AI Act: what you need to know as a customer

Fine-tuning your model with LoRA uses only a fraction of the compute originally spent training the underlying base model. Under the Regulation, fine-tuning only turns the fine-tuner into a provider of its own base model once it reaches a substantial share of that original training compute; the LoRA adaptation falls far short of that threshold.

Assigning the provider and deployer roles depends on the circumstances of the individual case. If in doubt, have your specific role reviewed by a lawyer.

• Machine-readable, signed labelling of outputs

  Every answer your system returns through its interface carries structured metadata that marks it as AI generated and cryptographically signs its origin. This fulfils the provider obligation under Article 50(2) and follows the voluntary EU code of practice on labelling of 10 June 2026, which recommends signed metadata as the benchmark.

• A verification tool

  Use it to verify the signature of any output independently of IonKon, at any time.

• This documentation

  This overview and the accompanying technical documentation support you in meeting your own obligations as deployer (see our terms and conditions, section 14).

• A visible notice in the interface

  If you use the interface delivered by IonKon, the notice that users are writing with an AI system is already set up; you do not need to implement it yourself. It appears in both German and English as a bar along the top edge, on every page and permanently, and cannot be dismissed. It is produced by a component of our delivery, not by the interface. If you use your own interface, you must ensure there that users can recognise they are interacting with an AI system, unless this is already obvious (Article 50(1)).

• Purpose limitation

  Use the system within its intended purpose: as an internal assistant system based on your own company documents.

• AI literacy of your staff

  Providers and deployers must take measures to foster the AI literacy of the people who work with the system or use its outputs (Article 4). Since the Digital Omnibus recast, no specific level of knowledge for individuals is prescribed. The obligation has applied since 2 February 2025; market surveillance authorities have been enforcing it since 2 August 2026.

• No prohibited or high-risk use without prior agreement

  Using the system for practices prohibited under the AI Act, or as a high-risk AI system under Annex III, requires prior agreement with IonKon (see our terms and conditions, section 14).

The verification tool delivered with your system reads an output's metadata and checks its signature against your system's public key; the result is a simple yes or no. A short guide is included with the delivery.

The Digital Omnibus (Regulation (EU) 2026/1744, in force since 27 July 2026) postponed the high-risk obligations and recast Article 4. The transparency obligations under Article 50 are not affected.