How AI is used

Version of 13 September 2026. Earlier versions are available on request.

Article 50 of the EU AI Act has applied since 2 August 2026. It requires that people are told when they are interacting with an AI system, and that synthetic text published to inform the public is disclosed as such. This page is that disclosure, and it also says the things the Act does not require but a customer deserves to know.

What is generated, and what is not

• Generated: the first draft of a site’s words — headings, service descriptions, the about section — produced by a large language model from a brief you fill in.
• Not generated: the design. Layout, typography, colour and structure come from design systems written by hand. The model never emits HTML, CSS, class names or colours; it fills in a fixed content shape and nothing else.
• Not generated: facts. Opening hours, phone numbers, prices, certifications and addresses come from what you enter or from public registers.

The rule the generator is held to

It is not allowed to invent a verifiable claim. Certifications, testimonials, prices, years in business and staff numbers are either supplied and confirmed, or marked as unconfirmed and listed for you to check with the business owner. A site cannot be published while any placeholder remains — that is enforced by the pre-publish checks, not by good intentions.

This matters because the damage runs in one direction. A website that claims a tradesman is certified when he is not is a problem for him first, and he did not write it.

A person publishes, always

Nothing reaches the public automatically. Every generated draft is reviewed and published by a human being, who is responsible for its accuracy. There is no automated decision-making producing legal or similarly significant effects within the meaning of Article 22 of the GDPR.

Risk classification

Livery uses general-purpose AI models through an API to generate text. It is not a high-risk system under Annex III: it does not assess creditworthiness, employment, education, essential services, law enforcement or migration. Our obligation is the transparency one in Article 50, discharged here and in the product.

What is sent to the model

The business brief: trade, town, services, opening hours and any notes you add. Not your account details, not your client list, and not the contents of other sites. The provider and its location are named on the sub-processor page. An operator running a local model sends nothing outside their own infrastructure.

Who owns the output

You and your client do. We claim no rights in the copy generated for a site. Note that purely machine-generated text may not attract copyright at all in some jurisdictions, which cuts both ways: it is yours to use, and it may be harder to stop someone else reusing it. Text a person edits meaningfully is on firmer ground.

Mistakes

Language models produce fluent text that is sometimes wrong, and fluency is exactly what makes the errors hard to catch. The pre-publish checks catch unfilled placeholders and unverified claims; they cannot catch a confident sentence that is merely inaccurate. Read the draft. You are the one whose name is on the invoice to the business.