What is the problem with AI that guesses?
You ask the model about a procedure your company agreed on two years ago. You get a coherent answer, but not one that reflects your own agreement. Without access to company notes, the answer may sound convincing even though it is not based on what you agreed. It's hard to tell knowledge from guesswork by tone alone.
In Task Manager, you indicate the sources for an answer before asking a question: tasks, a chat with a client, a note with a procedure, a folder from Files, or a calendar. If there's more material than the model can take in at once, the assistant tells you what it didn't include. It first drops general scopes. Tasks indicated explicitly are kept. So you know where the scope of the answer ends. In the Pro variant, the assistant also draws on the whole company's knowledge base, which the program builds in the background.
Where does the assistant get its answers?
There is a Sources button above the conversation. Browse the selection window as you would Files: from the Workspace through tasks and chats to a thread and its folders. Click a row to select a source, and use the arrow to go one level deeper. When you select a task or chat, the Assistant also takes into account files added to it - both from messages and from the Files module. The number of these files appears next to the thread name.
Below the source list, the usage bar shows how much of the selected knowledge has fit. You don't need to split a long note: the Assistant reads it in sections and retrieves more as needed. A file selected as a knowledge source is read up to 300 000 characters, in sections of around 24 000.
A separate safeguard applies to actions. When you ask for a reminder to be set or a task to be created, and the assistant's reply claims it's been done, the program checks whether that action was actually carried out. If not, a yellow "Correction" block appears below the reply. It tells you what happened and what phrase to use to get the action done. This is checked by the program itself using fixed rules, not by a second AI model that could make the same mistake as the first.
What do you gain?
You don't need to check every answer from scratch. When you know which note and thread the Assistant used, you can go back to the specific agreements instead of trying to recall the whole matter from memory.
A new team member doesn't need to ask several colleagues for different versions of the same procedure. They can rely on what someone has written down. If they can't find an agreement in the available materials, they can check with the person who documented it rather than taking a guess as company policy.
You also don't have to assume a reminder exists just because the Assistant said it set one. If the action didn't happen, you'll see a correction before you rely on it.
Accountancy firm, second half of January
A client calls with a question about changing their tax status. The call is answered by someone who has been working at the firm for a month. They know the agreement is somewhere between the chat with the client, a note from last year's conversation and a letter in a DOCX file added to the task.
They select three sources: the client's thread, the procedure note and a folder in Files. Selecting the thread also includes the letter attached to it as a DOCX file. They ask what was agreed and what the deadline is. Before passing the answer on to the client, they can compare it with the note and the letter. If they can't find a deadline in them, they still need to check with the accountant.
Frequently asked questions
Is the content of my documents sent to the AI provider?
Yes. The content of the selected source is sent to an external AI provider so the model can use it to answer. The administrator sets the provider and API key under Settings → AI. The options are OpenAI, Gemini, Claude and Grok.
Can the assistant see documents I don't have access to?
No. The Assistant uses materials you have access to. A source you lose access to will be omitted from the next question. The administrator can access all tasks and chats, but cannot see another user's private note unless it has been shared with them.
Can I use my own Mail as a source?
Yes, if the administrator enables this option under Settings → AI. It is switched off by default. Once enabled, everyone can select only their own mailboxes. The administrator cannot select someone else's Mail either.
What definitely won't the assistant read?
The Assistant won't read older DOC files in binary format or PDFs that are scans without a text layer, as it does not have OCR. It reads up to around 50 KB of text from an email attachment. If it's longer, it sees only the beginning.