1.You are talking to an AI
Every agent built on integrable.cloud is an AI system. It is not a person, it is not a person typing quickly, and it is not a person supervising each message before it is sent.
We are the provider of that system under the EU AI Act, and the customer whose website you are on is the deployer. The duty to tell you that you are interacting with an AI is ours.
2.How you are told
In every agent since 7 October 2026
Article 50 of the EU AI Act became applicable on 2 August 2026. It requires the disclosure to be made clearly and distinguishably, at the latest at the first interaction. Every agent built on integrable.cloud carries it, in the visitor’s own language:
- An “AI agent” label on the chat’s header, on every message the agent offers before anyone types, on the agent island, and on the launcher’s message - visible before the first answer, not after it.
- In a spoken conversation the label stays on screen beside the voice, with a note that the audio is not stored.
- It is not a setting, and the styling a customer can add has no way to name it. Naming an agent and giving it a personality stays available; presenting it as a human does not.
- Asking an agent whether it is a person or an AI always gets a straight answer: it says it is an AI agent.
Our Acceptable Use Policy also prohibits customers from hiding it, denying it, or presenting AI-generated output as human-written.
3.How an answer is produced
Understanding this explains both what the agent is good at and how it fails.
- Your question is matched against the business's own content — their pages, documents and help articles — by meaning rather than keyword.
- The most relevant passages are retrieved.
- Those passages and your question are sent to a large language model, which writes an answer from them.
- The answer is returned with the source it came from, so it can be checked.
The important consequence: the agent is answering from that business’s content, not from general knowledge about the world. If their page is wrong or out of date, the answer will be too — and the citation is what makes that visible and fixable rather than mysterious.
4.Which models, and what they receive
| Provider | Role | What it receives |
|---|---|---|
| Groq | The default model that generates answers | The visitor's message and the retrieved passages from the business's content. |
| Anthropic | Fallback and premium models | The same, when routed there. |
| Groq, Mistral AI | Hearing a visitor who talks instead of typing | The recording, turned into text and never stored. Only the words reach the conversation. |
| Fish Audio, xAI, Microsoft, DeepInfra | The agent's spoken voice | Only the words the agent says - never the visitor's voice. |
This transfer happens on every answered message and it is inherent to how the product works rather than a feature that can be switched off. The DPA states this as clause 6 because a customer needs to know it before they deploy, not after. The voice rows apply only when a visitor talks to an agent with voice on.
A business can give its agent its own voice: someone reads a statement of consent aloud, and that recording is both the proof of consent and the voice the agent learns. A clip that does not read the statement is refused. Nobody is recognised by their voice - not the visitor, and not the speaker.
The full list of everything that touches data, and whether it is in the path always or only when connected, is at subprocessors.
5.What is not done with it
No training, by anyone in the chain
Conversations, knowledge base content and captured leads are not used to train, fine-tune or evaluate any AI model — not by us, and not by the model providers, whose commercial terms with us prohibit it.
This is a contractual commitment rather than a setting. It does not vary by plan, it is not something you have to opt out of, and it applies on the free plan exactly as it does on the most expensive one.
6.What it gets wrong
Every AI agent vendor has a page like this. Most of them use it to reassure. Here is the actual list.
- It can be confidently wrong. Grounding an answer in real content and citing the source makes errors visible and fixable. It does not make them impossible. Nothing does.
- It can miss something that is there. If a passage is phrased very differently from the question, retrieval can fail even though the answer exists on the site.
- It does not know what changed this morning. It knows what has been indexed.
- It has no judgement about consequences. It cannot tell that one question is routine and the next one matters enormously to the person asking.
- It should say when it does not know, and it is designed to. It will not always manage it.
Nothing an agent says is legal, medical, financial or professional advice, and no business using one may present it as such.
7.Human oversight
Any conversation can be handed to a person. Handover here is absolute: once a human takes a conversation over, nothing automated replies into it again, including a scripted flow. An agent that keeps interrupting a human colleague is worse than no agent.
Businesses deploying an agent should give people a route to a human for anything consequential. We build the route; using it is their design decision.
8.If you deploy an agent, these duties are yours
You are the deployer. Some obligations cannot be discharged for you by the provider, and it is better to read them here than to meet them in an enforcement letter.
- AI literacy. AI Act Article 4 has applied since February 2025 and requires you to ensure a sufficient level of AI literacy among staff who operate the system.
- Disclosing AI-generated public-interest text. Article 50(4) applies to you, not us, if you publish AI-generated text to inform the public on matters of public interest.
- Your privacy notice and lawful basis. The agent runs on your site and talks to your visitors. You are the controller — see the DPA.
- Not using it for a prohibited or high-risk purpose. The list is in the Acceptable Use Policy, and it is the clause we terminate for.
9.Reporting a bad answer
If an agent told you something wrong, the business running it can fix it fastest — they own the content the answer came from. Tell them first.
If it said something harmful, deceptive, or claimed to be human, tell us at abuse@integrable.cloud. You do not need an account and you can report anonymously. We acknowledge within one business day.