How we use artificial intelligence
We sell checks on AI transparency duties. Not declaring our own would be odd. This page says where AI enters our work and where it does not. It is also the model of what we deliver to a client who asks us for the same thing: if you need a declaration like this for your company, take it.
Where it enters, and where it does not
| Activity | AI systems | Who answers for the result |
|---|---|---|
| Site copy, guides, study | Yes, drafted with an assistant | The controller, who checks and signs before publishing |
| Code of the inspection tools | Yes, written with the same assistant | The controller, who runs it and puts it through automated tests |
| The measurements in reports: WCAG violations, DNS records, headers, certificate | No | axe-core and network queries: same site, same result |
| Composition of the report | No | A document template filled with the measured figures |
| Reply emails to whoever asks for a measurement | Assisted draft | The controller reads it and sends it |
| Chat, assistants or automated replies on the site | None | There are none |
| Decisions on prices, outcomes, accepting a client | None | No automated decision within the meaning of Article 22 GDPR |
The third row is the one that matters most. The figures in our reports were not written by a model: they come from axe-core against the WCAG 2.2 rules, together with DNS queries and header reading. These are deterministic tools: on the same site they give anyone the same result, including you, and you can run the measurement yourself.
What we do not do
- We publish no generated images or video depicting real people, places or events.
- No review, testimonial or client case is invented or generated. As of 9 September 2026 we publish none at all, for the simplest reason: we have no clients to cite yet.
- The figures in the study and in the reports come from measurements, not from a model’s estimate. When a measurement is not reliable we say so, instead of filling the gap with a guess.
- No client content is used to train models.
The provider
The writing and coding assistant we use is Claude, by Anthropic, on a Max subscription. That is a consumer service: the provider does not act as a processor on our behalf and there is no agreement under Article 28 GDPR, as there is with Resend, Vercel and Stripe. We say so because it is the reason the rule below is not an extra precaution: it is the condition that makes using the tool workable.
An account setting the provider calls “Model Improvement” decides whether conversations and coding sessions are used to improve the models, and it covers the command-line tool used to write beaudited’s code. It is off, verified on 9 September 2026; with the setting off, retention at the provider is thirty days, with it on it would rise to five years.
The rule we set ourselves on data
What goes into the work with the assistant is the domain to be analysed and the technical measurements of the site, which are public data: the same anyone can read by opening the page. What does not go in are names, email addresses, client correspondence and access credentials. Temporary credentials, where a job requires them, are used only on the client’s systems and revoked when the work ends.
Why this page exists
Article 50(4) of regulation (EU) 2024/1689 provides that, for text published to inform the public on matters of public interest, the duty to declare artificial generation falls away where there is human review with editorial responsibility. Here there is, and this page makes it verifiable rather than leaving it implied. The other parts of Article 50 do not concern us, because the site has no systems interacting with visitors: the guide explains when they do concern you.
If you find something wrong
If you find content on our pages that looks generated without review, or a figure that does not add up, write to email. We correct or remove it, and the correction stays recorded below. It is the same reporting channel we ask clients to open on their own sites.
Corrections
None as of 9 September 2026.
Last updated: 11 September 2026. First review due at the first paying client, or sooner if anything changes.