Updated on 26 October 2026
It opens the day before you ask
In the morning, the assistant has already written up what is waiting: the day's meetings, the news that concerns you and that you haven't read yet, the replies that have come in on your requests, the client files that are pressing, the ones that are late. The text is short.
Every piece of information there is clickable: the client file mentioned opens, the news item is marked as read without leaving the page. This daily briefing is regenerated only when the facts change, instead of calling the model every time you open it.
It answers the way the office next door would
The question is asked in plain language. Before searching, the assistant works out what it is about, and when it falls to an agent, it hands it over. If it hesitates between two pieces of software, it asks rather than guesses. Then it searches the documentation that has been connected and the company's memory, and it answers citing its source with the link.
It also goes and looks at live data when the question calls for it: your diary, the client files in your scope, the news, your open requests. A team member never sees a client file that isn't theirs, and neither does the assistant: scope is checked on the server, not in the instruction given to the model.
A screenshot can be pasted into the conversation, and it reads it. A question can also be dictated into the microphone, using the browser's speech recognition; when the browser does not offer it, dictation can go through your instance and an engine you choose.
It says when it doesn't know
This is the one point we will not compromise on. An assistant that invents a procedure wastes more time than it saves, and it destroys trust in a single answer.
When the answer exists nowhere, the question does not die for all that: if you decide so, it goes to your subject expert, to your IT person, or to the software vendor's support. It becomes a tracked request, and its answer, once validated, joins the knowledge base.
It points you to the right person
Some questions need a human. The assistant knows who on the team has already handled a similar case or has declared the skill, and says so. Whatever someone has approved as an answer can then be viewed on their profile, under "What they have shared with the team".
It suggests, you click
When an answer calls for an action, the assistant offers it in a card with a button: book a meeting, create a task, write to a colleague, draft the reply to a client, mark a request as handled. Nothing happens without your click.
What it doesn't do
It does not decide for you and it signs nothing. It writes no reply to a client that goes out without a human having read and approved it, except those certified by an authorised person, which are then signed as automatic. It does not replace the expert-comptable (chartered accountant), the lawyer or the IT person: it takes off their hands the questions whose answer already existed.
It answers, and it routes. When a question belongs to an agent, the IT specialist for a piece of software, the financial analyst for your figures, it hands the question over and brings the answer back to you. Running a task from start to finish, working through an inbox or an after-sales ticket, is the job of the AI agents, which have their own page.
