Article 50 disclosure

The AI disclosure is added by the runtime, not by a prompt.

OpusDone adds the Article 50 disclosure after the model has produced its text and before the message is stored, once per conversation and in the customer’s language. Nothing in the prompt can take it out: you can edit the wording, but you cannot remove it. Article 50 applies from 2 August 2026, and penalties under the Act reach €15M or 3% of worldwide turnover.

Request access See how it works

Applies from 2 August 2026
Prepended after the model writes
Once per conversation, marked on record

Reviewed 27 August 2026

The gap

Why is an instruction not a disclosure?

Most assistants handle this by asking the model to mention it. That is a request, and requests get dropped.

A prompt line is a hope

Telling a model to open by saying it is an AI works most of the time. Most of the time is not a control, and the one conversation where it forgets is the one you will hear about.

A custom greeting can quietly replace it

Someone rewrites the opener to be warmer, or shorter, or more on-brand. The disclosure goes with it, and nobody notices for weeks because everything still looks fine.

The word “assistant” is not enough

A customer reading “your assistant” has been told nothing about what they are talking to. In OpusDone that word alone deliberately does not count as a disclosure.

Where it is added

What does OpusDone do with every first reply?

01
The model writes its answer
Normally, within its scope, using only knowledge marked live. At this point nothing has been shown to the customer and nothing has been stored.
02
Your own wording is checked first
If the business’s text already discloses that this is an AI, nothing is added and the conversation is simply marked. You are not made to say it twice because you wrote a good opener.
03
The line is prepended, not appended
It goes in front of the reply, before the message is stored. A customer reads it before they read the answer, rather than finding it under a paragraph about your opening hours.
04
It is written in the customer’s language
The disclosure is rendered in the language the customer is writing in, in line with the rest of the conversation. A disclosure nobody can read is not doing its job.
05
It is marked on the conversation record
It appears once per conversation, and the fact is recorded against that conversation. A customer who comes back next week starts a new conversation and is told again.

What can a prompt edit remove from this?

Nothing. The disclosure is added after generation, so there is no instruction to override and no opener to overwrite. Change every word of your prompt and the customer is still told.

Request access Talk to us

Article 50

How does the OpusDone disclosure actually work?

This is a description of what the product does. It is not legal advice.

The disclosure is added after the model writes, before storing.
It is prepended to the message, not tacked on the end.
No prompt edit can remove it. It is not in the prompt.
Once per conversation, marked on the conversation record.
A returning customer is told again in the next conversation.
Written in the customer’s language, across all nine languages.
If your own wording already discloses, nothing is added.
The word “assistant” alone does not count as disclosure.
Article 50 applies from 2 August 2026.
Penalties reach €15M or 3% of worldwide turnover.

Guardrails

Five things that cannot happen here.

This is the part of the product where “usually” would be the wrong standard, so it is not the standard.

No prompt edit removes it

The disclosure is not generated by the model, so editing what the model is told changes nothing about whether it appears.

No client-authored opener replaces it

A rewritten greeting either already discloses, in which case the conversation is marked, or it does not, in which case the line is prepended anyway.

It is never buried at the end

Appending it would put it after the answer, where it is read last or not at all. It goes first, every time.

It is never English-only

The customer’s language decides the wording, the same rule the rest of the reply follows.

A returning customer is never assumed to remember

The marker is on the conversation, not the person, so a new conversation gets the disclosure again.

EU AI Act questions, answered

When does Article 50 start to matter?

It applies from 2 August 2026. Penalties under the AI Act reach €15M or 3% of worldwide turnover. We have built the disclosure into the runtime rather than the prompt so that the date does not require you to change anything.

What if I already tell people it is an AI?

Then nothing is added. The check looks at whether the business’s own wording discloses that this is an AI; if it does, the conversation is marked and the customer reads your version. The word “assistant” on its own does not satisfy the check.

What about conversations that need a person?

Escalation has five reasons, and sensitive is locked on in the interface because it is enforced in the runtime rather than offered as a setting. If a customer asks for a human and the model does not escalate, the runtime creates the escalation itself.

Is this legal advice?

No. This page describes what the product does and nothing more. Whether and how the Act applies to your business is a question for your own adviser, and we will not pretend otherwise.

Does the disclosure appear on every message?

No, once per conversation. It is prepended to the first reply and marked on the conversation record so it is not repeated on every turn. A new conversation, including one from a returning customer, gets it again.

What is built on the data side?

Business data is isolated and every query is scoped to it. Credentials are encrypted at rest with AES-256-GCM, and the application, database and file storage run on servers in Germany. Erasure works end to end, and export produces a ZIP in JSON and CSV with internal team notes excluded.

Keep reading