AI use policy

• The AI agent answers people who chat or message your business, from the content you give it, and hands over to a person when asked or when it cannot help.
• ATON Assist helps your team: it drafts replies, rewrites them, summarises conversations and translates both ways. Nothing it writes is sent without a person choosing to send it.
• A handover summary gives the person taking over the conversation so far, so your customer never repeats themselves.

All of them run on AI models made by Anthropic.

Every answer from the AI agent is marked as AI, in the chat and on every messaging channel, and the way to reach a person is always there. This is the same on every plan, and white label never removes it.

The AI never claims to be a person, and you may not set it up to. When nobody is available, it says so and offers to take a message.

To write an answer or a draft, the AI is given only what that answer needs:

• the conversation so far;
• the parts of your training content that match the question: website pages you asked us to scan, files, answers, help articles, catalogue feeds, and past chats a person has approved;
• your settings for it, such as its tone, its name and topics to avoid;
• what your own website or app says about a signed-in customer, if you set that up.

It never sees other workspaces' information, and never sees payment card details, which ATON Desk does not hold.

• Answer from your content, and say when it does not know.
• Hand the conversation to a person, or take a message when nobody is there.
• Run the actions you define, such as looking up an order. Before an action changes anything, it asks your customer to confirm.

What visitors write, and the pages and files it learns from, are treated as information, never as instructions. The limits on what it may do are built into the actions themselves, not left to what it is told.

The AI's answers are written by Anthropic, outside Canada, mainly in the United States, even for workspaces with Canada data storage. Anthropic handles what it is given only to answer, and does not use it to train its models. Finding the parts of your content that match a question is done by a search model on our own servers in Toronto, Canada.

We do not use your workspace's content to train AI models either, ours or anyone's. What the AI writes in a chat is kept as long as that chat and deleted with it; the privacy notice gives every period.

The AI can be wrong, even when it sounds sure. Check anything that matters, keep your content up to date, and have a person review answers on topics where a mistake would hurt.

You are responsible for what your workspace says to your customers, including the AI's answers.

• Each plan includes a number of AI conversations a month. An AI conversation is one in which the AI answered at least once.
• ATON Assist includes 500 suggestions a seat a month, shared across your workspace.
• Your admins are warned at 80% and at 100%. Nothing is charged beyond your plan without your choice.

If your workspace uses QA scorecards, every agent is told that their chats are reviewed, and sees their own reviews. Nobody is scored without knowing.

Your admins can turn the AI agent and ATON Assist off. With the AI agent off, every conversation goes to a person.

When this policy changes, we publish a new version with its date, and earlier versions stay readable on this page. Questions: privacy@aton-network.org.