AI disclosure

An AI employee has a job (for example Bookings, Social or Research), a list of skills and a list of limits. When you give it a task, our server writes instructions from that job, adds what you have saved that is relevant, sends it all to a language model, and saves the answer for you. The names, faces and levels are there to make the product easy to use; they do not mean there is a person or a mind behind them.

Today the work is done by Claude models made by Anthropic: Claude Haiku, Claude Sonnet and Claude Opus, depending on the workstation tier you chose for that employee. We reach them through Anthropic directly, or through OpenRouter, a routing service, when the direct route is not configured. Reading a web page and searching the web are done with Anthropic's own tools and are not available on the OpenRouter route.

We may change models or providers as better ones appear. The Sub-processors page always shows who is used.

• The employee's job, skills and limits, and your business's name, kind and city.
• Your business brain: company profile, projects, recent updates and the passages of your saved knowledge that fit the task.
• Names and notes of up to 60 clients from Memory, without their contact details.
• Corrections you gave that employee before, and the task you wrote.
• Pages it read for the task: links you pasted and, in jobs that search, the pages it found.
• From tables you connected and shared with that employee: rows of the columns that were not held back as personal.
• From an app you connected and allowed that employee to use (a mailbox, a calendar, a code repository): what it read there for the task. There is no personal-data filter on a mailbox.
• For the website assistant: the text you wrote for it and the visitor's conversation. For the setup chat: your conversation with it.

Not sent: your password, your payment details, the contents of your vault, your CRM records, and connection secrets such as a HubSpot key.

What the providers do with it: [CONFIRM PROVIDER TERMS ON TRAINING AND RETENTION]

• A model can state something false with complete confidence: a price, a date, a name, a quotation, a source, a number.
• It can misread a document or a table, miss what matters, or answer a different question from the one you asked.
• It knows only what it was given and what it learned in training, which has a cut-off date. It does not know your business beyond what you saved.
• Text on a web page or in an email it reads may try to give it instructions. The employees are told to treat everything they read as material and never as orders, but no such defence is perfect.
• Two people who ask for similar things may get similar text.

So read the work before you use it. Check numbers, names, dates and anything with legal or financial weight. The work is a draft from an assistant, not advice from a professional.

• Work comes back to you as text. Nothing is sent, published, booked or paid by an AI employee on its own.
• When an employee wants to do something in an app you connected (send an email, add a calendar event, open an issue or a pull request), it writes a proposal. You see every detail and approve or decline. An approved proposal is carried out once; nothing that deletes, moves money or changes settings or sharing is offered at all.
• You choose which employee may read which source, and can take that back at any time.
• You can ask for a redo with a note, and the employee keeps the note for next time.
• Our own staff do not routinely read your tasks or answers.

The website assistant is the one place where an AI answers people other than you without your reading each answer first. It answers only from the text you wrote or approved for it, it says in its window that it is an AI assistant and that messages are saved for the business, and when a visitor wants a person it takes a message for you instead of improvising. Daily caps limit how much it can say. You can read every conversation in Settings and switch it off at any time.

• We do not train AI models of our own on your content.
• We do not use AI to make decisions about you that have legal or similarly significant effects.
• We do not let an AI employee see another customer's data.
• We do not present AI output as written by a person.

If an AI employee produced something harmful or clearly wrong, we want to know: [CONTACT EMAIL].