctness, uniqueness, non-infringement or fitness for a regulated purpose.

4. Model licensing and documentation

We identify the selected model version and applicable licences when scoping the inference service. Commercial-use terms, redistribution, adapters and hosted add-on components can have different conditions. An open base model does not imply permission to use every feature of a provider’s hosted system.

Benchmark reports should state the model, hardware, output settings, sample size and methodology. Changes such as quantisation, distillation and reduced sampling steps can affect quality and cannot be assumed to be lossless.

5. Data protection and security

Data protection and AI-system rules address different obligations. Defined network boundaries can support control over data paths, but do not by themselves establish a lawful basis, satisfy transparency duties or eliminate model risks. Privacy notices, processing contracts, access controls and appropriate assessments remain relevant.

For our handling of website enquiries, read the privacy notice. Customer inference processing is governed by the relevant service and processing agreements.

6. Official guidance and questions

Relevant official resources include the European Commission’s AI regulatory framework, the Commission’s guidance on general-purpose AI models, and the UK ICO’s guidance on AI and data protection. Links to these sources are provided below.

For questions about a proposed inference service, contact hi@perchy.ai. Provide the use case and territory so the relevant obligations can be discussed in context.